Deep learning for time series classification: a review

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Poor random initializations can have a significant effect on a DNN's performance.

“Finally, we analyze how poor random initializations can have a significant effect on a DNN's performance.”

May combine two claims.

FCN does not hold any pooling nor a regularization operation.

“We can see that FCN does not hold any pooling nor a regularization operation.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Similarly to the FCN model, the layers (except the final one) in the ResNet architecture have an invariant number of parameters across different datasets.

“Similarly to the FCN model, the layers (except the final one) in the ResNet architecture have an invariant number of parameters across different datasets.”

May combine two claims.

ResNet significantly outperforms the other approaches on the 85 univariate UCR/UEA datasets with an average rank of almost 2.

“The ResNet significantly outperforms the other approaches with an average rank of almost 2.”

May combine two claims.

Time-CNN is the smallest model: it contains a very small number of parameters by design with only 18 filters compared to the 512 filters of FCN.

“This suggestion is also supported by the fact that Time-CNN is the smallest model: it contains a very small number of parameters by design with only 18 filters compared to the 512 filters of FCN.”

May combine two claims. The quote only partly states this claim.

ResNet and FCN dominate with not much of a difference when grouped by training set size.

“Again, ResNet and FCN still dominate with not much of a difference.”

May combine two claims.

ResNet is the most accurate DNN of the study with similar runtime to FCN, the second most accurate DNN.

“our empirical study strongly suggests to use ResNet instead of any other deep learning algorithm - it is the most accurate one with similar runtime to FCN (the second most accurate DNN).”

May combine two claims.

Although the datasets used in the experiments are relatively small compared to billions of labeled images, the deepest networks reached competitive accuracies on the UCR/UEA archive benchmark.

“Although the datasets used in our experiments are relatively small compared to the billions of labeled images (such as ImageNet (Russakovsky et al., 2015) and OpenImages (Krasin et al., 2017) challenges), the deepest networks did reach competitive accuracies on the UCR/UEA archive benchmark.”

May combine two claims.

MCNN and t-LeNet architectures yielded very low accuracies with only one win for the Earthquakes dataset.

“The MCNN and t-LeNet architectures yielded very low accuracies with only one win for the Earthquakes dataset.”

May combine two claims.

FCN wins on 36 datasets whereas Encoder wins only on 17.

“FCN wins on 36 datasets whereases Encoder wins only on 17”

May combine two claims.

On DiatomSizeReduction, ResNet and FCN achieved the worst accuracy (30%) while Time-CNN reached the best accuracy (95%).

“ResNet and FCN achieved the worst accuracy (30%) on this dataset while Time-CNN reached the best accuracy (95%).”

May combine two claims.

TWIESN did reach competitive accuracies on several long time series datasets, such as reaching a 96.8% accuracy on Meat whose time series length is equal to 448.

“However, TWIESN did reach competitive accuracies on several long time series datasets such as reaching a 96.8% accuracy on Meat whose time series length is equal to 448.”

May combine two claims.

Overall, on average, ResNet is the best architecture with FCN and Encoder following as second and third respectively.

“Overall, our results show that, on average, ResNet is the best architecture with FCN and Encoder following as second and third respectively.”

May combine two claims.

FCN's learned representation has too many data points close to the decision boundary, whereas ResNet's hidden features enable projecting data points further away from the decision boundary.

“In other words, FCN's learned representation has too many data points close to the decision boundary whereas ResNet's hidden features enables projecting data points further away from the decision boundary.”

• AssumptionRests on an assumption: MDS plots with Euclidean distance on GAP vectors preserve class separability sufficiently to interpret model performance.

May combine two claims.

FCN and ResNet accuracy on Wine are 58.7% and 74.4% respectively.

“which is evident by the relatively low accuracy of both models FCN and ResNet which is equal respectively to 58.7% and 74.4%.”

May combine two claims.

FCN's CAM contains smoother regions with cyan, green and yellow, whereas ResNet's CAM contains more dark red and blue subsequences.

“In other words, FCN's CAM contains much more smoother regions with cyan, green and yellow regions, whereas ResNet's CAM contains more dark red and blue subsequences”

May combine two claims.

