Cognitive Psychology for Deep Neural Networks: A Shape Bias Case Study

By Samuel Ritter, David G. T. Barrett, Adam Santoro, Matthew Botvinick. 2017

207 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Safety, Robustness & Interpretability

Citations in the corpus

ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness (2018, background): “Ritter et al. (2017) discovered that CNNs develop a so-called “shape bias” just like children, i.e. that object shape is more important than colour for object classiﬁcation (although see Hosseini et al. (2018) for contrary evidence).”
This Looks Like That, Because ... Explaining Prototypes for Interpretable Image Recognition (2020, background): “It has been shown that CNNs trained on ImageNet are strongly biased towards recognizing texture [16], although other work shows that CNNs can be biased towards shape [33] or colour [20].”
Few-Shot Learning with Metric-Agnostic Conditional Embeddings (2018, methodology): “While children, and some neural networks, share a bias to classify based on shape (Landau et al., 1988; Ritter et al., 2017), we can also recognize the potential that some classiﬁcations would require a re-weighting of attribute importance as in the classiﬁcation of non-solids where shape is less…”
Evaluating Compositionality in Sentence Embeddings (2018, background): “We see a role for cognitive science in designing benchmarks that better probe the competences of AI systems, much in the same way that cognitive scientists have been probing the competences of humans (Ritter et al., 2017; Lake et al., 2018).”
ConvNets and ImageNet Beyond Accuracy: Explanations, Bias Detection, Adversarial Examples and Model Criticism (2017, background): “Concomitantly, several studies have demonstrated the lack of robustness of deep neural networks to adversarial examples [42, 16, 7] and raised questions about their tendency to exhibit (undesirable) biases [37].”
Semantic Adversarial Examples (2018, methodology): “In fact, models trained on ImageNet are shown to display shape bias as well [12].”
Visual Analytics in Deep Learning: An Interrogative Survey for the Next Frontiers (2018, unknown): “Other domains have recognized that interpretable deep learning research may require evaluation techniques for their interpretations, and argue that there is a large body of work from fields such as philosophy, cognitive science, and social psychology that could be utilized [62], [111].”
Assessing Shape Bias Property of Convolutional Neural Networks (2018, unknown): “In [3], the authors studied shape bias property in CNNs by performing experiments on small datasets, in which images were arranged in triples of a probe, a shape-match and a color-match.”
Using cognitive psychology to understand GPT-3 (2022, methodology): “Finally, methods from cognitive psychology have also been applied to understand deep learning models’ behavior more generally (51).”
Understanding Grounded Language Learning Agents (2017, methodology): “These conclusions can be incorporated with those of Ritter et al. (2017), who observe a shape bias in convolutional networks trained on the ImageNet Challenge training set.”
Understanding Deep Learning Performance through an Examination of Test Set Difficulty: A Psychometric Case Study (2017, methodology): “One method for interpreting deep neural networks (DNNs) is to examine model predictions for spe-ciﬁc input examples, e.g. testing for shape bias as in Ritter et al. (2017).”
Learning Not to Learn: Training Deep Neural Networks With Biased Data (2018, background): “As mentioned in [17], neural networks prefer to categorize images based on shape rather than color.”
ConvNets and ImageNet Beyond Accuracy: Understanding Mistakes and Uncovering Biases (2017, background): “networks to adversarial examples [15,16,17] and raised questions about their tendency to exhibit (undesirable) biases [18].”
The Origins and Prevalence of Texture Bias in Convolutional Neural Networks (2019, background): “Previous studies found preferences for shape over color [77] and perceptual shape over physical shape [58].”
Exploring Corruption Robustness: Inductive Biases in Vision Transformers and MLP-Mixers (2021, background): “Research indicates that humans tend to classify objects based on shape rather than color or texture while convolutional neural networks are more biased towards texture (Ritter et al., 2017).”
REPAIR: Removing Representation Bias by Dataset Resampling (2019, background): “Second, the learning procedure could give rise to biased deep learning algorithms [3, 25].”
Psychlab: A Psychology Laboratory for Deep Reinforcement Learning Agents (2018, background): “Third, several recent studies have also taken the perspective of probing convolutional neural networks along lines motivated by psychology and cognitive science but, unlike our study, mostly have used networks trained by supervised learning (Zoran et al., 2015; Ritter et al., 2017; Volokitin et al.,”
Language models show human-like content effects on reasoning (2022, methodology): “More broadly, we demonstrate the value of using established theories and experimental paradigms from cognitive science to probe the increasingly complex behaviors of machine learning models (cf. Binz and Schulz, 2022; Ritter et al., 2017).”