The DOI matches its Crossref record; 213 of the 278 references it cites were found in a registry or an index; none of the 100 sources its claims cite is checked for support yet.
Bringing AI to edge: From deep learning’s perspective
Open accessOverview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
72 Not checked yet · 33 No basis given · 26 Own interpretation · 1 Own result
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Most of the above-mentioned structure pruning methods are rule-based, i.e., some heuristic algorithms devised according to their own criterion.”
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“the probability of the correct class dominates the probability distribution generated by the teacher network (e.g., [0.97, 0.1, 0.0, 0.2]), which significantly limits the knowledge transferring capability.”
§ Bringing AI To Edge: From Deep Learning's Perspective > C. Knowledge Distillation
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“All of the above-mentioned approaches mainly search for the best operators for cells or layers where their width and depth are fine-tuned manually.”
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“the majority of existing methods use platform-independent metrics, i.e., FLOPs and MACs as the constraint to design the model.”
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
In the paper
“as shown in [50], the width and depth of a DNN have a critical impact on its accuracy and latency.”
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
What the sources say
- Not checked yetcites [50] source
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“quantization becomes a promising approach to address the aforementioned issue, which encodes full-precision (FP32) weights and activations with low-precision ones (e.g., FP16, INT8, binary) while preserving even the same level of accuracy.”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Quantization
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“searching in a discrete space using RL and EA is really costly”
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: unclear — It is unclear why the paper says it holds.
In the paper
“knowledge distillation can also be used to design robust networks [151].”
§ Bringing AI To Edge: From Deep Learning's Perspective > C. Knowledge Distillation
What the sources say
- Not checked yetcites [151] source
plainconditional- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
In the paper
“Knowledge distillation demonstrates better transferability for the small models, i.e., for a given dataset, learning the small model from knowledge distillation outperforms learning it from scratch.”
§ Bringing AI To Edge: From Deep Learning's Perspective > C. Knowledge Distillation
What the sources say
- Not checked yetcites [150]
boosteddirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
In the paper
“capable of compressing a DNN model, in some cases more than 40x, while only degrading the prediction accuracy slightly [76].”
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
What the sources say
- Not checked yetcites [76]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- It may state two separate claims as one. 38
- Probably not a substantial claim. 22
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“during the run-time, DNN applications share the computational units and communication bandwidth with other applications and as a result the availability of computing and communication resources may significantly affect DNNs' inference time.”
§ Bringing AI To Edge: From Deep Learning's Perspective > VI. ADAPTIVE MODELS
hedgedconditional- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
In the paper
“ResNet-50 [7] has 26 million parameter weights, generates 16 million activations in one inference, requires around 168 MB memory space, and needs at lease 3GB/s memory bandwidth.”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Quantization
What the sources say
- Not checked yetcites [7] source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Almost all of pruning methods are hardware-agnostic and depend on hardware-independent metrics, like MACs and FLOPs.”
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: unclear — It is unclear why the paper says it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Post-training quantization is flexible and can be broadly applied to existing DNN models.”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Quantization
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
In the paper
“since operations with lowbit representation usually consume less energy and execute faster, quantization reduces energy consumption and latency as well on some hardware platforms [129].”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Quantization
What the sources say
- Not checked yetcites [129] source
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“Knowledge distillation is another technique to conduct model compression, where a more compact student model can learn the knowledge from a complicated and powerful teacher model.”
§ Bringing AI To Edge: From Deep Learning's Perspective > C. Knowledge Distillation
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“a diverse of hardware accelerators have emerged for boosting the execution of DNNs, but these various systems demonstrate different capability to handling network complication.”
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
plaincorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
In the paper
“some edge systems are energy-constrained, battery powered or supplied by sustainable energy, like solar energy (intermittent computing [171]), and in these cases energy variance will change the system's status, such as scaling down frequency and power which affects the inference time of DNN applications as well.”
§ Bringing AI To Edge: From Deep Learning's Perspective > VI. ADAPTIVE MODELS
What the sources say
- Not checked yetcites [171] source
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Model compression provides a 'large-to-small' method to generate an efficient and compact model from a complex model for edge devices.”
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Quantization has become a standard means to compress memory-hungry DNN models for resourceconstrained edge systems.”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Quantization
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“neural architecture search, which has demonstrated its capability to design more accurate DNNs without tedious parameter tuning”
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
boosteddirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper may give no reason why it holds.
In the paper
“distilling knowledge from large networks can improve the accuracy of small models which may be directly applicable to edge systems.”
§ Bringing AI To Edge: From Deep Learning's Perspective > C. Knowledge Distillation
What the sources say
- Not checked yetcites [149]
hedgeddirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
In the paper
“Pruning is the well-studied topic in model compression and many methods have been proposed to advance the network pruning [103].”
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
What the sources say
- Not checked yetcites [103] source
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“NAS defines a new paradigm to design DNN models, frees DNN designers from tedious hyperparameter tuning and automates the costly DNN design procedure.”
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds. It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“the state-of-the-art models have more than billions of parameters and at the same time during inference a model produces a large portion of intermediate results (activation/feature maps) which usually occupy a large memory space.”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Quantization
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“network compression involves a costly and time-consuming compression-retraining procedure to retain the accuracy of the pruned model.”
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper may give no reason why it holds.
None of the sources cited for this claim have been checked yet.
In the paper
“MnasNet and MobileNet V3 on Edge TPU perform very low latency, even lower than the high-end GPU”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Hardware-software co-design
What the sources say
- Not checked yetcites [53] source
- Not checked yetcites [45]
- Not checked yetcites [21]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“However, DNN models, the current core component of artificial intelligence, are highly computation-intensive and are even growing 'wider' (more filters within a layer) and 'deeper' (more layers) with billions of parameters and millions of float point operations (FLOPs)”
§ Bringing AI To Edge: From Deep Learning's Perspective > I. INTRODUCTION
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: unclear — It is unclear why the paper says it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“we also discuss adaptive DNN models which are also not reviewed systematically in previous articles.”
§ Bringing AI To Edge: From Deep Learning's Perspective > A. Motivation Behind This Paper
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper may give no reason why it holds.
In the paper
“However, developing chips is a costly process, 30-80 million dollars and 2-3 years to develop [19].”
§ Bringing AI To Edge: From Deep Learning's Perspective > I. INTRODUCTION
What the sources say
- Not checked yetcites [19]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“Depth-wise separable convolution can significantly reduce the computational complexity, thus shortening inference time on edge devices.”
§ Bringing AI To Edge: From Deep Learning's Perspective > A. Classification
What the sources say
- Not checked yetcites [43]
- Not checked yetcites [46]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Convolutional layers account for the major resource and time cost of a CNN model.”
§ Bringing AI To Edge: From Deep Learning's Perspective > II. PRELIMINARIES
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.
In the paper
“SqueezeNet achieves AlexNet-level accuracy while greatly reducing the model complexity.”
§ Bringing AI To Edge: From Deep Learning's Perspective > A. Classification
What the sources say
- Not checked yetcites [41]
plaincorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
In the paper
“Different methodologies are devised to approach this goal, like novel light-weight DNN model design, model compression, and the emergent neural architecture search (NAS) [22]”
§ Bringing AI To Edge: From Deep Learning's Perspective > I. INTRODUCTION
What the sources say
- Not checked yetcites [22]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“However, inference from cloud suffers from responsiveness (e.g., latency) and privacy issues which are critical in some scenarios.”
§ Bringing AI To Edge: From Deep Learning's Perspective > I. INTRODUCTION
What the sources say
- Not checked yetcites [5]
- Not checked yetcites [11]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
In the paper
“some customized edge accelerators achieve high efficiency at the cost of generality, e.g., edge TPU only supports a limited number of operators [21] and some new DNN architectures cannot be wellsupported.”
§ Bringing AI To Edge: From Deep Learning's Perspective > I. INTRODUCTION
What the sources say
- Not checked yetcites [21]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“less-capable edge systems”
- Also stated in: § I. INTRODUCTION · plain
“resource-constrained edge systems”
§ Bringing AI To Edge: From Deep Learning's Perspective > Abstract
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Also stated in: § I. INTRODUCTION · plain
None of the sources cited for this claim have been checked yet.
