August 2025

Cooperative actions of interneuron families support the hippocampal spatial code

• Methods: Multi-shank silicon probes + optogenetic tagging of four interneuron families (Pvalb, Sst, Vip, Id2) in hippocampus and neocortex.
• Classification: Machine learning on physiological fingerprints achieved >89% accuracy, bridging genetics and in vivo physiology.
• Findings: Distinct interneuron families differentially sculpt pyramidal place fields; Pvalb suppressed early spikes, Sst/Id2 suppressed late, suggesting time-division control.
• Conclusion: Cooperative interneuron interactions are central to the hippocampal spatial map, beyond excitatory synaptic plasticity.

Speed modulations in grid cell information geometry

• Approach: Introduced Gaussian Process with Kernel Regression (GKR) to model neural manifolds + noise covariance in grid cell populations.
• Manifold dilation: Faster running speed stretched the toroidal-like manifold of grid codes.
• Noise trade-off: Noise also increased, but information gain from dilation outpaced noise costs.
• Result: Fisher information and decoding accuracy improved at higher speeds, while noise correlations projecting onto the manifold impaired coding.
• Implication: Population-level coding, not just single cells, explains why navigation accuracy improves with movement speed.

Tonic dopamine and biases in value learning

• Problem: Psychiatric disorders often feature biased future predictions (pessimism in depression, optimism in addiction).
• Model: RL framework with basal ganglia circuit architecture + dopamine receptor pharmacodynamics.
• Mechanism: Differences in D1 vs. D2 receptor affinities and sigmoidal dose–response curves mean tonic dopamine shifts sensitivity to positive vs. negative reward prediction errors.
• Evidence: Explains optimistic and pessimistic learning in both humans and mice.
• Significance: Provides a biologically grounded mechanism linking tonic dopamine to learning biases and psychiatric symptoms.

Reconciling flexibility and efficiency: compositional cognitive maps in MEC

• Model: Cognitive maps constructed from predictive object representations (PORs) perturbing a baseline open-space grid.
• Composition: PORs are translation/rotation invariant and recombinable, enabling efficient compositional map building.
• Neural mapping:
  • Object vector cells = encode POR building blocks.
  • Grid cells = baseline efficient spatial metric.
• Insight: MEC supports planning-ready compositional maps, reconciling flexible recombination with efficient path planning.

Competitive integration of time and reward in foraging

• Task: Head-fixed mice performed a VR patch-foraging paradigm with probabilistic, diminishing water rewards.
• Behavioral model: Mice integrated elapsed time (positive contribution) vs. received rewards (negative contribution), scaled by a latent patience state.
• Deviation from theory: Decisions systematically diverged from Marginal Value Theorem but were captured by competitive integration models.
• Neural evidence: Neuropixels across frontal cortex showed ramping activity encoding the integrated decision variable, oppositely modulated by time vs. reward.
• Conclusion: Frontal ramping dynamics provide a neural substrate for value-sensitive patch leaving.