The paper shows that recurrent neural networks trained to clean up noisy sequences of musical notes develop internal signals that look like prediction. In other words, when the networks learn to make sense of noisy input, they also come to carry information about what will happen next. The authors argue this offers a possible reason why the brain might use prediction to support perception.

The researchers trained recurrent neural networks (RNNs) — a type of model that processes sequences by feeding back its own activity over time — to denoise a tokenised version of Johann Sebastian Bach’s compositions. Tokenised here means the music was converted into a sequence of discrete symbols, like short words. The networks were trained at several levels of noise, so they had to recover the original token sequence from inputs that had been randomly corrupted.

To test whether the trained networks were implicitly predicting the next token, the team froze the RNN weights and trained a simple linear readout on the network’s internal states to predict the next token. They compared this readout against an independently trained linear benchmark model that did not rely on the RNN states. The linear readout from the RNNs performed better than the benchmark at moderate noise levels. The authors take this as evidence that the RNNs had come to rely on predictive information to support the denoising task.

The paper also reports that the networks’ responses to inputs were proportional to prediction error — that is, the size of the internal response matched how surprising the input was compared with what the network implicitly expected. This pattern is one of the neural signatures often associated with predictive processing in the brain. Importantly, these signatures appeared even though the networks were never explicitly trained to predict; they were only trained to improve perception (denoising).