In the paper
“The off-the-shelf hardware and software cannot efficiently execute those compressed formats, so specialized hardware and software libraries are required to execute sparsified DNN models [83], [84].”
§ Bringing AI To Edge: From Deep Learning's Perspective > Filters Channels Filters Channels
What the sources say
- Not checked yetcites [83]
- Not checked yetcites [84] source
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“the accuracy of the pruned model is inherently upper bounded by the pre-trained model, i.e., the compressed model cannot have better accuracy than the pre-trained model.”
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper may give no reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Since onestage methods are less computational than two-stage methods, all of these methods mentioned above target one-stage method for improving efficiency of object detection DNNs.”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Object Detection
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“Experimental results show deep compression can significantly reduce the model size with no or negligible accuracy loss.”
§ Bringing AI To Edge: From Deep Learning's Perspective > Kernel Weights
What the sources say
- Not checked yetcites [75] source
- Not checked yetcites [76]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“it is capable of fitting the whole model into on-chip memory of edge devices such that the high overhead occurred by off-chip memory access can be mitigated.”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Quantization
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
In the paper
“The one-stage approaches can trade off accuracy for efficiency, but the practical deployment on edge devices is still behind the expected performance [70].”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Object Detection
What the sources say
- Not checked yetcites [70] source
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“Deep Compression uses three steps, prunning, quantization, and Huffman Encoding, to compress over-parameterized DNN models, such as VGG [77], AlexNet [1].”
§ Bringing AI To Edge: From Deep Learning's Perspective > Kernel Weights
What the sources say
- Not checked yetcites [75] source
- Not checked yetcites [76]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“The two-stages methods provide high accuracy at the cost of efficiency, and even on powerful GPUs they cannot guarantee real-time constraints (e.g., 25fps for video).”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Object Detection
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“It is worthy noting that the three approaches are not mutual exclusive to each other and in many cases they are combined to maximally compress the redundant models.”
§ Bringing AI To Edge: From Deep Learning's Perspective > IV. NETWORK COMPRESSION
plaincorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Since structured pruning does not lead to irregular pattern, the compressed network pruned by structure pruning can directly accelerate its inference on off-theshelf hardware platforms without specialized software library support.”
§ Bringing AI To Edge: From Deep Learning's Perspective > Filters Channels Filters Channels
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“different layers have different functions, thereby likely benefiting from employing different pruning criteria at different layers.”
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
hedgedconditional- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Rests on an assumption: Different layers have different functions.
None of the sources cited for this claim have been checked yet.
In the paper
“edge systems can update models locally by using incremental learning [226] [227], where techniques in [228] are deployed to improve the accuracy for new data and infer unknown classes.”
§ Bringing AI To Edge: From Deep Learning's Perspective > E. The data challenge at edge
What the sources say
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
In the paper
“It is known that accuracy of DNN models are highly related to its width (channels or feature maps) and depth (layers or blocks) [50]”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Hardware-software co-design
What the sources say
- Not checked yetcites [50] source
boostedcorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
In the paper
“As stated in [40], DL algorithms are approaching the computational limits of computing systems”
§ Bringing AI To Edge: From Deep Learning's Perspective > VIII. CONCLUSION
What the sources say
- Not checked yetcites [40]
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim.
In the paper
“Such variance may influence the quality of service (QoS) of applications without rigorously temporal requirement, like voice recognition, face recognition, machine translation, etc [20] and on the other hand may lead to catastrophic consequence for those with rigorously temporal requirement, (i.e., real-time requirement), like autonomous driving, UAV, etc.”
§ Bringing AI To Edge: From Deep Learning's Perspective > VI. ADAPTIVE MODELS
What the sources say
- Not checked yetcites [20] source
hedgedconditional- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
In the paper
“FNAS can achieve the same accuracy level under different resource constraints, while significantly reducing the search cost and inference latency.”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Hardware-software co-design
What the sources say
- Not checked yetcites [204]
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“The academia and industry have a consensus that the breakthroughs in hardware architecture will bring DNNs to a new level and boost DNN adoption [15] [205].”
§ Bringing AI To Edge: From Deep Learning's Perspective > B. Hardware-software co-design
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
In the paper
“Attention technique mimics human visual systems to locate their focal point on the important area and up-weights the areas of interest to improve the model accuracy.”
§ Bringing AI To Edge: From Deep Learning's Perspective > VI. ADAPTIVE MODELS
What the sources say
- Not checked yetcites [174] source
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
None of the sources cited for this claim have been checked yet.
In the paper
“Like [218], an efficient point-voxel CNN is proposed for efficient 3D learning, and this model can help to implement 3D AR/VR applications and SLAM [219] of autonomous driving on edge systems.”
§ Bringing AI To Edge: From Deep Learning's Perspective > C. Lightweight models for other applications
plainconditional- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
In the paper
“a recent study in [9] raises a concern regarding the environment effect of training complex NLP models.”
§ Bringing AI To Edge: From Deep Learning's Perspective > C. Lightweight models for other applications
What the sources say
- Not checked yetcites [9] source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
In the paper
“instead of sharing the data with a centralized server clients (e.g., an edge system) in FL only share the learned gradient with others which will not leak data.”
§ Bringing AI To Edge: From Deep Learning's Perspective > E. The data challenge at edge
What the sources say
- Not checked yetcites [229] source
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“Recently, few works study to employ incremental learning [227] and federated learning [230] [231] with edge systems, but the research in this context is still in its infancy.”
§ Bringing AI To Edge: From Deep Learning's Perspective > E. The data challenge at edge
What the sources say
- Not checked yetcites [227]
- Not checked yetcites [230]
- Not checked yetcites [231]
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“Attention has shown its successful application in CV [174] and NLP [180].”
§ Bringing AI To Edge: From Deep Learning's Perspective > VI. ADAPTIVE MODELS
What the sources say
- Not checked yetcites [174] source
- Not checked yetcites [180]
plaincorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“However, on edge systems, some system dynamics, memory contention, cache miss, bandwidth unavailability, etc, also impact the execution of DNN application”
§ Bringing AI To Edge: From Deep Learning's Perspective > Methods
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper may give no reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“data privacy is one motivation of edge systems, so it may be impossible to collect data from different edge systems and share them with each other.”
§ Bringing AI To Edge: From Deep Learning's Perspective > E. The data challenge at edge
hedgedconditional- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. It may state two separate claims as one.
The paper relies on this without support.
At present, edge intelligence systems are used mainly for computer vision tasks.
“currently computer vision models are the major application for edge intelligence systems”
The paper relies on this without support.
Real-world edge deployments face rapidly changing conditions.
“Edge systems installed in real world operate in highly dynamic environment”
The paper relies on this without support.
Manual design of lightweight DNN models depends heavily on expert knowledge and requires lengthy hyperparameter exploration.
“The manual design of lightweight DNN models is considerably dependent on experts' knowledge and also needs a time-consuming hyperparameter exploration.”
The paper relies on this without support.
Different layers have different functions.
“different layers have different functions”
The paper relies on this without support.
Reductions in FLOPs and compression ratio do not directly translate into performance improvement.
“the reduced FLOPs and compression ratio cannot directly translate to the performance improvement.”
The paper relies on this without support.
Different hardware architectures demonstrate different degrees of parallelism.
“different hardware architectures demonstrate different degrees of parallelism”
The paper relies on this without support.
The design space for DNN is too large for exhaustive search.
“In addition, design space for DNN is too large to have an exhaustive search.”
The paper relies on this without support.
There is a lack of measurement of trade-offs between metrics, leading to high search cost or sub-optimal design.
“However, we still lack the measurement of the trade-off between each metric and this will lead to either high searching cost or sub-optimal design result.”
The paper relies on this without support.
Different hardware accelerators have diverse underlying characteristics, and most DNNs were designed without considering underlying hardware features.
“However, different hardware accelerators features diverse underlying characteristics, and the majority of DNNs were designed without consideration of underlying hardware features.”
The paper relies on this without support.
The majority of works on deep learning for edge systems target computer vision tasks such as image classification and object detection.
“Currently, the majority of works regarding deep learning on edge systems target computer vision tasks, i.e., image classification and object detection.”
The paper relies on this without support.
