Neural population geometry and optimal coding of tasks with shared latent structure
This work addresses how neural activity geometry contributes to efficient learning across tasks, providing insights for neuroscience and AI, though it is incremental in building on existing theories of neural coding.
The authors developed an analytical theory linking neural population geometry to generalization performance in multi-task learning with shared latent structure, showing that four geometric measures determine performance and that experimentally observed disentangled representations emerge as optimal solutions, validated with macaque ventral stream recordings.
Humans and animals can recognize latent structures in their environment and apply this information to efficiently navigate the world. However, it remains unclear what aspects of neural activity contribute to these computational capabilities. Here, we develop an analytical theory linking the geometry of a neural population's activity to the generalization performance of a linear readout on a set of tasks that depend on a common latent structure. We show that four geometric measures of the activity determine performance across tasks. Using this theory, we find that experimentally observed disentangled representations naturally emerge as an optimal solution to the multi-task learning problem. When data is scarce, these optimal neural codes compress less informative latent variables, and when data is abundant, they expand these variables in the state space. We validate our theory using macaque ventral stream recordings. Our results therefore tie population geometry to multi-task learning.