Receptive Fields - Surprise & Sparsity

In this study, I investigate how auditory receptive fields can emerge through synaptic adaptation driven by surprising sounds, guided by efficient coding principles in sensory processing. I developed a computational model using an unsupervised single-layer network that adapts synaptic weights based on the level of surprise, maximizing the neural response to unexpected sounds while minimizing overall activity. Using an autoregressive generative model trained on speech data, I define surprise as the negative log probability of observed energy across time-frequency bins. Synaptic weights adjust through three factors—alpha, beta, and gamma—to simulate processes like Long-Term Potentiation and synaptic depression. This approach leads to tuning properties similar to those found in auditory cortex neurons, capturing a balance between stability and adaptability. My results suggest a biologically plausible mechanism for the development of auditory receptive fields in response to surprise, providing insight into fundamental principles of sensory representation.

(Note: We submitted this work to Cosyne 2025 and is currently under review!)