Semantic-Decoding

Minimal SRM-based semantic decoding pipeline for podcast ECoG data.

This folder contains only the core code needed to:

1. Load podcast high-gamma neural data and text embeddings
2. Build subject-wise neural targets from word onsets
3. Optionally fit SRM on selected subjects
4. Train a text-to-neural decoder per subject
5. Compare performance with cosine similarity on the test split

Run

python srm_cosine_pipeline.py --config config.example.json 

Input data format

The pipeline expects a pickle file with:

{
  "significant_elecs": (n_channels,), 
  "word_data": { 
    "subj": (n_channels,),
    "highgamma": (timepoints, n_channels),
    "onset": (n_words,), 
    "word": (n_words,)   
  } 
}

And a text embedding .npy file with shape:

(n_words, embedding_dim)

Outputs

Results are saved per subject under the configured output directory:

• test_metrics.json
• hyperparameters.json

Notes

• Text PCA is fit on the train split only.
• SRM standardization is computed per subject using train samples only.
• Final target z-scoring is also computed from the train split only.
• Test performance is reported with cosine similarity and MSE.