Hosted by the Intelligent Systems Laboratory / Interactive AI Centre for Doctoral Training
We propose a machine learning approach to quantifying multi-day stability of neural representations in local field potential (LFP) data. We trained deep neural networks (DNN) to decode position from wavelet decomposition images of LFP simultaneously recorded from hippocampus, prefrontal cortex and parietal cortex of four rats. Data were collected while each rat learned a maze-based memory and decision-making task over >20 days, with different DNN trained using data from separate days. By calculating the decoding accuracy of a DNN trained on data from different days, we quantified the representational stability of spatial encoding in these brain regions. This method enables a holistic, unbiased analysis of neural codes, requiring only minimally-processed LFP input and negating sampling or sorting bias inherent in spike-based metrics.