Evaluating Neural Network Representations Against Human Cognition

Explore the intersection of cognitive science and machine learning in this UC Berkeley lecture by Tom Griffiths on evaluating neural network representations against human cognition. Delve into key concepts like categorization, family resemblance, and psychological spaces. Examine the differences between prototypes and exemplars, and learn about multidimensional scaling techniques. Investigate the limitations of spatial representations and the importance of features in cognitive models. Analyze convolutional neural networks and their performance in predicting human similarity judgments. Discover how hierarchical clustering can provide insights into human representations of categories. Compare vector space models with human word associations and explore the challenges in modeling relational similarity. Gain a deeper understanding of the complexities involved in bridging artificial and human cognition through this comprehensive exploration of representation learning.

Intro
Categorization
Family resemblance
Prototypes vs. exemplars
Psychological spaces
Multidimensional scaling
Features not spaces
Violation of triangle inequality
Objectives
Convolutional neural networks
Collecting similarity judgments
Prediction performance
Hierarchical clustering
A basis for human representations
Birds vs. Planes
Category representations
Vector space models
Objections to spatial representations
Word association
The inadequacy of the cosine
Testing the parallelogram model
Testing on a wide range of relations
Human relational similarity judgments
Asymmetry in analogies
Violations of the triangle inequality
Conclusions