ResNet performed very well in general except for the ECG datasets where it was outperformed by FCN.

“ResNet performed very well in general except for the ECG datasets where it was outperformed by FCN.”

May combine two claims.

The weight's initial value can easily decrease the accuracy of FCN whereas ResNet maintained a relatively high accuracy when taking the worst initial weight values.

“In other words, the weight's initial value can easily decrease the accuracy of FCN whereas ResNet maintained a relatively high accuracy when taking the worst initial weight values.”

May combine two claims.

There is small variance between the approaches that replace the GAP layer with an FC dense layer (MCDCNN, CNN), which also showed similar performance to TWIESN and MLP.

“We found small variance between the approaches that replace the GAP layer with an FC dense layer (MCDCNN, CNN) which also showed similar performance to TWIESN and MLP.”

May combine two claims.

Both FCN and ResNet managed to project the GunPoint data into a linearly separable space.

“We note that both deep learning models (FCN and ResNet) managed to project the data from GunPoint into a linearly separable space”

• AssumptionRests on an assumption: MDS plots with Euclidean distance on GAP vectors preserve class separability sufficiently to interpret model performance.

May combine two claims.

DNNs for TSC suffer from a significant decrease (increase) in accuracy when initialized with bad (good) random weights.

“we observed in our experiments, that DNNs for TSC suffer from a significant decrease (increase) in accuracy when initialized with bad (good) random weights.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

DNNs have revolutionized computer vision, especially with deeper architectures such as Residual and Convolutional Neural Networks.

“DNNs have indeed revolutionized the field of computer vision especially with the advent of novel deeper architectures such as Residual and Convolutional Neural Networks.”

• Also stated in: § 1 Introduction · p. 2 · plain

  “Deep Convolutional Neural Networks (CNNs) have revolutionized the field of computer vision (Krizhevsky et al., 2012).”

• cites Krizhevsky et al., 2012Not checked yet

  Not checked in the free run.

May combine two claims.

HIVE-COTE is hugely computationally intensive and impractical to run on a real big data mining problem.

“To achieve its high accuracy, HIVE-COTE becomes hugely computationally intensive and impractical to run on a real big data mining problem (Bagnall et al., 2017).”

• Also stated in: § 1 Introduction · p. 2 · hedged

  “The approach requires training 37 classifiers as well as cross-validating each hyperparameter of these algorithms, which makes the approach infeasible to train in some situations (Lucas et al., 2018).”

• cites Bagnall et al., 2017 sourceNot checked yet

  Not checked in the free run.

• cites Lucas et al., 2018Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Sequential data such as text and audio can be processed with DNNs to reach state-of-the-art performance for document classification and speech recognition.

“Apart from images, sequential data such as text and audio can also be processed with DNNs to reach state-of-the-art performance for document classification and speech recognition.”

• Also stated in: § 1 Introduction · p. 2 · plain

  “Following the success of deep neural networks (DNNs) in computer vision, a huge amount of research proposed several DNN architectures to solve natural language processing (NLP) tasks such as machine translation (Sutskever et al., 2014; Bahdanau et al., 2015), learning word embeddings (Mikolov et al., 2013; Mikolov et al., 2013) and document classification (Le and Mikolov, 2014; Goldberg, 2016). DN”

• cites Sutskever et al., 2014Not checked yet

  Not checked in the free run.

• cites Hinton et al., 2012 sourceNot checked yet

  Not checked in the free run.

• cites Le and Mikolov, 2014Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The paper used a Wilcoxon signed-rank test with Holm's alpha correction for pairwise post-hoc analysis.

“Once we have established that a statistical difference exists within the classifiers' performance, we followed the pairwise post-hoc analysis recommended by Benavoli et al. (2016) where the average rank comparison is replaced by a Wilcoxon signed-rank test (Wilcoxon, 1945) with Holm's alpha correction (Holm, 1979; Garcia and Herrera, 2008).”

• cites Benavoli et al., 2016Not checked yet

  Not checked in the free run.

