Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Predictions emerge in neural networks trained to perceive Bach's music

Akanksha Gupta¹, Alejandro Tabas²; ¹Université d'Aix-Marseille, ²Basque Center on Cognition, Brain and Language

Presenter: Alejandro Tabas

Predictive processing proposes that the prior knowledge relevant for inference is compressed into a prediction on the immediately future states. Here we inquire whether neural networks trained to infer the current latent state in musical sequences develop a set of internal predictions on what comes next. We used noisy tokenized Bach compositions as sensory inputs and trained RNN as models of neural circuits. We first trained the networks to infer the current latent state (token of the composition without noise) given a stream of observations (tokens of the composition with noise). After the training, we inspected whether the internal states of the network stored predictive information on the next token. To do this, we fitted a linear readout from the hidden states of the network optimized to predict the next latent state. To ensured that the predictions were stored in the network and not computed by the linear readout, we compared the predictive performance of the network with that of a linear network trained to predict the next latent state based on the current latent state. The results confirm that neural circuits optimised to perceive the current state learn to predict future sensory input, suggesting that predictive capabilities emerge as a natural consequence of such optimization. These findings offer computational evidence for predictive processing and provide insights into how biological systems might compress their prior knowledge and use it to navigate in noisy environments.

Topic Area: Predictive Processing & Cognitive Control

Extended Abstract: Full Text PDF