Shogo-Noguchi/PredANNpp-Acoustic

PredANN++ (Acoustic, MuQ) — Encoder-only checkpoint

Model description

This model is a finetuned EEG encoder for song identification from 3-second EEG segments (10-class classification) on the NMED-T dataset.

The encoder is trained with multitask pretraining (masked prediction) using MuQ acoustic embeddings, followed by finetuning with cross-entropy on song ID.

• Input: EEG (128 channels, 125 Hz, 3 seconds)
• Output: Song ID logits (10 classes)

Intended use

• Research use for EEG-based music recognition
• Comparing the effect of acoustic neural representations (MuQ) with expectation-related representations

Not intended use

• Medical diagnosis
• Any clinical decision making
• Commercial usage without verifying upstream non-commercial licenses

Training data

• NMED-T (Naturalistic Music EEG Dataset – Tempo), 10 songs, 20 subjects, trial=1.

Training procedure (high-level)

1. Multitask pretraining: encoder-decoder masked prediction of MuQ acoustic tokens (50% masking)
2. Finetuning: encoder-only training for Song ID classification

Evaluation

The repository contains an evaluation script:

• codes_3s/analysis/evaluate.py

License and upstream dependencies (IMPORTANT)

This checkpoint is trained using features computed with:

• MuQ model weights (CC-BY-NC 4.0)

Therefore, this checkpoint is distributed as CC-BY-NC 4.0.

Citation

If you use this model, please cite:

• The PredANN++ paper (to be added upon publication)
• NMED-T dataset paper
• MuQ model

Acknowledgements

We borrow and adapt code from multiple repositories. Please see:

• THIRD_PARTY_NOTICES.md in the GitHub repository.