Source-linked AI summary
Cognitive Psychology for Deep Neural Networks: A Shape Bias Case Study
Samuel Ritter, David G. T. Barrett, Adam Santoro, Matt M. Botvinick
TL;DR
DNNs perform well but remain difficult to understand, motivating methods that reveal the properties underlying their solutions. This paper adapts cognitive-psychology theories and experiments to probe one-shot learning models, finding human-like shape bias alongside substantial variation across models. The results support behavioral probes as tools for exposing DNN properties and provide a candidate computational account of human one-shot word learning.
Problem
DNN solutions are poorly understood, creating a need for methods that reveal the predictive properties underlying their behavior.
Method
The paper applies cognitive psychology’s hypothesis-driven theories and probe-dataset experiments to Matching Networks and Inception models performing one-shot word learning.
Results
Inception networks and Matching Networks exhibit strong shape bias, while identical architectures with different seeds can converge to qualitatively different biases.
Takeaways & Limitations
Cognitive-psychology behavioral probes can expose hidden computational properties of DNNs and offer a candidate computational account of human one-shot word learning.
Takeaways & Limitations
Behavioral probes reveal computational-level behavior but do not elucidate the neural mechanisms producing shape bias.
Abstract
from arXiv · showhide
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.
1. Introduction
The paper proposes using cognitive psychology’s theories and experimental tools to make DNN solutions more interpretable. A shape-bias case study applies developmental-psychology methods to one-shot learning models.
- DNNs achieve strong performance across complex tasks, but their solutions remain poorly understood and are often treated as black boxes.This opacity obstructs both research aimed at improving models and real-world applications.
- The authors propose adding psychology-inspired methods to existing approaches for understanding DNNs.Their proof-of-concept study examines Matching Networks and an Inception Baseline on one-shot word learning.
- Behavioral probes from cognitive psychology provide a complementary way to investigate otherwise opaque operations in DNNs.The paper positions these probes alongside existing neuroscience-inspired and interpretability methods.
- Developmental-psychology research finds that children tend to assign the same word to similarly shaped objects rather than objects sharing color, texture, or size.The study tests whether DNNs exhibit this human-like shape bias using stimuli and procedures inspired by the original experiments.
- Inception networks trained on ImageNet display a strong shape bias, while differently seeded but architecturally identical networks vary substantially in that bias.Matching Networks also show a strong shape bias that closely mimics the bias of their input Inception model.
2. Inductive Biases, Statistical Learners and Probe Datasets
The paper frames DNN interpretability as identifying human-describable properties that trained models use predictively. It proposes hypothesis-driven probe experiments to test those candidate inductive biases.
- A DNN develops inductive biases by discovering input properties that predict labels while minimizing a loss function.These properties can include object shape, tumor-node count, or patterns in medical measurements.
- Real-world predictive properties are often complex feature conjunctions that are difficult for humans to describe intuitively.Intermediate-layer representations may exist numerically but remain too arcane to interpret.
- The proposed workflow begins by hypothesizing a human-evaluable property hp(x) that the model may use.Candidate properties should be relevant to the task, such as object shape or tumor-node count.
- Researchers then derive behavioral predictions, construct probe examples spanning the property’s range, and test whether model behavior supports the hypothesis.The experiment can support or fail to support the proposed bias.
- The study applies a theory–experiment pair from cognitive psychology to state-of-the-art one-shot learning models.This approach transfers psychological tools for studying human cognition to DNN analysis.
3. The problem of word learning; the solution of inductive biases
One-shot word learning is difficult because a single example supports many possible word meanings. Developmental psychologists address this ambiguity by proposing cognitive biases that eliminate unlikely hypotheses, which the paper tests in DNNs.
- Quine’s word-learning problem highlights that a novel label for one object permits an enormous number of possible interpretations.Possible meanings range from the object’s category to its color, the individual object, or its parts.
- Developmental psychologists proposed biases that narrow the hypothesis space, including whole-object, taxonomic, mutual-exclusivity, and shape biases.These biases were tested by asking whether a novel name applies to other objects.
- The resulting computational-level account describes word learning as using biases to eliminate unlikely hypotheses when inferring new-word meanings.Association-learning and Bayesian approaches are identified as contrasting or complementary theories.
- The paper leaves association-learning and Bayesian theories for future work and focuses on hypothesis-elimination theory and methodology.This is a stated scope choice for applying psychological analysis to deep-learning models.
- The experiments ask which predictive properties DNNs use, whether models share them, whether they are human-interpretable, and how they change during training.Analogous experiments test shape bias in Matching Networks, Inception networks, and different instances of each architecture.
4. One-shot word learning models and training
The paper evaluates one-shot word-learning models that assign a novel probe an existing support-set label after a single example, using Inception-based and Matching Nets architectures.
- One-shot word learning task: One-shot learning assigns a novel probe image the correct label from support-set labels after only a single example.The probe is unlabelled, while the support set contains images and associated labels.
- Inception baseline: The Inception Baseline labels each probe with the nearest support-set neighbour using cosine distance between last-layer features.Its Inception features support 87.6% one-shot classification accuracy on ImageNet.
