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 first introduce a novel synthetic data framework, Retinal Optic Flow Learning (ROFL), which enables control over motion statistics and their causes. We then present a new hierarchical VAE and test it against alternative models on two downstream tasks: (i) predicting ground truth causes of retinal optic flow (e.g., self-motion); and (ii) predicting the responses of neurons in the motion processing pathway of primates. We manipulate the model architectures (hierarchical versus non-hierarchical), loss functions, and the causal structure of the motion stimuli.

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 first introduce a novel synthetic data framework, Retinal Optic Flow Learning (ROFL), which enables control over motion statistics and their causes. We then present a new hierarchical VAE and test it against alternative models on two downstream tasks: (i) predicting ground truth causes of retinal optic flow (e.g., self-motion); and (ii) predicting the responses of neurons in the motion processing pathway of primates. We manipulate the model architectures (hierarchical versus non-hierarchical), loss functions, and the causal structure of the motion stimuli.

A good explanation among many others of logical consequence is that given a possibly empty set of assumptions A that are said to entail x, written A x, one is bound to believe x whenever it is the case that she believes the conjunction of all elements of A. This explanation shows a weak correspondence with the Tarskian semantic definition of logical consequence, according to which A entails x if and only if x is true under all interpretations (models) that make all elements of A true. But the relevant term here is "bound" ("is obliged to"), and the talk is of a normative conception of logical consequence. But where does this command come from? To understand some of the implications of what we call here a normative account of logical consequence, we will consider a notion which is not very well understood until the present day: that of "defeasible subsumption". To grasp this notion, it suffices to consider the following example.