• cites Wilcoxon, 1945 sourceNot checked yet

  Not checked in the free run.

• cites Holm, 1979Not checked yet

  Not checked in the free run.

May combine two claims.

Deep learning models can handle the curse of dimensionality by leveraging different degrees of smoothness in compositional function and parallel computations of GPUs.

“The multivariate evaluation shows another benefit of deep learning models, which is the ability to handle the curse of dimensionality (Bellman, 2010; Keogh and Mueen, 2017) by leveraging different degrees of smoothness in compositional function (Poggio et al., 2017) as well as the parallel computations of the GPUs (Lu et al., 2015).”

• cites Bellman, 2010 sourceNot checked yet

  Not checked in the free run.

• cites Poggio et al., 2017 sourceNot checked yet

  Not checked in the free run.

• cites Lu et al., 2015Not checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: The UCR/UEA archive datasets are representative of TSC problems.

May combine two claims.

RNNs were rarely applied for time series classification because they are designed mainly to predict an output for each time stamp.

“Apart from time series forecasting, we found that these neural networks were rarely applied for time series classification which is mainly due to three factors: (1) the type of this architecture is designed mainly to predict an output for each element (time stamp) in the time series (L¨ angkvist et al., 2014);”

• cites L¨ angkvist et al., 2014Not checked yet

  Not checked in the free run.

May combine two claims.

RNNs are considered hard to train and parallelize.

“(3) RNNs are considered hard to train and parallelize which led the researchers to avoid using them for computational reasons (Pascanu et al., 2013).”

• cites Pascanu et al., 2013Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

ESNs were designed to mitigate the challenges of RNNs by eliminating the need to compute the gradient for hidden layers, reducing training time and avoiding the vanishing gradient problem.

“They were designed to mitigate the challenges of RNNs by eliminating the need to compute the gradient for the hidden layers which reduces the training time of these neural networks thus avoiding the vanishing gradient problem.”

• cites Gallicchio and Micheli, 2017Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Deep learning has seen very successful applications in the last years.

“This is surprising as deep learning has seen very successful applications in the last years.”

• Also stated in: § 1 Introduction · p. 2 · plain

  “After having established the current state-of-the-art of non deep classifiers for TSC (Bagnall et al., 2017), we discuss the success of Deep Learning (LeCun et al., 2015) in various classification tasks which motivated the recent utilization of deep learning models for TSC (Wang et al., 2017b).”

• cites LeCun et al., 2015 sourceNot checked yet

  Not checked in the free run.

• cites Wang et al., 2017bNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Deep learning approaches for TSC can be separated into two main categories: generative and discriminative models.

“Deep learning approaches for TSC can be separated into two main categories: the generative and the discriminative models (as proposed in L¨ angkvist et al. (2014)).”

• cites L¨ angkvist et al., 2014Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The paper used the Friedman test to reject the null hypothesis when comparing classifiers over multiple datasets.

“As for comparing the classifiers over multiple datasets, we followed the recommendations in Demˇ sar (2006) and used the Friedman test (Friedman, 1940) to reject the null hypothesis.”

• cites Demˇ sar, 2006Not checked yet

  Not checked in the free run.

• cites Friedman, 1940 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Generative models usually exhibit an unsupervised training step that precedes the learning phase of the classifier.

“Generative models usually exhibit an unsupervised training step that precedes the learning phase of the classifier (L¨ angkvist et al., 2014).”

• cites L¨ angkvist et al., 2014Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The decision taken by HIVE-COTE's 37 classifiers cannot be interpreted easily by domain experts.

“Finally, adding to the huge runtime of HIVE-COTE, the decision taken by 37 classifiers cannot be interpreted easily by domain experts, since researchers already struggle with understanding the decisions taken by an individual classifier.”

May combine two claims.

MLP does not have any layer whose number of parameters is invariant across time series of different lengths, which means that the transferability of the network is not trivial: the number of parameters (weights) of the network depends directly on the length of the input time series.

