The relationship between perception and inference, as postulated by Helmholtz in the 19th century, is paralleled in modern machine learning by generative models like Variational Autoencoders (VAEs) and their hierarchical variants. Here, we evaluate the role of hierarchical inference and its alignment with brain function in the domain of motion perception. We introduce a novel synthetic data framework, Retinal Optic Flow Learning (ROFL), that enables control over motion statistics and their causes. We then test the performance of unsupervised models on two critical downstream tasks: predicting ground truth variables (e.g., object motion) and predicting the responses of single neurons from area MT of the primate dorsal pathway. Importantly, this framework allows us to manipulate both the architectures of the models and the causal structure of the world they are trained on. We found that a single inductive bias, hierarchical latent structure, yields several improvements. First, it improves the linear decodability of ground truth variables and does so in a sparse and disentangled manner. Second, hierarchical VAEs outperform previous state-of-the-art models in predicting MT neuron responses with a performance gain of over 2x and result in sparse latent-to-neuron relationships. Third, these results depend on the causal structure of the world, indicating that alignment between brains and artificial neural networks depends not only on architecture but on matching the stimulus statistics of the organism. Collectively, these results suggest that the brain engages in hierarchical Bayesian inference to comprehend the state of the world and that this functional organization can be effectively captured by hierarchical VAEs.

—

To request the Zoom link email jteeters@berkeley.edu.  Also indicate if you would like to be added to the mailing list for announcements about Redwood Seminars.