On-device learning is challenging because training is more computation-intensive and memory-hungry than inference, while edge systems are resource-constrained.
“Ondevice learning is a challenging task, because training is a more computation-intensive and memory-hungry procedure compared to inference, whereas edge systems are resourceconstrained in many settings.”
The paper relies on this without support.
Limited energy supply of some edge systems makes on-device learning even more difficult.
“Moreover, limited energy supply of some edge systems will make this issue even more difficult.”
The paper relies on this without support.
The success of DNN models heavily relies on high-quality and large-scale datasets.
“The success of DNN models heavily relies on high-quality and large-scale datasets, such as ImageNet”
The paper relies on this without support.
For some edge applications, collecting and labeling massive data is difficult or expensive.
“for some edge applications, e.g., edge surveillance systems in wild life, defect detection in manufacturing process, etc, it is difficult or expensive to collect massive amount of data and label them to train a good model.”
The paper relies on this without support.
During long-term operation, edge systems are likely to collect data with different distributions or new classes, requiring model updates.
“during the long-term operation, edge systems are likely to collect data with different distributions from the original training data or data pertaining to a new class which is not included in the original training data, we may need to update the model on edge systems”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Machine learning frameworks: ML frameworks, such as TensorFlow Lite and PyTorch, are used to develop and run machine learning models on edge devices [202].”
From Edge AI: A survey (2023) · cites this paper as [202]
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
“HERE, COST REFERS TO THE TIME SPENT TO SEARCH THE NETWORK, WHILE WE USE GPU Hour AS UNIT.”
§ Bringing AI To Edge: From Deep Learning's Perspective > V. HARDWARE-AWARE NEURAL ARCHITECTURE SEARCH
“Cost (GPU days) Metrics Target Edge Systems MnasNet [53] Measure RL 40000”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Bringing AI To Edge: From Deep Learning's Perspective > A. Different metrics oriented DNN models
“Fig. 11 shows the latency of 100000 network architectures generated using DARTS [166]”
§ Bringing AI To Edge: From Deep Learning's Perspective > VIII. CONCLUSION
“Fig. 11.Latency vs FLOPs: The latency measurements of 10000 different architectures with different FLOPs count.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
“BinaryConnect [111] 1-bit Yes No 2016”
§ Bringing AI To Edge: From Deep Learning's Perspective > References
“111. M. Courbariaux, Y. Bengio, and J.-P. David, "Binaryconnect: Training deep neural networks with binary weights during propagations," in Advances in neural information processing systems, 2015, pp. 3123- 3131.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Bringing AI To Edge: From Deep Learning's Perspective > Methods
“BranchyNet [189] Early-exit Gradient 2017”
§ Bringing AI To Edge: From Deep Learning's Perspective > References
“189. S. Teerapittayanon, B. McDanel, and H.-T. Kung, "Branchynet: Fast inference via early exiting from deep neural networks," in 2016 23rd International Conference on Pattern Recognition (ICPR).”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
“Li et al. [87] Inference-based Layer Rule-based 2016”
§ Bringing AI To Edge: From Deep Learning's Perspective > Methodologies
“Li et al. [87] adopt a global pruning method”
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 3×
“For instance, training ResNet50 [7] with ImageNet on 8 Nvidia Tesla P100 takes 29 hours [8], where ImageNet dataset [2] has 1.28 M images of 1000 classes for training.”
BackgroundMethodInfluentialVery Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 1×
“Deep Compression uses three steps, prunning , quantization , and Huffman Encoding , to compress over-parameterized DNN models, such as VGG [77], AlexNet [1].”
MethodBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
ImageNet: A large-scale hierarchical image database
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · 2009 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2009.5206848 · Cited by 65,333 · cited 2×
“For instance, training ResNet50 [7] with ImageNet on 8 Nvidia Tesla P100 takes 29 hours [8], where ImageNet dataset [2] has 1.28 M images of 1000 classes for training.”
BackgroundMethodInfluentialBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325 · cited 1×
“These models, like BERT [217], are known to be highly complex, even more complicated than DNN models for image classi-fication, and a recent study in [9] raises a concern regarding the environment effect of training complex NLP models.”
MethodYou Only Look Once: Unified, Real-Time Object Detection
Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.91 · Cited by 53,803 · cited 3×
“To boost the efficiency of object detection on resource-limited devices, some works strive to reduce the computational complexity of backbone part of object detection models, like Tiny-dsod [58] for DSOD [69], Tiny-SSD [59] for SSD [65], and Tiny-YOLO [57] for YOLO.”
MethodA fast and elitist multiobjective genetic algorithm: NSGA-II
Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002) · IEEE Transactions on Evolutionary Computation · doi:10.1109/4235.996017 · Cited by 50,053
Microsoft COCO: Common Objects in Context
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-10602-1_48 · Cited by 45,460 · cited 1×
“EXPERIMENTAL RESULTS ARE UPON TWO WIDELY-USED BENCHMARK, PASCAL VOC2007 [54] AND MS COCO [55].”
MethodSqueeze-and-Excitation Networks
Hu, J., Shen, L., & Sun, G. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00745 · Cited by 35,667 · cited 1×
“To improve the performance and training speed, modern DNN models also have other operation layers, like batch normalization layer [38], squeeze-and-excitation layer [39], etc.”
MethodRich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation
Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014) · 2014 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2014.81 · Cited by 32,684 · cited 1×
“RCNN [66], Fast RCNN [67], and 3 https://https://www.automl.org/ Faster RCNN [68] are representative examples of two-stage methods.”
Method
Research works printed but not linked: 37
- A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep convolutional neural networks, in: P.L. Bartlett, F.C.N. Pereira, C.J.C. Burges, L. Bottou, K.Q. Weinberger (Eds.), Advances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems 2012. Proceedings of a meeting held December 3-6, 2012, Lake Tahoe, Nevada, United States, 2012, pp. 1106–1114. URL:https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html
- Goodfellow, 2016, vol. 1
- Ieee computer society’s top 12 technology trends for 2020, URL: https://www.computer.org/press-room/2019-news/ieee-computer-societys-top-12-technology-trends-for-2020, access: 2020-5-14.
- Wei, 2021, Wireless edge caching based on content similarity in dynamic environments, Journal of Systems Architecture, 102000
- Choudhary, 2020, A comprehensive survey on model compression and acceleration, Artificial Intelligence Review, 1
- Y. Guo, H. Wang, Q. Hu, H. Liu, L. Liu, M. Bennamoun, Deep learning for 3d point clouds: A survey, IEEE Transactions on Pattern Analysis and Machine Intelligence.
- M. Feurer, A. Klein, K. Eggensperger, J.T. Springenberg, M. Blum, F. Hutter, Efficient and robust automated machine learning, in: C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, R. Garnett (Eds.), Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7–12, 2015, Montreal, Quebec, Canada, 2015, pp. 2962–2970. URL:https://proceedings.neurips.cc/paper/2015/hash/11d0e6287202fced83f79975ec59a3a6-Abstract.html.
- J. Wang, T.A. Bohn, C.X. Ling, Pelee: A real-time object detection system on mobile devices, in: S. Bengio, H.M. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, R. Garnett (Eds.), Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3–8, 2018, Montréal, Canada, 2018, pp. 1967–1976. URL:https://proceedings.neurips.cc/paper/2018/hash/9908279ebbf1f9b250ba689db6a0222b-Abstract.html.
- S. Ren, K. He, R.B. Girshick, J. Sun, Faster R-CNN: towards real-time object detection with region proposal networks, in: C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, R. Garnett (Eds.), Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7–12, 2015, Montreal, Quebec, Canada, 2015, pp. 91–99. URL:https://proceedings.neurips.cc/paper/2015/hash/14bfa6bb14875e45bba028a21ed38046-Abstract.html.
- M. Denil, B. Shakibi, L. Dinh, M. Ranzato, N. de Freitas, Predicting parameters in deep learning, in: C.J.C. Burges, L. Bottou, Z. Ghahramani, K.Q. Weinberger (Eds.), Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Proceedings of a meeting held December 5–8, 2013, Lake Tahoe, Nevada, United States, 2013, pp. 2148–2156. URL: https://proceedings.neurips.cc/paper/2013/hash/7fec306d1e665bc9c748b5d2b99a6e97-Abstract.html.