“MLP does not have any layer whose number of parameters is invariant across time series of different lengths (denoted by # invar in Table 1) which means that the transferability of the network is not trivial: the number of parameters (weights) of the network depends directly on the length of the input time series.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The careful design of Encoder's attention mechanism enabled the invariance across all layers which encouraged the authors to implement a transfer learning approach.

“One final note is that the careful design of Encoder's attention mechanism enabled the invariance across all layers which encouraged the authors to implement a transfer learning approach.”

• cites Serr` a et al., 2018Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Residual Networks (ResNets) add linear shortcut connections for the convolutional layers, potentially enhancing the model's accuracy.

“Residual Networks (ResNets) (Wang et al., 2017b; Geng and Luo, 2018) which add linear shortcut connections for the convolutional layers potentially enhancing the model's accuracy (He et al., 2016).”

• cites He et al., 2016Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The paper adopted accuracy to measure model performance on test data.

“To measure the performance of the model on the test data (generalization), we adopted the accuracy measure (similarly to Bagnall et al. (2017)).”

• cites Bagnall et al., 2017 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

ESNs were originally proposed for time series forecasting, but Tanisaro and Heidemann (2016) proposed a variant of ESNs that uses directly the raw input time series and predicts a probability distribution over the class variables.

“Although ESNs were originally proposed for time series forecasting, Tanisaro and Heidemann (2016) proposed a variant of ESNs that uses directly the raw input time series and predicts a probability distribution over the class variables.”

• cites Tanisaro and Heidemann, 2016Not checked yet

  Not checked in the free run.

May combine two claims.

The replacement of the traditional final FC layer with a Global Average Pooling (GAP) layer reduces drastically the number of parameters in a neural network while enabling the use of the CAM that highlights which parts of the input time series contributed the most to a certain classification.

“In addition, one of the main characteristics of this architecture is the replacement of the traditional final FC layer with a Global Average Pooling (GAP) layer which reduces drastically the number of parameters in a neural network while enabling the use of the CAM (Zhou et al., 2016) that highlights which parts of the input time series contributed the most to a certain classification.”

• cites Zhou et al., 2016Not checked yet

  Not checked in the free run.

May combine two claims.

The latter approach (pre-train then fine-tune) reached higher accuracy thus benefiting from the transfer learning technique.

“The latter approach reached higher accuracy thus benefiting from the transfer learning technique.”

• cites Serr` a et al., 2018Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

End-to-end deep learning aims to incorporate the feature learning process while fine-tuning the discriminative classifier.

“In contrast to feature engineering, end-to-end deep learning aims to incorporate the feature learning process while fine-tuning the discriminative classifier (Nweke et al., 2018).”

• cites Nweke et al., 2018 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The popularity of CNNs for TSC is probably due to their robustness and the relatively small amount of training time compared to complex architectures such as RNNs or MLPs.

“which is probably due to their robustness and the relatively small amount of training time compared to complex architectures such as RNNs or MLPs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

For memory usage limitations over a single GPU, we left the MTS dataset Performance Measurement System (PeMS) out of our experimentations.

“For memory usage limitations over a single GPU, we left the MTS dataset Performance Measurement System (PeMS) out of our experimentations.”

May combine two claims.

Unlike ResNet and FCN, t-LeNet does not have much invariant layers (2 out of 4) due to the use of an FC layer instead of a GAP layer, thus increasing drastically the number of parameters needed to be trained which also depends on the length of the input time series.

“Unlike ResNet and FCN, this approach does not have much invariant layers (2 out of 4) due to the use of an FC layer instead of a GAP layer, thus increasing drastically the number of parameters needed to be trained which also depends on the length of the input time series.”

Classified as probably not a substantial claim.

The accuracy of deep learning methods in human activity recognition depends highly on the quality of the extracted features.

“This has also been the consensus in the human activity recognition literature, where the accuracy of deep learning methods depends highly on the quality of the extracted features (Nweke et al., 2018).”

• cites Nweke et al., 2018 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

We also evaluated all deep learning models on Baydogan's archive that contains 13 MTS classification datasets.

“We also evaluated all deep learning models on Baydogan's archive (Baydogan, 2015) that contains 13 MTS classification datasets.”

