Neural Decoder for Behavioral Choice

Neuropixels recordings give hundreds of simultaneously recorded neurons — noisy, correlated, and high-dimensional. The question: can left vs right choice still be decoded above chance from a short post-stimulus window, without pretending the decoder is a causal circuit map?

Shows whether behaviorally relevant structure lives in a low-dimensional neural subspace. Cross-validation alone is not enough; a permutation null asks whether the effect beats chance label structure.

• Steinmetz et al. 2019 — Mouse Richards session from DANDI 000017.
• 778 neurons × trial firing rates in a 0–300 ms post-stimulus window.
• 174 left/right choice trials after filtering for analyzable behavioral outcomes.

Build a trial × neuron firing-rate matrix → PCA to compress correlated population activity → L2-regularized logistic regression on the top principal components → 5-fold cross-validated ROC-AUC, plus a 500-shuffle label-permutation null to estimate chance.

• Implemented the full decoding pipeline from firing-rate features through figures.
• Chose PCA + L2 logistic regression to control overfitting with 778 features and only 174 trials.
• Ran a 500-iteration label permutation test so “above chance” is quantified, not assumed.
• Swept principal-component count and showed performance saturates around ~11 PCs.

• Primary metric: ROC-AUC for ranking left vs right trials (reported with fold mean ± std).
• Secondary metric: classification accuracy under the same CV splits.
• Null distribution mean AUC ≈ 0.505 — confirms the chance baseline sits at ~0.50.
• Regularization and dimensionality reduction are required; raw-rate decoding overfits in this regime.

• p ≫ n features vs trials — without PCA, the classifier memorizes noise.
• Population correlations mean many neurons carry overlapping information; the useful subspace is smaller than 778.
• Decoding choice is not the same as identifying a specific causal circuit.

• ROC-AUC 0.884 ± 0.025; accuracy 0.799 ± 0.039 under 5-fold CV.
• Permutation p-value < 0.0001 versus the label-shuffled null (mean null AUC ≈ 0.505).
• Decoder performance saturates near ~11 principal components — most usable signal lives in a compact subspace.

Cross-validation measures stability; permutation testing asks whether the effect beats chance. Both belong in a decoding write-up.