- Y. LeCun, J.S. Denker, S.A. Solla, Optimal brain damage, in: Advances in neural information processing systems, 1990, pp. 598–605.
- S. Han, J. Pool, J. Tran, W.J. Dally, Learning both weights and connections for efficient neural network, in: C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, R. Garnett (Eds.), Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7–12, 2015, Montreal, Quebec, Canada, 2015, pp. 1135–1143. URL: https://proceedings.neurips.cc/paper/2015/hash/ae0eb3eed39d2bcef4622b2499a05fe6-Abstract.html.
- S. Han, H. Mao, W.J. Dally, Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding, International Conference on Learning Representations (ICLR).
- D. Molchanov, A. Ashukha, D.P. Vetrov, Variational dropout sparsifies deep neural networks, in: D. Precup, Y.W. Teh (Eds.), Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6–11 August 2017, Vol. 70 of Proceedings of Machine Learning Research, PMLR, 2017, pp. 2498–2507. URL: http://proceedings.mlr.press/v70/molchanov17a.html.
- Deb, 2001, vol. 16
- H. Li, A. Kadav, I. Durdanovic, H. Samet, H.P. Graf, Pruning filters for efficient convnets, in: 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24–26, 2017, Conference Track Proceedings, OpenReview.net, 2017. URL: https://openreview.net/forum?id=rJqFGTslg.
- W. Wen, C. Wu, Y. Wang, Y. Chen, H. Li, Learning structured sparsity in deep neural networks, in: D.D. Lee, M. Sugiyama, U. von Luxburg, I. Guyon, R. Garnett (Eds.), Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5–10, 2016, Barcelona, Spain, 2016, pp. 2074–2082. URL: https://proceedings.neurips.cc/paper/2016/hash/41bfd20a38bb1b0bec75acf0845530a7-Abstract.html.
- M. Courbariaux, Y. Bengio, J. David, Binaryconnect: Training deep neural networks with binary weights during propagations, in: C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama, R. Garnett (Eds.), Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7–12, 2015, Montreal, Quebec, Canada, 2015, pp. 3123–3131. URL: https://proceedings.neurips.cc/paper/2015/hash/3e15cc11f979ed25912dff5b0669f2cd-Abstract.html.
- A. Tarvainen, H. Valpola, Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results, in: I. Guyon, U. von Luxburg, S. Bengio, H.M. Wallach, R. Fergus, S.V.N. Vishwanathan, R. Garnett (Eds.), Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4–9, 2017, Long Beach, CA, USA, 2017, pp. 1195–1204. URL: https://proceedings.neurips.cc/paper/2017/hash/68053af2923e00204c3ca7c6a3150cf7-Abstract.html.
- H. Cai, C. Gan, T. Wang, Z. Zhang, S. Han, Once-for-all: Train one network and specialize it for efficient deployment, in: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26–30, 2020, OpenReview.net, 2020. URL: https://openreview.net/forum?id=HylxE1HKwS.
- C. Li, Z. Yu, Y. Fu, Y. Zhang, Y. Zhao, H. You, Q. Yu, Y. Wang, C. Hao, Y. Lin, HW-nas-bench: Hardware-aware neural architecture search benchmark, in: International Conference on Learning Representations, 2021. URL: https://openreview.net/forum?id=_0kaDkv3dVf.
- J. Lin, Y. Rao, J. Lu, J. Zhou, Runtime neural pruning, in: I. Guyon, U. von Luxburg, S. Bengio, H.M. Wallach, R. Fergus, S.V.N. Vishwanathan, R. Garnett (Eds.), Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4–9, 2017, Long Beach, CA, USA, 2017, pp. 2181–2191. URL: https://proceedings.neurips.cc/paper/2017/hash/a51fb975227d6640e4fe47854476d133-Abstract.html.
- B.E. Bejnordi, T. Blankevoort, M. Welling, Batch-shaping for learning conditional channel gated networks, in: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26–30, 2020, OpenReview.net, 2020. URL: https://openreview.net/forum?id=Bke89JBtvB.
- T. Bolukbasi, J. Wang, O. Dekel, V. Saligrama, Adaptive neural networks for efficient inference, in: D. Precup, Y.W. Teh (Eds.), Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6–11 August 2017, Vol. 70 of Proceedings of Machine Learning Research, PMLR, 2017, pp. 527–536. URL: http://proceedings.mlr.press/v70/bolukbasi17a.html.
- K. Xu, J. Ba, R. Kiros, K. Cho, A.C. Courville, R. Salakhutdinov, R.S. Zemel, Y. Bengio, Show, attend and tell: Neural image caption generation with visual attention, in: F.R. Bach, D.M. Blei (Eds.), Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6–11 July 2015, Vol. 37 of JMLR Workshop and Conference Proceedings, JMLR.org, 2015, pp. 2048–2057. URL: http://proceedings.mlr.press/v37/xuc15.html.
- A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need, in: I. Guyon, U. von Luxburg, S. Bengio, H.M. Wallach, R. Fergus, S.V.N. Vishwanathan, R. Garnett (Eds.), Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4–9, 2017, Long Beach, CA, USA, 2017, pp. 5998–6008. URL: https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html.
- A. Veit, M.J. Wilber, S.J. Belongie, Residual networks behave like ensembles of relatively shallow networks, in: D.D. Lee, M. Sugiyama, U. von Luxburg, I. Guyon, R. Garnett (Eds.), Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5–10, 2016, Barcelona, Spain, 2016, pp. 550–558. URL: https://proceedings.neurips.cc/paper/2016/hash/37bc2f75bf1bcfe8450a1a41c200364c-Abstract.html.
- T. Chen, T. Moreau, Z. Jiang, L. Zheng, E. Yan, H. Shen, M. Cowan, L. Wang, Y. Hu, L. Ceze, et al., {TVM}) An automated end-to-end optimizing compiler for deep learning, in: 13th {USENIX})Symposium on Operating Systems Design and Implementation ({OSDI})18), 2018, pp. 578–594.
- Z. Liu, H. Tang, Y. Lin, S. Han, Point-voxel CNN for efficient 3d deep learning, in: H.M. Wallach, H. Larochelle, A. Beygelzimer, F. d’Alché-Buc, E.B. Fox, R. Garnett (Eds.), Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8–14, 2019, Vancouver, BC, Canada, 2019, pp. 963–973. URL: https://proceedings.neurips.cc/paper/2019/hash/5737034557ef5b8c02c0e46513b98f90-Abstract.html.
- I.J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A.C. Courville, Y. Bengio, Generative adversarial nets, in: Z. Ghahramani, M. Welling, C. Cortes, N.D. Lawrence, K.Q. Weinberger (Eds.), Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8–13 2014, Montreal, Quebec, Canada, 2014, pp. 2672–2680. URL: https://proceedings.neurips.cc/paper/2014/hash/5ca3e9b122f61f8f06494c97b1afccf3-Abstract.html.
- Y. Wang, Z. Jiang, X. Chen, P. Xu, Y. Zhao, Y. Lin, Z. Wang, E2-train: Training state-of-the-art cnns with over 80% energy savings, in: H.M. Wallach, H. Larochelle, A. Beygelzimer, F. d’Alché-Buc, E.B. Fox, R. Garnett (Eds.), Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8–14, 2019, Vancouver, BC, Canada, 2019, pp. 5139–5151. URL: https://proceedings.neurips.cc/paper/2019/hash/663772ea088360f95bac3dc7ffb841be-Abstract.html.
