Prof. Drew Linsley | Harmonizing the object recognition strategies of deep neural networks with humans

The past several years of deep learning have revealed that the performance of deep neural networks (DNNs) is largely a function of scale: the number of model parameters and data used for training. At the same time, DNNs have become the standard model for human perception and generating hypotheses on its underlying neural circuitry. Have recent trends in deep learning carried concomitant improvements in explaining human perception? In this talk, I will provide evidence that DNNs have progressively become worse models of human perception as their accuracies on ImageNet have increased due to scale. I will also describe a fix for this problem, the Neural Harmonizer: a simple addition to the standard deep learning toolkit that can significantly improve the alignment of DNNs with human perception. These findings provide an explanation for the reported suboptimality of state-of-the-art DNNs for explaining many facets of biological intelligence — from perceptual data to neural recordings — and offer a simple and "harmonized" way forward.

Biography:

Drew Linsley is an Assistant Professor (Research) in Computational Neuroscience at Brown University. His goal is to build artificial systems that can perceive and reason about the world as effectively as humans, and in the process, understand how biological brains accomplish the same thing. To take steps towards these goals, he focuses on computational problems that are easier for humans to solve than state-of-the-art algorithms for artificial intelligence, and seeks to reverse-engineer the neural algorithms that are responsible for this difference. By pursuing these fundamental problems of visual intelligence, his work spans machine learning, computer vision, computational neuroscience, and cognitive science.