Neural Noise- Predicting artificial neural network (ANN) representations can be used to learn a recognition model for music identification from brain recordings.1
- Decoding from brain recordings via ANN-derived representations extends to auditory and cognitive content and transfer learning.1
- Nature-published; method relevance for non-motor BCI.1 1
Weekly enrichment (2026-07-20)
- The method reverses prior brain–ANN alignment work: instead of regressing cortical responses from artificial neural network (ANN) representations, it uses ANN representations as a supervisory (teacher) signal to train an EEG recognition model on noisy non-invasive recordings.2 3
- The task is framed as 10-class music identification (chance level 0.1) using the publicly available NMED-T dataset of EEG from 20 subjects listening to 10 unique songs.3
- Training uses a contrastive-learning approach with dual classification heads on both the EEG and ANN branches, deliberately not aligning the ANN branch to the noisy EEG branch to preserve its discriminative power; SampleCNN serves as the base 1D CNN encoder.3
- With the optimal PredANN loss weight of 0.05 (seed 42), the model reached 0.482 classification accuracy versus 0.474 for the no-PredANN baseline, a statistically significant gain (McNemar’s test, p = 1.70 × 10⁻⁸).3
- Averaged across random seeds, predicting ANN representations improved accuracy by about 14.1% over baseline, indicating more robust, initialization-independent learning.3
- The largest gains occurred when assuming an approximately 200 ms brain-response delay to music, consistent with known auditory cortical latencies; longer EEG input durations further increased accuracy.2 3
- Performance varied by stimulus and subject: songs #2, #4, and #5 (electronic/distinctive features) consistently exceeded 85% accuracy, while individual subject accuracy ranged widely (top subject mostly >80%, lowest subject as low as 20%).3
- Published in Scientific Reports (2025), DOI 10.1038/s41598-025-02790-6; a follow-up (PredANN++) adds expectation-related teachers (surprisal, entropy) alongside acoustic ones, reporting single-model accuracy up to ~0.859 with complementary gains from combining representation types.3 4
- BCI implication: leveraging ANN teacher representations offers a route to train auditory/non-motor neural decoders from limited, noisy non-invasive EEG, relevant to brain-computer interfaces and music-cognition research.2 4