- Pan, 2009, A survey on transfer learning, IEEE Transactions on Knowledge and Data Engineering, 22, 1345, 10.1109/TKDE.2009.191 · DOI · Paper page
- Dosovitskiy, 2021, An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, International Conference on Learning Representations (ICLR)
- Luo, 2021, HSCoNAS: Hardware-Software Co-Design of Efficient DNNs via Neural Architecture Search, Design, Automation \& Test in Europe Conference \& Exhibition (DATE)
- Luo, 2021, Designing Efficient DNNs via Hardware-Aware Neural Architecture Search and Beyond, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD)
- Kong, 2022, HACScale: Hardware-Aware Compound Scaling for Resource-Efficient DNNs, Asia and South Pacific Design Automation Conference (ASP-DAC)
- Huai, 2021, ZeroBN: Learning Compact Neural Networks For Latency-Critical Edge Systems, ACM/IEEE Design Automation Conference (DAC)
Cited by
Sorted by citations
Singh, R., & Gill, S. S. (2023) · Internet of Things and Cyber-Physical Systems · doi:10.1016/j.iotcps.2023.02.004 · Cited by 572
Distributed Artificial Intelligence Empowered by End-Edge-Cloud Computing: A Survey
Duan, S., Wang, D., Ren, J., Lyu, F., Zhang, Y., Wu, H., & Shen, X. (2022) · IEEE Communications Surveys & Tutorials · doi:10.1109/comst.2022.3218527 · Cited by 439
“There have been some surveys investigating AI-powered cloud-edge orchestration [5], [6], [8], [9], [10], edge intelligence [12], [13], [14], [15], end-edge-cloud orchestration [7], [11], AI/ML for wireless edge network [16], [17], and privacy and security for distributed learning [18].”
BackgroundEdge AI: A Taxonomy, Systematic Review and Future Directions
Gill, S. S., Golec, M., Hu, J., Xu, M., Du, J., Wu, H., Walia, G. K., Murugesan, S. S., Ali, B., Kumar, M., et al. (2024) · Cluster Computing · doi:10.1007/s10586-024-04686-y · Cited by 280
“The study emphasises the significance and changing challenges of resource management [8], whereas a presentation framework is described that addresses the issue of sparse training data in emerging radio access networks by utilizing a range of methods, such as interpolation, domain-knowledge…”
BackgroundEdge deep learning in computer vision and medical diagnostics: a comprehensive survey
Xu, Y., Khan, T. M., Song, Y., & Meijering, E. H. W. (2025) · Artificial Intelligence Review · doi:10.1007/s10462-024-11033-5 · Cited by 187
E-YOLO: Recognition of estrus cow based on improved YOLOv8n model
Wang, Z., Hua, Z., Wen, Y., Zhang, S., Xu, X., & Song, H. (2023) · Expert Systems with Applications · doi:10.1016/j.eswa.2023.122212 · Cited by 128
AI augmented Edge and Fog computing: Trends and challenges
Tuli, S., Mirhakimi, F., Pallewatta, S., Zawad, S., Casale, G., Javadi, B., Yan, F., Buyya, R., & Jennings, N. R. (2023) · Journal of Network and Computer Applications · doi:10.1016/j.jnca.2023.103648 · Cited by 147
Green Edge AI: A Contemporary Survey
Mao, Y., Yu, X., Huang, K., Zhang, Y. A., & Zhang, J. (2024) · Proceedings of the IEEE · doi:10.1109/jproc.2024.3437365 · Cited by 140
“Resource-efficient on-device and collaborative edge inference strategies were elaborated in [45] and [46], respectively.”
BackgroundMachine Learning in Manufacturing towards Industry 4.0: From ‘For Now’ to ‘Four-Know’
Chen, T., Sampath, V., May, M. C., Shan, S., Jorg, O. J., Martín, J. J. A., Stamer, F., Fantoni, G., Tosello, G., & Calaon, M. (2023) · Applied Sciences · doi:10.3390/app13031903 · Cited by 110
“Several methods, including model pruning [83], model quantization [84], and knowledge distillation [85], have been suggested in the literature as ways to compress these dense models.”
MethodVukićević, A. M., Petrovic, M., Milošević, P., Peulić, A., Jovanovic, K. M., & Novakovic, A. (2024) · Artificial Intelligence Review · doi:10.1007/s10462-024-10978-x · Cited by 76
Enabling All In-Edge Deep Learning: A Literature Review
Joshi, P., Hasanuzzaman, M., Thapa, C., Afli, H., & Scully, T. (2023) · IEEE Access · doi:10.1109/access.2023.3234761 · Cited by 67
“Model size or memory footprint is computed having ‘MB’ as their unit of measurement [242], [243], [244], [245], [246].”
MethodMultisensory Metaverse-6G: A New Paradigm of Commerce and Education
ALSAMH, M. H., Hawbani, A., Kumar, S., & Alsamhi, S. H. (2024) · IEEE Access · doi:10.1109/access.2024.3392838 · Cited by 63
“Through the use of DNN [44] and reinforcement learning techniques, which scan large datasets to discover behavioural patterns and train models appropriately, AI technology allows edge computing to handle complicated intelligence-related concerns.”
MethodLu, K., Huang, J., Zeng, G., Chen, M.-R., Geng, G., & Weng, J. (2025) · IEEE Transactions on Dependable and Secure Computing · doi:10.1109/tdsc.2025.3545363 · Cited by 60
“MFLOPs is a widely-used and hardware-independent metrics to evaluate model complexity [69], so except for the above four indices, we include MFLOPs in the evaluation index for a comprehensive comparison between the state-of-the-art methods and the CNN model optimized by the MODEO-CNN.”
Silva, A. V., Camargo, J., Almeida, J., & Cugnasca, P. (2022) · IEEE Access · doi:10.1109/access.2022.3229233 · Cited by 51
“…AI area trend in using them because of their flexibility in providing somewhat accurate models for functions which are hard to be precisely specified [54], [55] and that can ultimately benefit from exploring and exploiting an operational environment for fine-tuning prior to revenue service [56].”
BackgroundBig Data and AI Algorithms for Sustainable Development Goals: A Topic Modeling Analysis
Nedungadi, P., Surendran, S., Tang, K., & Raghu Raman, R. (2024) · IEEE Access · doi:10.1109/access.2024.3516500 · Cited by 49
LocateUAV: Unmanned Aerial Vehicle Location Estimation via Contextual Analysis in an IoT Environment
Dilshad, N., Ullah, A., Kim, J., & Seo, J. (2022) · IEEE Internet of Things Journal · doi:10.1109/jiot.2022.3162300 · Cited by 45
“Network compression comprises several techniques that focus on this issue by eliminating redundant and overparameterized network links [24].”
BackgroundA survey on privacy for B5G/6G: New privacy challenges, and research directions
Sandeepa, C., Siniarski, B., Kourtellis, N., Wang, S., & Liyanage, M. (2022) · Journal of Industrial Information Integration · doi:10.1016/j.jii.2022.100405 · Cited by 49
Key Considerations for Real-Time Object Recognition on Edge Computing Devices
Surantha, N., & Sutisna, N. (2025) · Applied Sciences · doi:10.3390/app15137533 · Cited by 52
Vu, C. C. (2024) · Materials Today Physics · doi:10.1016/j.mtphys.2024.101376 · Cited by 37
Bhuiyan, M. S. M., Rafi, M. A., Rodrigues, G. N., Mir, M. N. H., Ishraq, A., Mridha, M. F., & Shin, J. (2025) · Array · doi:10.1016/j.array.2025.100390 · Cited by 65
Generative Adversarial Super-Resolution at the edge with knowledge distillation
Angarano, S., Salvetti, F., Martini, M., & Chiaberge, M. (2023) · Engineering Applications of Artificial Intelligence · doi:10.1016/j.engappai.2023.106407 · Cited by 39
Dhar, N., Deng, B., Lo, D. C.-T., Wu, X., Zhao, L., & Suo, K. (2024) · ACM Southeast Regional Conference (ACMSE) · doi:10.1145/3603287.3651205 · Cited by 45
Latency-Aware Offloading for Mobile Edge Computing Networks
Feng, W., Liu, H., Yao, Y., et al. (2021) · IEEE Communications Letters · doi:10.1109/lcomm.2021.3074621 · Cited by 34
Distributed Intelligence in Wireless Networks
Liu, X., Yu, J., Liu, Y., Gao, Y., Mahmoodi, T., Lambotharan, S., & Tsang, D. H. K. (2023) · IEEE Open Journal of the Communications Society · doi:10.1109/ojcoms.2023.3265425 · Cited by 41
“…that have tried to address this interdisciplinary problem of AI and wireless communications from the aspects of edge intelligence [31], [32], and distributed ML in wireless communications [9], [22], [23], [24], [25], [26], [27], [28], [29], [30], but their focus is different from our work.”