• cites Baydogan, 2015Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Any deep learning algorithm can benefit from modifications for solving imbalanced TSC tasks.

“However, any deep learning algorithm can benefit from this type of modification.”

Probably not a substantial claim: may not pass S3 (load-bearing).

We can easily pre-train a model on a source dataset, then transfer and fine-tune it on a target dataset without having to modify the hidden layers of the network.

“That being said, we can easily pre-train a model on a source dataset, then transfer and fine-tune it on a target dataset without having to modify the hidden layers of the network.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

This archive also exhibits datasets with different characteristics such as the length of the time series which, unlike the UCR/UEA archive, varies among the same dataset. This is due to the fact that the datasets in the UCR/UEA archive are already re-scaled to have an equal length among one dataset.

“This archive also exhibits datasets with different characteristics such as the length of the time series which, unlike the UCR/UEA archive, varies among the same dataset. This is due to the fact that the datasets in the UCR/UEA archive are already re-scaled to have an equal length among one dataset (Bagnall et al., 2017).”

• cites Bagnall et al., 2017 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

FCNs are mainly convolutional networks that do not contain any local pooling layers, which means that the length of a time series is kept unchanged throughout the convolutions.

“FCNs are mainly convolutional networks that do not contain any local pooling layers which means that the length of a time series is kept unchanged throughout the convolutions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Cui et al. (2016) were the first to introduce the Window Slicing (WS) method as a data augmentation technique.

“Cui et al. (2016) were the first to introduce the Window Slicing (WS) method as a data augmentation technique.”

• cites Cui et al., 2016Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A linear shortcut is added to link the output of a residual block to its input thus enabling the flow of the gradient directly through these connections, which makes training a DNN much easier by reducing the vanishing gradient effect.

“Actually, the difference with the usual convolutions (such as in FCNs) is that a linear shortcut is added to link the output of a residual block to its input thus enabling the flow of the gradient directly through these connections, which makes training a DNN much easier by reducing the vanishing gradient effect (He et al., 2016).”

• cites He et al., 2016Not checked yet

  Not checked in the free run.

May combine two claims.

The problem with an MLP approach is that temporal information is lost and the features learned are no longer time-invariant.

“The problem with an MLP approach is that temporal information is lost and the features learned are no longer time-invariant.”

May combine two claims.

The main goal of deep learning approaches is to remove the bias due to manually designed features, thus enabling the network to learn the most discriminant useful features for the classification task.

“We made this choice because we believe that the main goal of deep learning approaches is to remove the bias due to manually designed features (Ord´ o¨ nez and Roggen, 2016), thus enabling the network to learn the most discriminant useful features for the classification task.”

• cites Ord´ o¨ nez and Roggen, 2016Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The length of a time series is not an issue for time series classification problems.

“This form of pre-processing has also been used by Ratanamahatana and Keogh (2005) to show that the length of a time series is not an issue for TSC problems.”

• cites Ratanamahatana and Keogh (2005)Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

By using an FC layer before the softmax classifier, the transferability of MCDCNN is limited to the first and second convolutional layers.

“By using an FC layer before the softmax classifier, the transferability of this network is limited to the first and second convolutional layers.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Time-CNN uses mean squared error (MSE) instead of the traditional categorical cross-entropy loss function, and the final layer is a traditional FC layer with sigmoid as the activation function, which does not guarantee a sum of probabilities equal to 1.

“The first characteristic of Time-CNN is the use of the mean squared error (MSE) instead of the traditional categorical cross-entropy loss function, which has been used by all the deep learning approaches we have mentioned so far. Hence, instead of a softmax classifier, the final layer is a traditional FC layer with sigmoid as the activation function, which does not guarantee a sum of probabilities”

Probably not a substantial claim: may not pass S3 (load-bearing).

Unlike GAP, local pooling introduces invariance to small perturbations in the activation map by taking the maximum value in a local pooling window.

“Unlike GAP, local pooling introduces invariance to small perturbations in the activation map (the result of a convolution) by taking the maximum value in a local pooling window.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Since the authors did not train nor test any of the state-of-the-art non-deep learning algorithms, it is possible that they allowed much more training time for the described DNNs.

