Long overdue: hierarchical latent inference in neural circuits

I realize I didn’t announce Rylan’s work on neural circuits performing hierarchical latent inference when it appeared in NeurIPS last December! Here’s a brief summary:

Often we must make decisions between two alternatives, but which of the two is more likely to result in a payoff might change over time. The International Brain Laboratory trains animals in a setting where the payoff balance between the two switches discretely, and the switches are uncued. To make a good decision, animals must infer the unknown balance of payoff probabilities, by accumulating information over multiple trials. In addition, each trial requires integration of evidence. Rylan showed how a neural circuit can perform this latent hierarchical inference task, by training RNNs. The RNN finds solutions that are not Bayesian but rather simple linear filters over time. It also uses the same coupled activity subspace to integrate evidence both within-trial and across-trials, albeit with different time-constants.

We also introduce a new method for network “distillation”, RADD: training a small network based on the hidden states of a bigger network to achieve a highly compact network that can solve the same latent inference problem (when the small network trained directly on the problem cannot do so). Next up: comparison of model predictions with neural data.

Circuit inference biased in strongly recurrent networks and a possible solution

Abhranil’s new paper is out! Congrats Abhranil.

We show that connectivity inference from activity in strongly recurrent networks will be systematically biased regardless of data volume and even given access to every neuron in the network. But measuring activity far-out-of equilibrium after a simple low-dimensional suppressive input could ameliorate the bias.

Memories from patterns: a review

“Memory from patterns: Attractor and integrator networks in the brain”, with Mikail, is submitted. Comments and suggestions are welcomed!

The theory of how complex patterns emerge from simple interactions and constituents is one of the big ideas in biology, explaining animal coats and morphogenesis.

The same principles can produce dynamical states for computation in the brain, in the form of attractor networks. We review how attractor networks generate states for robust representation, integration, and memory.

Our review covers the conceptual ideas, the theory, and the potential utility of continuous and discrete attractor networks, then focuses on the empirical evidence that the brain computes using these structures. Finally, we discuss modern developments in combining the concepts of modularity and attractors, and list future challenges.

We hope the review provides a vista of a field of systems neuroscience driven by theoretical ideas, where theory and experiment have come together fruitfully and harmoniously.

What are the main sources of homing error in young and aging humans?

Matthias and Ingmar’s paper is out in Nature Communications!

Homing, or determining the straight path back to “home” after a winding outbound journey, is a critical but error-prone computation.

What are the main causes of human homing error, and how do they change with age?

We put humans into immersive VR, measured homing errors along winding paths, and modeled the time-resolved process of error accumulation with a Langevin-type diffusion equation.

We found that forgetful integration, biases in velocity estimation or integration, and reporting or readout errors do not limit homing ability; rather, the bottleneck is an accumulation of unbiased random error.

The random error accumulates with movement but not time, suggesting it is related to velocity sensing rather than integration.

Aging humans do worse; their diminished performance is not from new sources of error but an increase in the unbiased error already limiting young subjects.

Representing high-dimensional cognitive variables with grid cells

Mirko & Marcus’ paper is out in PLoS Computational Biology!

If grid cells encode non-spatial cognitive variables, they should be able to represent spaces of dimension greater than two.

Can grid cells construct unambiguous representations of higher-dimensional inputs without recurrent rewiring to form higher-dimensional grid responses, a cell-inefficient and inflexible mechanism?

We show how they could do so by combining low-dimensional random projections with “classical” two-dimensional hexagonal grid cell responses.

Without reconfiguration of the recurrent circuit, the same network can flexibly encode multiple variables of different dimensions while maximizing the coding range (per dimension) by automatically trading-off dimension with an exponentially large coding range.

This model achieves high efficiency and flexibility by combining two powerful concepts, modularity and mixed selectivity, in what we call “mixed modular coding”.