BackgroundLightNAS: On Lightweight and Scalable Neural Architecture Search for Embedded Platforms
Luo, X., Liu, D., Kong, H., Huai, S., Chen, H., & Liu, W. (2022) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2022.3208187 · Cited by 32
“As such, they end up with the architecture with promising accuracy on target task, which, however, comes at the cost of high computational complexity on target hardware, thereby impeding the practical deployment of DNNs, especially on resource-constrained embedded platforms [21].”
BackgroundInfluentialSelf-citationKreković, D., Krivić, P., Žarko, I. P., Kušek, M., & Le-Phuoc, D. (2025) · Internet of Things · doi:10.1016/j.iot.2025.101553 · Cited by 33
A Benchmark of Deep Learning Models for Multi-leaf Diseases for Edge Devices
Anh, P. T., & Duc, H. T. M. (2021) · International Conference on Advanced Technologies for Communications · doi:10.1109/atc52653.2021.9598196 · Cited by 31
Deep Learning on Energy Harvesting IoT Devices: Survey and Future Challenges
Lv, M., & Xu, E. (2022) · IEEE Access · doi:10.1109/access.2022.3225092 · Cited by 26
“Recently, new DNNs with small footprints were developed to run on MCUs with very limited resource [21].”
BackgroundSurvey of Deep Learning Accelerators for Edge and Emerging Computing
Alam, S., Yakopcic, C., Wu, Q., Barnell, M., Khan, S., & Taha, T. M. (2024) · Electronics · doi:10.3390/electronics13152988 · Cited by 38
Pazhanivel, D. B., Narayanan, V. A., & Sivakumar, P. B. (2024) · Sensors · doi:10.3390/s24155069 · Cited by 30
Sada, A. B., Khelloufi, A., Naouri, A., et al. (2024) · Cluster Computing · doi:10.1007/s10586-024-04578-1 · Cited by 22
Federated Edge Intelligence and Edge Caching Mechanisms
Karras, A., Karras, C., Giotopoulos, K. C., Tsolis, D. K., Οικονόμου, Κ., & Sioutas, S. (2023) · Information · doi:10.3390/info14070414 · Cited by 22
A lightweight method for face expression recognition based on improved MobileNetV3
Liang, X., Liang, J., Yin, T., & Tang, X. (2023) · IET Image Processing · doi:10.1049/ipr2.12798 · Cited by 29
Luo, X., Liu, D., Kong, H., Huai, S., Chen, H., Xiong, G., & Liu, W. (2024) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3701728 · Cited by 27
Self-citationQuantization Aware Attack: Enhancing Transferable Adversarial Attacks by Model Quantization
Yang, Y., Lin, C., Li, Q., Zhao, Z., Fan, H., Zhou, D., Wang, N., Liu, T., & Shen, C. Q. (2024) · IEEE Transactions on Information Forensics and Security · doi:10.1109/tifs.2024.3360891 · Cited by 26
“R ECENT progress in edge computing [1] has triggered a large demand for deep neural networks (DNNs) that can be deployed on edge devices.”
BackgroundNasir, T., & Malik, M. K. (2023) · Information Processing & Management · doi:10.1016/j.ipm.2023.103544 · Cited by 21
Deep learning based approaches to enhance energy efficiency in autonomous driving systems
Kwak, J., Lee, Y., Choi, M., & Lee, S. (2024) · Energy · doi:10.1016/j.energy.2024.132625 · Cited by 21
A Decade of Video Analytics at Edge: Training, Deployment, Orchestration, and Platforms
Bai, T., Zhao, H., Huang, L., Wang, Z., Kim, D. I., & Nallanathan, A. (2025) · IEEE Communications Surveys & Tutorials · doi:10.1109/comst.2025.3563377 · Cited by 25
“Furthermore, as the fundamental enabling technique of VA, AI in the edge computing scenarios, or the so-called edge intelligence , has been reviewed in the existing research efforts [17], [18], [19], [20], [21], [22], [24], [25], [26], [27], [28], [29], [30], [31].”
PLiNIO: A User-Friendly Library of Gradient-Based Methods for Complexity-Aware DNN Optimization
Pagliari, D. J., Risso, M., Motetti, B. A., & Burrello, A. (2023) · Forum on specification and Design Languages · doi:10.1109/fdl59689.2023.10272045 · Cited by 18
“…NAS in ii) is analogous to structured pruning [12], PLiNIO supports three of the most common complexity-driven DNN optimizations in the state-of-the-art, i.e., Quantization, Pruning and NAS [2], [3] under a unified API thus enabling a complete Design Space Exploration (DSE) for DNN workloads.”
BackgroundAttention-Driven Transfer Learning Model for Improved IoT Intrusion Detection
Abdelhamid, S., Hegazy, I., Aref, M. M., & Roushdy, M. I. (2024) · Big Data and Cognitive Computing · doi:10.3390/bdcc8090116 · Cited by 21
Huang, Y., Xu, G., Song, X., & Xu, Y. (2024) · IEEE Transactions on Services Computing · doi:10.1109/tsc.2024.3433534 · Cited by 15
Synthetic data generation method for data-free knowledge distillation in regression neural networks
Zhou, T., & Chiam, K. (2023) · Expert Systems with Applications · doi:10.1016/j.eswa.2023.120327 · Cited by 19
Designing Efficient DNNs via Hardware-Aware Neural Architecture Search and Beyond
Luo, X., Liu, D., Huai, S., Kong, H., Chen, H., & Liu, W. (2021) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2021.3100249 · Cited by 17
Self-citationTransparency and Accountability
Pappachan, P., Moslehpour, M., Bansal, R., & Rahaman, M. (2024) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3693-3860-5.ch006 · Cited by 15
Raza, S. M., Abidi, S. M. H., Masuduzzaman, M., & Shin, S. Y. (2025) · Neurocomputing · doi:10.1016/j.neucom.2025.131357 · Cited by 13
Adaptive Deep Learning Architectures for Real-Time Data Streams in Edge Computing Environments
Sıcakyüz, Ç., Almufti, S. M., & Rustamova, N. R. (2024) · Qubahan Techno Journal · doi:10.48161/qtj.v3n2a25 · Cited by 13
“This strategy allows multiple applications to coexist without significant degradation of performance in any one application[16], [17].”
HSCoNAS: Hardware-Software Co-Design of Efficient DNNs via Neural Architecture Search
Luo, X., Liu, D., Huai, S., & Liu, W. (2021) · Design, Automation, and Test in Europe · doi:10.23919/date51398.2021.9473937 · Cited by 14
“…to pursue competitive accuracy, DNNs are evolving deeply with more layers as well as widely with more channels, thereby incurring the computational gap between complicated DNNs and resource-limited hardware like edge devices, which are deemed as the key computing platform for future AI [5], [7].”
Self-citationDwivedi, U. K., Wiwatcharakoses, C., & Sekimoto, Y. (2022) · International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME) · doi:10.1109/iceccme55909.2022.9987855 · Cited by 13
“However, for most of the tasks, AI on Edge devices [10] were used due to their superior on-device computation power for deep learning models.”
BackgroundNext-Word Prediction: A Perspective of Energy-Aware Distributed Inference
Wang, S., Shao, Z., & Lui, J. C. S. (2023) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2023.3310536 · Cited by 12
Girdhar, N., Raj, A., Sharma, D., Singh, V. K., Doucet, A., & Renz, M. (2025) · Soft Computing · doi:10.1007/s00500-025-10854-y · Cited by 18
“…to be light-weight and portable, making them suitable for deployment on resource-constrained devices such as mobile phones, IoT sensors, and edge servers Jubair et al. ( 2022 ); Mora-Cantallops et al. ( 2021 ); Liu et al. ( 2022 ). economical and modest for a wider range of users and applications.”
