FriSem

Violet Xiang, PhD Student, Department of Psychology, Stanford University, 2021 Fall First-Year Project Neuroscience Student

Title: Modeling human learning with contrastive learning algorithms
Abstract: Deep convolutional neural networks trained on large annotated visual datasets such as ImageNet have become the state-of-the-art models both for visual recognition tasks and for predicting neuronal responses in primates' visual stream. Until recently, the large amount of annotated labels required to train these systems made them implausible as models of real visual development and learning. Substantial progress, however, in deep unsupervised learning using contrastive learning objectives has closed the gap to their supervised counterparts both in performance on visual tasks and adult neural predictivity, in
turn opening the possibility that contrastive training processes themselves might model real visual learning. In this work, we evaluate the effectiveness of contrastive learning algorithms in modeling human learning. We propose a method to compare the change of visual categorization behaviors of humans and models when the same stimuli sequences are presented. We find that all algorithms have imperfections, exhibiting mismatches to human real-time performance changes in the short term. We further show that adding the use of negative samples to a state-of-the-art contrastive algorithm can greatly increase its performance our metric. Taken together, we show that choices in how memory is handled by contrastive algorithms strongly impact their ability to match real human learning, and quantitatively expose an open problem space for improved unsupervised learning algorithms that are more human-like in their learning.

You can find this information on the talk here.