context-contrasting-model

Computational model code accompanying Seignette et al. (2026).

Summary

Pyramidal cells (PyCs) in visual cortex change their responses with experience, but not all in the same way: after familiarization, some cells respond less to the images they see (-NO), some respond more (+NO), and some begin to respond even when the image is occluded and only its context is present (+O). The computational model in this repository shows that all three transitions can emerge from one small canonical circuit: a two-compartment PyC that combines feedforward input at the soma with contextual feedback at its apical dendrite, and a PV interneuron that divisively inhibits it. We propose three opposing plasticity rules acting together during familiarization: feedforward synapses depress with repeated activity, apical feedback synapses strengthen with repeated presentation of overlapping contextual inputs, and feedforward/lateral inhibition strengthens with correlated PV and PyC activity. Which transition a cell undergoes depends on its starting synaptic balance, and the same rules reproduce the diverse set of empirically-observed transitions for both familiar and novel images.

Change in occluded (O) vs. non-occluded (NO) response after familiarization: empirical cells (left) and the three familiar-image model transitions (right).

Model code for six mean-field models of PyC/PV transitions: familiar and novel, each with +NO, +O and -NO transitions. Every model is one YAML file in configs/; trial timing, counts and contextual breadth are in configs/protocol.yaml.

Setup

uv sync

Layout

Path	Purpose
configs/	One parameter set per model, plus protocol.yaml
src/model.py	One PyC/PV pair: dynamics and three plastic synapses
src/simulate.py	Naive probes, familiar training, expert probes, acute silencing
src/robustness.py	One-at-a-time parameter and initial-weight sweeps
plotting/	plot.py and final_results.ipynb for the final panels

Run

From the repository root:

uv run python -m src.simulate      # results/model.csv, results/model_trial_responses.csv
uv run python -m src.robustness    # results/model_robustness.csv, figures/robustness/
uv run jupyter notebook plotting/final_results.ipynb   # figures/final_layout/

Simulation defaults to 50 test trials and seeds 0–9 for every config (--test-trials, --seeds, --jobs). Each config/seed pair runs as its own process, and results do not depend on the worker count.

Outputs

• results/ground_truth.csv: empirical sector membership, per-neuron traces and transition vectors (fixed input, exported from the thesis data analysis).
• results/model.csv: trial-averaged traces and transition vectors per config and seed (naive, expert, FB/PV/combined silencing).
• results/model_trial_responses.csv: per-trial NO/O responses.
• results/model_robustness.csv: one row per sweep point (figures go to figures/robustness/).