Poster Presentation

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Representational drift dominates cross-session decoder degradation in mouse primary visual cortex

Santiago Tobio¹; ¹Universidad de San Andres

Presenter: Santiago Tobio

Primary visual cortex (V1) carries rich stimulus representations, yet mice often fail to fully exploit this information during behavior. What constrains the translation from neural representation to behavioral readout? Using the Allen Institute Visual Behavior Two-Photon dataset (32 mice, 204 sessions, Slc17a7-Cre line, GCaMP6f), we trained five decoder architectures (logistic regression, multilayer perceptrons, and a temporal convolutional network) to classify eight natural images from V1 population activity. Intra-session decoding achieved a median accuracy of 85.2% (chance: 12.5%), with logistic regression matching nonlinear models in over 50% of sessions, indicating that V1 representations are largely linearly separable. Surprisingly, decoder accuracy was negatively correlated with mouse behavioral performance (hit rate: ρ = -0.295, p < 0.001, controlling for population size), and arousal proxies (pupil area, running speed) did not predict decoding accuracy (Δ R² = 0.004, p = 0.80). To investigate temporal constraints on readout, we developed a decomposition framework that isolates three sources of cross-session accuracy loss: neuron loss from imperfect tracking, representational drift, and gain/scale shifts. Across 241 session pairs, representational drift accounted for 59.5% of the total accuracy drop (∼19 percentage points), compared to 31.6% for neuron loss and 8.9% for gain shifts. Drift was significantly greater for novel than familiar stimuli (p < 10⁻⁸), scaled with inter-session interval (ρ = 0.258), and was pair-specific rather than a stable mouse trait (ICC = 0.287). These results suggest that representational drift, rather than arousal-gated processing or limited neural information, poses the primary challenge for stable downstream readout from V1.

Topic Area: Methods, Tools, Theory & Neural Coding