Brain-like computation and intelligence

Week 4: Path integration and attractor models

Today we had a first glimpse at task-driven modeling, we’ll see more in the next weeks for different domains (vision, audition, proprioception, language and motor). 

The basic premise of this approach is that neural circuits are optimized through evolution for particular functions and thus, if we optimize models on particular functions they might make nontrivial predictions that relate to the brain. That's vague, but we will saw one concrete example that should give you some idea. Again, we'll see multiple examples over the next weeks.

Path integration is an important brain function; in mammals, the hippocampal formation supports this computation via specialized cell types. We saw behavioral evidence that ants, zebrafish and flies can do this. From the latter two species, we also discussed some recent recordings and perturbations to give you an idea of the current state of knowledge.

1. Can one engineer a system to carry out path integration?

We showed that ring attractor models can implement path integration (for an angular variable). This presentation was based on the classic paper by Zhang et al, J Neuro 1996. Check it out for more details. We also illustrated how this work was generalized to grid cells by many people. 

2. Can one learn to path integrate and will it work like an attractor model?

We discussed that learning to path integrate converges to similar solutions (with the right constraints). Here we discussed the Sorcher et al.'s model (Neuron 2023). Together with Markus Frey, and Mackenzie Mathis, I also wrote a perspective piece that puts this in its bigger context.  You will also work with the code from this model in the exercises. 

Overall, attractor models are powerful models of brain function. Attractor models are a first “brain-like circuit” in this class. Think about how this system computes vs., for instance a CPU or how you would do path integration using calculus.