Recent advances in Machine Learning based Advanced Driver Assistance System applications
Tatar, G., Bayar, S., Çiçek, İ., & Niar, S. (2024) · Microprocessors and Microsystems · doi:10.1016/j.micpro.2024.105101 · Cited by 13
A deep learning detection framework for gas pipeline leaks suitable for low-power edge devices
Shi, J., Xu, Y., & Lyu, P. (2025) · Engineering Applications of Artificial Intelligence · doi:10.1016/j.engappai.2025.113269 · Cited by 10
Busia, P., Minakova, S., Stefanov, T. P., Raffo, L., & Meloni, P. (2021) · IEEE Access · doi:10.1109/access.2021.3115243 · Cited by 8
“However, the ongoing research in the field of Edge AI is exploring alternative, more efficient ways to execute CNNs on heterogeneous edge platforms [7].”
BackgroundA Dynamic Distributed Scheduler for Computing on the Edge
Hu, F., Mehta, K., Mishra, S., et al. (2024) · 2024 International Symposium on Parallel Computing and Distributed Systems (PCDS) · doi:10.1109/pcds61776.2024.10743743 · Cited by 9
Tomić, S. S., Nikolić, J., Perić, Z., & Aleksić, D. (2022) · Mathematical Problems in Engineering · doi:10.1155/2022/1463094 · Cited by 9
“Quantization is significantly beneficial for NN implementation on resource-limited devices as it can enable fitting the whole NN model into an on-chip memory of edge devices so that the high overhead occurred by the off-chip memory access can be mitigated [2].”
BackgroundTAB: Unified and Optimized Ternary, Binary, and Mixed-precision Neural Network Inference on the Edge
Zhu, S., Duong, L. H. K., & Liu, W. (2022) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3508390 · Cited by 10
“So the CNN models for edge devices need special optimization methods to obtain high accuracy and low latency inference with less storage and computation cost [27].”
MethodSelf-citationZhao, K., Chang, Y., Wu, W., Luo, H., Li, Z., He, S., & Guo, D. (2023) · Journal of Real-Time Image Processing · doi:10.1007/s11554-023-01375-8 · Cited by 8
“The number of CNN-based applications executing on edge devices for image processing is increasing [22], driven by the growing demand for edge AI.”
BackgroundVision Transformer based knowledge distillation for fasteners defect detection
Sampath, V., Maurtua, I., Martín, J. J. A., et al. (2022) · International Conference on Electrical, Computer and Energy Technologies (ICECET) · doi:10.1109/icecet55527.2022.9872566 · Cited by 8
“To compress these heavy networks, several techniques such as model pruning [6], model quantization [7], and knowledge distillation [8] have been proposed in the literature.”
MethodEfficient DNN Execution on Intermittently-Powered IoT Devices With Depth-First Inference
Lv, M., & Xu, E. (2022) · IEEE Access · doi:10.1109/access.2022.3203719 · Cited by 8
On-Device Deep Learning: Survey on Techniques Improving Energy Efficiency of DNNs
Boumendil, A., Bechkit, W., & Benatchba, K. (2024) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2024.3430028 · Cited by 9
“They focus on mobile devices and specific tasks as point cloud, video, and NLP. Liu et al. [21] discuss useful techniques for edge intelligence systems, including model compression, hardware-aware NAS, handcrafted models, as well as adaptive models where different channels or layers are activated…”
Wang, X., Wu, Z., Xiao, G., Han, C., & Fang, C. (2025) · Frontiers in Plant Science · doi:10.3389/fpls.2024.1503033 · Cited by 10
OCAP: On-device Class-Aware Pruning for personalized edge DNN models
Ma, Y.-D., Zhi-chao, Z., Liu, D., He, Z., & Zhou, W. (2023) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2023.102956 · Cited by 8
Self-citationLocalized Knowledge Distillation Helps IoT Devices Provide High-performance Visual Services
Yu, J., Ding, C., Li, H., Jin, Y., & Li, Y. (2023) · IEEE International Conference on Web Services (ICWS) · doi:10.1109/icws60048.2023.00036 · Cited by 6
Hu, F., Mehta, K., Mishra, S., et al. (2023) · Proceedings of the IEEE/ACM 16th International Conference on Utility and Cloud Computing · doi:10.1145/3603166.3632546 · Cited by 6
“In recent years, Edge AI systems [15, 16] have emerged as a promising solution for handling the growing demand for AI workloads by bringing computation closer to the data sources and reducing latency [5, 19].”
BackgroundEdgeCompress: Coupling Multidimensional Model Compression and Dynamic Inference for EdgeAI
Kong, H., Liu, D., Huai, S., Luo, X., Subramaniam, R., Makaya, C., Lin, Q., & Liu, W. (2023) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2023.3276938 · Cited by 7
“CNN compression: Computational redundancy widely exists in CNNs [28].”
Self-citationAdvancement of an Integrated Technological Platform for Wildfire Management through Edge Computing
Marić, L., Chandramouli, K., Martinez, J. R., et al. (2023) · International Conference on Smart and Sustainable Technologies (SpliTech) · doi:10.23919/splitech58164.2023.10193111 · Cited by 6
“Computation on the Edge[3] is a very important asset for the fight against wildfires in all its phases.”
BackgroundEntrust: usEr ceNtric plaTform foR continoUS healThcare
D’Aniello, G., Greco, L., Pupo, F., & Savaglio, C. (2024) · International Conference on Distributed Computing in Smart Systems and the Internet of Things (DCOSS-IoT) · doi:10.1109/dcoss-iot61029.2024.00093 · Cited by 6
“In recent years, some techniques and optimization methods have been proposed, ranging from lightweight deep learning models, network compression or efficient neural architecture search [16].”
MethodNikolić, J., Aleksić, D., Perić, Z., & Dincic, M. R. (2021) · Mathematics · doi:10.3390/math9233091 · Cited by 5
Challenges of Network Forensic Investigation in Fog and Edge Computing
Spiekermann, D., & Keller, J. (2023) · Future Internet · doi:10.3390/fi15100342 · Cited by 5
Alkanan, K., Zhong, Y., Luo, X., et al. (2026) · Kuwait Journal of Science · doi:10.1016/j.kjs.2026.100548 · Cited by 5
Self-citationEmotion on the edge: An evaluation of feature representations and machine learning models
Black, J. T., & Shakir, M. Z. (2025) · Natural Language Processing Journal · doi:10.1016/j.nlp.2025.100127 · Cited by 8
Discovering AI-enabled convergences based on BERT and topic network
Kim, J. H., Lee, S. Y., & Lee, W. S. (2023) · KSII Transactions on Internet and Information Systems · doi:10.3837/tiis.2023.03.018 · Cited by 6
Toward the Predictability of Dynamic Real-Time DNN Inference
Pang, W., Jiang, X., Lv, M., Gao, T. F., Liu, D., & Yi, W. (2021) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2021.3120329 · Cited by 5
“…data, some research focus on the tradeoff of accuracy, latency, energy, memory, and security, etc. DNN applications require DNN models to be adaptive to the different Execution variances (input data, system status like low power mode) for guaranteeing real-time constraints or performance [30].”
BackgroundSelf-citationSurvey of Deep Learning Accelerators for Edge and Emerging Computing
Alam, M. S., Yakopcic, C., Wu, Q., Barnell, M., Khan, S., & Taha, T. M. (2024) · doi:10.20944/preprints202407.0025.v1 · Cited by 38
RACENet: Real-time Adaptive Class-aware Early-exit Networks for Edge Devices
Ayyat, M., Osman, M., & Nadeem, T. (2025) · IEEE International Conference on Pervasive Computing and Communications · doi:10.1109/percom64205.2025.00035 · Cited by 5
“Several studies have explored the deployment of DNNs on edge devices, focusing on optimizing models to fit the constrained resources [21].”
Generating Synthetic Time-Series Data on Edge Devices Using Generative Adversarial Networks
Yousuf, M. F., & Mahmud, M. S. (2024) · 2024 International Conference on Computing, Networking and Communications (ICNC) · doi:10.1109/icnc59896.2024.10556140 · Cited by 5
“Machine learning on edge devices is becoming increasingly popular [7].”
BackgroundMalicious Traffic Classification via Edge Intelligence in IIoT
Wang, M., Zhang, B., Zang, X., Wang, K., & Ma, X. (2023) · Mathematics · doi:10.3390/math11183951 · Cited by 5
“where A(t) denotes the total task volume, αL(t) denotes the ratio of tasks processed locally, L refers to the necessary CPU cycles for executing the one-bit task [38], and fL(t) denotes the corresponding CPU-cycle frequency of the user device [39].”
