Neural population geometry and optimal coding of tasks with shared latent structure | Nature Neuroscience
The authors analytically determine how neuronal correlations and geometry collectively determine readout generalization across tasks and show how these geometric features follow distinct trajectories over the course of learning.
Main Humans constantly solve different instances of similar problems. We brake at stop signs, stop at red lights and slow down in crowded streets. We do these things effortlessly and efficiently learn to use new sensory cues to regulate our behavior. This is possible because we are able to recognize overt symbols such as road signs as well as more abstract visual cues such as the crowdedness of a street. More generally, humans and other animals learn to recognize latent variables in their environment and use them to guide their behavior across contexts and tasks. Recent experimental…
saved by
related reading
- Evidence that recurrent circuits are critical to the ventral stream’s execution of core object recognition behaviornature.com
- Toy Models of Superpositiontransformer-circuits.pub
- Dimensionality reduction in neural data analysisxcorr.net
- [2104.09743] Neural tuning and representational geometryarxiv.org
- Complementary task representations in hippocampus and prefrontal cortex for generalizing the structure of problemsnature.com
- The World Inside Neural Networksgoodfire.ai
- [2409.06843] Universal scale-free representations in human visual cortexarxiv.org
- 2205.08413.pdfarxiv.org
- Scaling Properties of Dimensionality Reduction for Neural Populations and Network Models | PLOS Computational Biologyjournals.plos.org
- Spatially embedded recurrent neural networks reveal widespread links between structural and functional neuroscience findingsnature.com
- Neural tuning and representational geometry | Nature Reviews Neurosciencenature.com
- Neural population-based approaches have opened new windows into neural computations and behaviorthetransmitter.org