“Since we did not train nor test any of the state-of-the-art non deep learning algorithms, it is possible that we allowed much more training time for the described DNNs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Deeper neural networks are much more successful than shallower architectures in computer vision.

“our findings are in agreement with the deep learning for computer vision literature where deeper neural networks are much more successful than shallower architectures (He et al., 2016).”

• cites He et al., 2016Not checked yet

  Not checked in the free run.

May combine two claims.

DiatomSizeReduction is the smallest dataset in the UCR/UEA archive with 16 training instances, which suggests that ResNet and FCN are easily overfitting this dataset.

“Interestingly, DiatomSizeReduction is the smallest datasets in the UCR/UEA archive (with 16 training instances), which suggests that ResNet and FCN are easily overfitting this dataset.”

Classified as probably not a substantial claim. May combine two claims.

Linear interpolation preprocessing is very important for deep learning models whose architecture depends on input time series length, such as MLP, and for parallel computation over GPUs.

“This step is very important for deep learning models whose architecture depends on the length of the input time series (such as a MLP) and for parallel computation over the GPUs.”

• AssumptionRests on an assumption: The MTS archive contains time series of unequal length.

May combine two claims.

HIVE-COTE's classification time is bounded by a linear scan of the training set due to employing a nearest neighbor classifier, whereas the trivial GPU parallelization of DNNs provides instant classification.

“In addition to the huge training time, HIVE-COTE's classification time is bounded by a linear scan of the training set due to employing a nearest neighbor classifier, whereas the trivial GPU parallelization of DNNs provides instant classification.”

May combine two claims.

Because increasing the number of convolutional layers increases the path length viewed by the CNN model, ResNet and FCN managed to outperform other approaches whose filter length is longer, such as Encoder.

“since increasing the number of convolutional layers will increase the path length viewed by the CNN model (Vaswani et al., 2017), ResNet and FCN managed to outperform other approaches whose filter length is longer (21) such as Encoder.”

• cites Vaswani et al., 2017Not checked yet

  Not checked in the free run.

May combine two claims.

The FC layer reduces the effect of learning time invariant features, which explains why MLP, Time-CNN and MCDCNN exhibit very similar performance.

“This FC layer reduces the effect of learning time invariant features which explains why MLP, Time-CNN and MCDCNN exhibit very similar performance.”

Probably not a substantial claim: may not pass S2 (checkable).

Allowing a much more thorough search of DTW's warping window would lead to a fairer comparison between deep learning approaches and other state-of-the-art TSC algorithms.

“Therefore, we believe that allowing a much more thorough search of DTW's warping window would lead to a fairer comparison between deep learning approaches and other state-of-the-art TSC algorithms.”

May combine two claims.

ResNet's hyperparameters were not tuned for each dataset; rather, the same architecture was used for the whole benchmark.

“unlike HIVE-COTE, ResNet's hyperparameters were not tuned for each dataset but rather the same architecture was used for the whole benchmark”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

In order to achieve high accuracies while training a DNN, a large training set is needed.

“One final observation that is in agreement with the deep learning literature is that in order to achieve high accuracies while training a DNN, a large training set is needed.”

May combine two claims.

NN-DTW-WW and EE are no longer competitive with current state-of-the-art algorithms for TSC.

“suggesting that these methods are no longer competitive with current state-of-the-art algorithms for TSC.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Further investigation of ResNet's hyperparameters should improve DNNs' accuracy for TSC.

“suggesting further investigation of these hyperparameters should improve DNNs' accuracy for TSC.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Discovering patterns in a one-dimensional space (time) should be an easier task for CNNs than in two-dimensional space, thus requiring less data to learn from.

“it is only natural to think that discovering patterns in a one dimensional space (time) should be an easier task for CNNs thus requiring less data to learn from.”

Classified as probably not a substantial claim.

Taking the mean accuracy over 10 runs with different random weight initializations reduces the bias due to the weights' initial values.

“which enables us to take the mean accuracy over the 10 runs in order to reduce the bias due to the weights' initial values.”