BackgroundOnce For All Skip: Efficient Adaptive Deep Neural Networks
Yang, Y., Liu, D., Fang, H., et al. (2022) · Design, Automation, and Test in Europe · doi:10.23919/date54114.2022.9774567 · Cited by 4
Self-citationStitchingNet and deep transfer learning method for sewing stitch defect detection
Jung, W., Kang, J.-G., Kwon, W., & Kim, H. (2025) · Journal of Computational Design and Engineering · doi:10.1093/jcde/qwaf037 · Cited by 5
Shrinking the giants: Paving the way for TinyAI
Nakhle, F. (2024) · Device · doi:10.1016/j.device.2024.100411 · Cited by 5
Reinforcement Learning for Real-Time Federated Learning for Resource-Constrained Edge Cluster
Rajashekar, K., Paul, S., Karmakar, S., & Sidhanta, S. (2024) · Journal of Network and Systems Management · doi:10.1007/s10922-024-09857-1 · Cited by 4
“To this end, many researchers have proposed machine learning-based approaches to solve different variations of the assignment problem [68, 69].”
Green Edge Intelligence Scheme for Mobile Keyboard Emoji Prediction
Tang, Y., Hou, J., Huang, X., Shao, Z., & Yang, Y. (2023) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2023.3243955 · Cited by 4
“Comparedwiththeabovetwopredictionschemes,apromis-ingalternativeistoconductmobilekeyboardemojiprediction withtheaidofedgecomputing[8],i.e., edge-assisted prediction .”
Optimization and Security of AI Models for Deployment at Edge: A Comprehensive Review
Rahaman, H., Giri, C., Roy, S. K., & Chakrabarti, A. (2025) · IEEE Computer Society Annual Symposium on VLSI · doi:10.1109/isvlsi65124.2025.11130213 · Cited by 5
“Running computationally intensive AI models, such as convolutional neural networks (CNNs) or transformer architectures, on such devices requires innovative techniques to optimize model size, inference speed, and energy consumption [6].”
Edge accelerated reconstruction using sensitivity analysis for single-lens computational imaging
Wang, X., Feng, T., Xing, Y., Zhao, Z., Dun, X., Wang, Z., & Cheng, X. (2025) · Advanced imaging. · doi:10.3788/ai.2025.10003 · Cited by 5
Ghosh, S., Mondal, A. S., & Chatterjee, A. (2025) · Measurement · doi:10.1016/j.measurement.2024.116619 · Cited by 4
Intelligent Drone Design for Precision Cashew Farming
Rajagopal, M. K., Bala Murugan, M. S., Paventhan, A., Rajurkar, C., Anand, R., & Kumar, S. (2024) · International Conference on Control and Robotics Engineering (ICCRE) · doi:10.1109/iccre61448.2024.10589729 · Cited by 4
“There are several critical points to consider when bringing a deep architecture onto a tiny processor [35], [36], [37]: 1.”
BackgroundOnboard AI: Constraints and Limitations
Miller, K., & Lohn, A. J. (2023) · doi:10.51593/2022ca008 · Cited by 7
Efficient On-Device Incremental Learning by Weight Freezing
Wang, Z.-H., He, Z., Fang, H., Huang, Y.-X., Sun, Y., Yang, Y., Zhang, Z., & Лю, Д. (2022) · Asia and South Pacific Design Automation Conference (ASP-DAC) · doi:10.1109/asp-dac52403.2022.9712563 · Cited by 8
“Recently, a new trend is to implement DNNs on resource-constrained systems, like IoT devices and edge devices for ubiquitous AI adoption, such that the system latency can be guaranteed and the risk of confidential data leakage can be minimized [2].”
BackgroundNo Free Lunch: Balancing Learning and Exploitation at the Network Edge
Mason, F., Chiariotti, F., & Zanella, A. (2022) · International Conference on Communications · doi:10.1109/icc45855.2022.9838277 · Cited by 3
“The machine learning research community is starting to become aware of this issue [8], focusing on model compression and lightweight learning techniques to reduce the burden on the edge hardware [9] and considering the cost of learning in the design of neural networks meant to operate on resourceconstrained devices [1…”
BackgroundLuo, X., Liu, D., Kong, H., Huai, S., Chen, H., Li, S., Xiong, G., & Liu, W. (2024) · Asia and South Pacific Design Automation Conference (ASP-DAC) · doi:10.1109/asp-dac58780.2024.10473971 · Cited by 3
“As a result, the increasing network complexity enlarges the computational gap between computation-intensive CNNs and resource-constrained embedded hardware systems [2].”
BackgroundInfluentialSelf-citationInterpretable deep model pruning
Huang, H., & Pao, H.-K. (2025) · Neurocomputing · doi:10.1016/j.neucom.2025.130485 · Cited by 3
Li, C., Liu, J., Liu, Q., Li, Y., Cao, Y., & Li, Y. (2025) · Communications Engineering · doi:10.1038/s44172-025-00535-x · Cited by 5
“Additionally, these methods typically rely on specialized software libraries and hardware support 43 .”
Deep Learning Model Compression Techniques Performance on Edge Devices
Rachmanto, R. D., Nabhaan, A. N. L., & Setyanto, A. (2024) · Matrik Jurnal Manajemen Teknik Informatika dan Rekayasa Komputer · doi:10.30812/matrik.v23i3.3961 · Cited by 3
“Deep learning (DL) advancements have fueled its adoption across diverse sectors [1].”
BackgroundEdge AI-Enabled Smart Parking Systems
Pham‐Quoc, C., Loc, N. T., & Vung, D. M. (2024) · International Conference on Advanced Technologies for Communications · doi:10.1109/atc63255.2024.10908301 · Cited by 3
ChatGPT or the Language Genie: Perspectives and Lived Experiences of Students in TunisiaPreprint
Abdeljaoued, M. (2024) · Research Square · doi:10.21203/rs.3.rs-5205322/v1 · Cited by 3
TsAFN: A two-stage adaptive fusion network for multimodal sentiment analysis
Liu, J., Wang, Y., Yang, J. B., Yu, X., & Zhao, M. (2025) · Applied Intelligence · doi:10.1007/s10489-025-06577-0 · Cited by 3
“It has been widely applied in computer vision, natural language processing [40], speech recognition, and other fields [41].”
Variable-precision neuromorphic state space model for on-edge activity classification
Leto, B., Urgese, G., Macii, E., & Fra, V. (2025) · Future generations computer systems · doi:10.1016/j.future.2025.108193 · Cited by 5
HACScale: Hardware-Aware Compound Scaling for Resource-Efficient DNNs
Kong, H., Лю, Д., Luo, X., Liu, W., & Subramaniam, R. (2022) · Asia and South Pacific Design Automation Conference (ASP-DAC) · doi:10.1109/asp-dac52403.2022.9712593 · Cited by 4
Self-citationHierarchical multi-scale parametric optimization of deep neural networks
Zhang, S., Vassiliadis, V. S., Dorneanu, B., & Arellano‐García, H. (2023) · Applied Intelligence · doi:10.1007/s10489-023-04745-8 · Cited by 2
Metadata
- Authors
- Liu, D., Kong, H., Luo, X., Liu, W., & Subramaniam, R.
- Affiliations
- Nanyang Technological University (SG)
- Hewlett-Packard (United States) (US)
- Corresponding author
- Luo, Xiangzhong
- Venue
- Neurocomputing
- Publisher
- Elsevier BV
- Published
- 2022-05
- Volume
- 485
- Pages
- 297-320
- ISSN
- 0925-2312 (ISSN-L)
- OpenAlex
- W3110617657
- Cited by
- 171
- References
- 278
- Citations by year
- 2026: 17 · 2025: 51 · 2024: 40 · 2023: 31 · 2022: 24
- FWCI
- 8.95
- Type
- journal article
- Language
- en
- Open access
- bronze
- doi.org
- Licence
- Locations
- Neurocomputing (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- DR-NTU (Nanyang Technological University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- Nanyang Technological University
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Keywords
- edge AI, edge computing, deep learning, lightweight models, model compression, neural architecture search, hardware-aware neural architecture search, adaptive deep learning, edge devices
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
