Learning to Complement Humans

Across the tested models, joint training for complementarity performs at least as well as fixed training.

“The joint models, which optimize for complementarity, uniformly outperform or tie their fixed counterparts.”

Humans and machines can make qualitatively different kinds of errors.

“We find that humans and machines may make qualitatively different kinds of errors.”

The performance gap between fixed and joint models increases when error costs are asymmetric.

“The gaps between the fixed and joint models grow under asymmetric costs.”

In two real-world domains, scientific discovery and medical diagnosis, human-machine teams built using the proposed methods achieve better performance than machines or people alone.

“We demonstrate in two real-world domains (scientific discovery and medical diagnosis) that human-machine teams built via these methods outperform the individual performance of machines and people.”

For CAMELYON16, a region defined by two features holds over 68% of human errors while containing only 10% of instances.

“Over 68% of human errors are concentrated in a region containing just 10% of instances, identified using two features.”

This improvement comes at the cost of more errors in the far-left leaf, where the human is almost perfectly accurate.

“This comes at the cost of increased errors in the far-left leaf; however, in this region the human is almost perfectly accurate.”

Under asymmetric costs on CAMELYON16, joint approaches outperform fixed counterparts by up to 10% for discriminative models and 4.8% for VOI models.

“In Figure 2(c) (asymmetric costs), the joint approaches now outperform their fixed counterparts by up to 10% (discriminative family) and 4.8% (VOI).”

Joint training yields larger benefits when model capacity is limited.

“Overall, joint training provides larger benefits with limited model capacity.”

• Also stated in: § 1 Introduction · plain

  “First, training for complementarity is most important when the ML model has limited capacity, forcing it to pick parts of the task to focus on.”

Under joint training, the ML model's errors shift because it focuses more on cases that people find hard.

“Moreover, the errors made by the ML model change under joint training as the model places more emphasis on instances that are difficult for humans.”

The trade-off from joint optimization yields a 2% reduction in overall loss.

“Overall, this tradeoff made by the joint optimization leads to a 2% overall reduction in loss.”

For Galaxy Zoo, and for CAMELYON16 discriminative models, every fixed-versus-joint difference is statistically significant (Student t-test, p < 10^-3).

“All differences between fixed and joint models are statistically significant for Galaxy Zoo, and on the CAMELYON16 task for the discriminative models (Student t-test, p < 10 -3 ).”

The direct relaxation of Equation 1 often produced unstable training.

“In experiments, this direct relaxation often produced unstable training; intuitively, the predictions and query policy may be spiky in some regions, giving a rapidly changing training signal.”

With equal costs in Figure 2(b), the two-layer discriminative and VOI methods performed equally.

“For example, in Figure 2(b) (equal costs), the two-layer model performance of discriminative or VOI approaches were previously tied.”

Fitting the decision tree in Figure 4 reveals clear structure in CAMELYON16 human errors.

“Here, we find clear structure in the human errors, uncovered by fit- ting the decision tree shown in Figure 4 (for the uniform-cost task with two-layer models and 50 hidden units).”

Using the simplification m(x,h)=h yields stronger empirical performance than allowing m(x,h) to output a different prediction.

“However, we observe stronger empirical performance using the simplification m θ (x, h) = h (though training a separate model for m θ (x, h) results in similar qualitative conclusions).”

Complementarity training yields greater gains when error costs are asymmetric.

“Second, training for complementarity has larger benefits when there is an asymmetric cost to errors (e.g., false negatives are more costly than false positives).”

Team-performance emphasis is especially needed for hard tasks that machines cannot fully solve by themselves.

“This suggests that an emphasis on team performance is particularly necessary for difficult tasks that machines cannot perfectly master on their own.”

Joint training makes human and machine errors differ in structured ways that the methods can exploit to improve team performance.

“Via joint training, human and machine errors become different in structured ways that can be leveraged by the methods to improve team performance.”

Complementarity-focused training matters most when the ML model's capacity is small and it must select which task parts to prioritize.

“First, training for complementarity is most important when the ML model has limited capacity, forcing it to pick parts of the task to focus on.”

Raghu et al. (2019) keep the ML model fixed rather than training it for complementarity.

“However, the ML model is always fixed, instead of being trained for complementarity.”

• cites Raghu et al., 2019Not checked yet

  Not checked in the free run.

When different errors must be prioritized, optimizing team utility pays off more.

“The need to prioritize among potential errors increases the returns of optimizing for team utility.”

Predicted probabilities must be calibrated to yield meaningful expected utilities.

“Calibration is necessary for the predicted probabilities to give meaningful expected utilities.”

• cites Platt, 1999Not checked yet

  Not checked in the free run.

• cites Niculescu-Mizil and Caruana, 2005Not checked yet

  Not checked in the free run.

Earlier systems have been built to decide when a human should be consulted.

“Prior work includes systems that determine when to con-sult humans [Horvitz and Paek, 2007;Kamar et al., 2012;Raghu et al., 2019].”

• cites Horvitz and Paek, 2007Not checked yet

  Not checked in the free run.

• cites Kamar et al., 2012Not checked yet

  Not checked in the free run.

• cites Raghu et al., 2019Not checked yet

  Not checked in the free run.

The joint model's error distribution shifts to complement the human's strengths and weaknesses.

“In other words, the distribution of errors incurred by the joint model shifts to complement the strengths and weaknesses of the human.”

Earlier predictive models were optimized for their own isolated accuracy, not for using the complementary strengths of people and machines.

“However, the predictive models are still trained to maximize their own, solitary performance, rather than to leverage the distinctive strengths of machines and humans.”

• cites Horvitz and Paek, 2007Not checked yet

  Not checked in the free run.

• cites Kamar et al., 2012Not checked yet

  Not checked in the free run.

• cites Raghu et al., 2019Not checked yet

  Not checked in the free run.

To the authors' knowledge, these methods are the first to jointly train ML systems and task-allocation policies so as to optimize human-AI teams.

“To our knowledge, the methods we present are the first to optimize human-AI teams by jointly training ML systems together with policies for allocating tasks to human experts versus machines.”

Focusing on team performance is particularly useful when trade-offs among different errors must be made.

“Optimizing combined team performance is especially helpful when it is necessary to prioritize between potential errors.”

Keeping the fixed VOI structure provides an inductive bias: the model begins with sound probabilistic reasoning and is then tuned for complementarity.

“Retaining the structure of the fixed VOI system can be viewed as an inductive bias which allows the model to start from wellfounded probabilistic reasoning and then to be fine-tuned for complementarity.”

Differences in error rates on classes 2 and 3 and in query behavior suggest that joint optimization alters the ML system's learning and decisions.

“Both the error rates of the two approaches on classes 2 and 3, and the way they query humans show differences, indicating that joint optimization changes how the ML system learns and makes decisions.”

Models with less capacity carry more bias, as their simpler hypotheses cannot match the ground truth as closely.

“Essentially, a lower-capacity model has more potential bias (since it represents less complex hypotheses which cannot fit the ground truth as closely).”