- Matching Nets: Matching Nets use attention over context-dependent probe and support embeddings to assign labels in a fully differentiable architecture.The embeddings are produced with LSTM components operating on Inception-derived features and the support set.
- Matching Nets training: Matching Nets are trained episodically by updating probe-classification accuracy and randomly permuting class labels after each update.The permutation prevents learning a fixed mapping between categories and labels, targeting classification from a single in-class support example.
- Evaluation: The trained Inception and Matching Nets models are then tested on probe datasets designed to reveal shape bias.The paper introduces the probe datasets after describing the ImageNet training procedure.
5. Data for bias discovery
The study uses controlled cognitive-psychology stimuli and a real-world probe set, arranging images in triples that contrast shape and colour matches to measure shape bias.
- Cognitive Psychology Dataset: The CogPsyc dataset contains 150 object images arranged as 50 triples across 5 backgrounds.Each triple includes a probe, a shape-match image, and a colour-match image.
- Cognitive Psychology Dataset: Each triple contrasts an image matching the probe in colour but not shape with one matching shape but not colour.This arrangement isolates whether models favour shape or colour when assigning the probe’s class.
- Cognitive Psychology Dataset: The controlled stimuli were supplied by Linda Smith and regulate object size and background content.The photographs were previously used in shape-bias experiments at Indiana University’s Cognitive Development Lab.
- Real-world probe dataset: The real-world dataset contains 90 Google Image Search photographs organized into 30 probe, shape-match, and colour-match triples.Probe objects were selected to be unlikely to appear in standard datasets such as ImageNet.
- Real-world probe dataset: A silver tuning fork, silver guitar capo, and black tuning fork illustrate colour-match and shape-match alternatives within one real-world triple.The colour match shares silver colour with the probe, whereas the shape match shares the tuning-fork shape.
- Real-world probe dataset: The real-world dataset was collected in roughly two days to strengthen confidence in CogPsyc results and demonstrate that probe datasets are easy to construct.Its bias-over-time pattern matched the established CogPsych dataset.
6. Results
ImageNet-trained Inception and Matching Networks show strong shape bias, but its magnitude varies across initializations and during training independently of near-identical classification accuracy.
- Bs = 0.68 for Inception on CogPsyc and Bs = 0.97 on the real-world dataset, indicating stronger shape than colour preference.
- Matching Networks achieved Bs = 0.7 on CogPsyc and Bs = 1 on the real-world dataset, inheriting bias from their fixed Inception inputs rather than changing during training.
- Figure 2 shows strong shape bias across 15 Inception seeds, within-model fluctuations up to three standard deviations, and higher variability at training’s start and end.
- All 15 Inception seeds achieved near-identical test accuracy during ImageNet training, despite differing shape-bias trajectories.
- Bs = 0.628 ± 0.049 on CogPsyc and Bs = 0.958 ± 0.037 on the real-world dataset at training’s end, showing initialization-dependent variability.
- ρ = 0.15 for CogPsyc and ρ = −0.06 for the real-world dataset, so bias fluctuations cannot be accounted for by classification-accuracy fluctuations.
7. Discussion
The shape-bias case study reveals that DNNs can inherit, vary, and propagate behavioral biases, while cognitive-psychology probes support both model interpretation and hypotheses about human word learning.
- A shape bias case study: Pre-trained ImageNet Inception and memory networks show a strong shape bias, warranting caution when applications depend on color features.The paper specifically highlights one-shot fruit classification as a setting where color may matter more than shape.
- A shape bias case study: Architecturally identical models can differ substantially in shape bias despite almost identical classification accuracy.The variability depends on random initialization and can also occur within models during training.
- A shape bias case study: Shape bias propagates from Inception to memory modules, so combining components can transmit unknown properties across the system.The memory network retains the downstream Inception network’s bias statistics.
- A shape bias case study: Unwanted biases may be addressed through architecture or training-data changes, or through post-hoc model selection using early stopping or initialization seeds.These alternatives are presented for cases where direct architectural or data manipulation is impractical.
- A shape bias case study: Behavioral probes expose computational-level properties but do not elucidate the neural mechanisms producing them.The authors suggest combining behavioral analysis with gradient-based visualization or neuron ablation.
- Modelling human word learning: Memory networks provide a computational-level account of human one-shot word learning by scaling to real-world images and reproducing hallmark findings.The model’s early training increase in shape bias parallels developmental differences observed in humans.
- Modelling human word learning: The model predicts that human shape bias varies across subjects and during development, without correlating with adult-level one-shot learning ability.These are proposed as testable predictions for human word-learning research.
- Cognitive Psychology for Deep Neural Networks: Applying cognitive-psychology techniques to DNNs opens a broader program for probing artificial cognitive properties and testing more human word-learning hypotheses.The paper notes that model probing can be faster than human behavioral experiments.
8. Conclusion
The paper demonstrates that cognitive-psychology techniques can reveal previously unknown properties of poorly understood DNNs. Its case study examines shape bias in Inception and memory networks and motivates further exploration using psychological methods.
- 8. Conclusion: Cognitive-psychology techniques revealed previously unknown properties of Inception and memory networks through a shape-bias case study.The conclusion presents this as a model for future exploration of DNNs using cognitive-psychology methods.