May combine two claims.

For a lazy machine learning algorithm such as NN-DTW, training time is zero when allowing maximum warping, whereas judicially setting the warping window can lead to a significant increase in accuracy.

“for a lazy machine learning algorithm such as NN-DTW, training time is zero when allowing maximum warping whereas it has been shown that judicially setting the warping window Dau et al. (2017) can lead to a significant increase in accuracy.”

• cites Dau et al., 2017Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

There is a need to have an equally large MTS classification archive in order to evaluate hybrid univariate/multivariate time series classifiers.

“These observations reinforces the need to have an equally large MTS classification archive in order to evaluate hybrid univariate/multivariate time series classifiers.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The fact that ResNet is not significantly different than COTE suggests that more datasets would give a better insight into these performances.

“but the fact that ResNet is not significantly different than COTE suggests that more datasets would give a better insight into these performances (Demˇ sar, 2006).”

• cites Demˇ sar, 2006Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Cliques formed by the Wilcoxon Signed Rank Test with Holm's alpha correction do not necessarily reflect the rank order.

“It is worthwhile noting that cliques formed by the Wilcoxon Signed Rank Test with Holm's alpha correction do not necessary reflect the rank order (Lines et al., 2018).”

• cites Lines et al., 2018 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Time-CNN and MCDCNN are the only architectures originally proposed for MTS data.

“Although Time-CNN and MCDCNN are the only architectures originally proposed for MTS data”

Probably not a substantial claim: may not pass S2 (checkable).

The success of ResNet is highly due to its deep flexible architecture.

“We believe that the success of ResNet is highly due to its deep flexible architecture.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Normalization is known to have a huge effect on DNNs' learning capabilities.

“especially since normalization is known to have a huge effect on DNNs' learning capabilities (Zhang et al., 2017).”

• cites Zhang et al., 2017Not checked yet

  Not checked in the free run.

The quote only partly states this claim.

The difference in accuracy between ResNet and FCN on Meat is probably explained by ResNet's ability to filter out non-discriminative and discriminative regions with higher confidence.

“which probably explains why FCN is less accurate than ResNet on this dataset.”

• AssumptionRests on an assumption: CAM visualizations faithfully indicate which time series regions contribute to class identification.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Most DNN architectures, with categorical cross-entropy as their cost function, employ mainly the same classifier: softmax, which is basically designed for multi-class classification.

“In fact, most DNNs architectures, with the categorical cross-entropy as their cost function, employ mainly the same classifier: softmax which is basically designed for multi-class classification.”

May combine two claims.

The shapelet transformation classifier (Hills et al., 2014) is an ensemble approach, which makes unclear how the shapelets affect the decision taken by the individual classifiers, whereas for an end-to-end deep learning model the classification can be directly explained by using the Class Activation Map.

“we stress that the shapelet transformation classifier (Hills et al., 2014) is an ensemble approach, which makes unclear how the shapelets affect the decision taken by the individual classifiers whereas for an end-to-end deep learning model we can directly explain the classification by using the Class Activation Map.”

• cites Hills et al., 2014 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The part where the actor raises his/her arm seems to be also a discriminative part of the data, which suggests that the deep learning algorithms are identifying more 'shapelets'.

“On the other hand, the part where the actor raises his/her arm seems to be also a discriminative part of the data which suggests that the deep learning algorithms are identifying more 'shapelets'.”

May combine two claims.

CNNs have the ability to localize a given discriminative shape regardless where it appears in the time series, which is evidence for CNNs' capability of learning time-invariant warped features.

“Finally, another interesting observation is the ability of CNNs to localize a given discriminative shape regardless where it appears in the time series, which is evidence for CNNs' capability of learning time-invariant warped features.”

Probably not a substantial claim: may not pass S3 (load-bearing).

This observation cannot confirm which classifier extracted the most discriminative subsequences especially because all algorithms achieved similar accuracy on the GunPoint dataset.

“We should note that this observation cannot confirm which classifier extracted the most discriminative subsequences especially because all algorithms achieved similar accuracy on GunPoint dataset.”