[Paper Review] Cognitive Psychology for Deep Neural Networks: A Shape Bias Case Study

Samuel Ritter, David G. T. Barrett+2 more|arXiv (Cornell University)|Jun 26, 2017

Explainable Artificial Intelligence (XAI)31 references61 citations

TL;DR

This paper applies cognitive psychology probing to deep neural networks to test for a shape bias in one-shot word learning; Inception and Matching Networks exhibit shape bias, with strong variance across seeds and during training, and bias propagates from input to downstream components.

ABSTRACT

Deep neural networks (DNNs) have achieved unprecedented performance on a wide range of complex tasks, rapidly outpacing our understanding of the nature of their solutions. This has caused a recent surge of interest in methods for rendering modern neural systems more interpretable. In this work, we propose to address the interpretability problem in modern DNNs using the rich history of problem descriptions, theories and experimental methods developed by cognitive psychologists to study the human mind. To explore the potential value of these tools, we chose a well-established analysis from developmental psychology that explains how children learn word labels for objects, and applied that analysis to DNNs. Using datasets of stimuli inspired by the original cognitive psychology experiments, we find that state-of-the-art one shot learning models trained on ImageNet exhibit a similar bias to that observed in humans: they prefer to categorize objects according to shape rather than color. The magnitude of this shape bias varies greatly among architecturally identical, but differently seeded models, and even fluctuates within seeds throughout training, despite nearly equivalent classification performance. These results demonstrate the capability of tools from cognitive psychology for exposing hidden computational properties of DNNs, while concurrently providing us with a computational model for human word learning.

Motivation & Objective

• Motivate interpretable analysis of DNNs by importing cognitive psychology methods and hypotheses.
• Test whether state-of-the-art one-shot learning models exhibit a shape bias similar to humans.
• Examine variability of shape bias across seeds and during training while maintaining high classification accuracy.
• Propose that shape bias may serve as a computational account for human one-shot word learning.

Proposed method

• Adapts cognitive psychology shape-bias experiments to DNNs to create probe datasets (CogPsyc) with shape-, color-, and probe-image triples.
• Evaluates an Inception Baseline (IB) one-shot classifier using pre-trained Inception features with nearest-neighbor classification.
• Uses Matching Networks (MN) with an attention-based embedding and memory module trained on ImageNet for one-shot learning.
• Computes shape bias B_s as the proportion of probes labeled by the shape-match, i.e., B_s = E(δ(ŷ − y_s)).
• Assesses bias across multiple seeds, datasets (CogPsyc and real-world), and training stages to analyze emergence and variability.

Experimental results

Research questions

• RQ1 Do state-of-the-art DNNs trained on ImageNet show a human-like shape bias in one-shot word learning tasks?
• RQ2 How does shape bias vary with initialization seeds and during training?
• RQ3 Is the observed bias consistent across model architectures (Inception vs Matching Networks) and input features?
• RQ4 Does shape bias propagate across model components when chaining models (IB to MN)?

Key findings

• Inception Baseline exhibits a shape bias B_s = 0.68 on CogPsyc data and B_s = 0.97 on real-world data.
• Matching Networks show shape bias B_s = 0.7 on CogPsyc data and B_s = 1 on real-world data.
• Shape bias varies substantially across seeds (IB: end-of-training mean B_s = 0.628, SD 0.049; real-world: mean 0.958, SD 0.037).
• Shape bias emerges early in training, before convergence, for IB models.
• MN inherits the IB bias from its input features and retains it during training (no significant change).
• Bias fluctuations are large within IB during training but not in MN, illustrating bias propagation across modules.