Applying new machine learning techniques to the analysis of neural time-series data

This challenge is focused on applying a novel machine learning technique for time-series analysis to neurological data, broadly construed. The technique in question involves the use of a recurrent neural network (RNN) model called a Legendre Memory Unit (LMU) that has recently been shown to have several computational advantages over past RNNs (e.g. LSTMs, GRUs, etc.), namely; reduced training times, reduced number of parameters, reduced number of nonlinearities, and greater memory capacity by several orders of magnitude. These benefits stem from a principled derivation of the optimal network weights needed to efficiently represent sliding windows of time. Because the LMU is so well-suited to tracking and compressing long time-scale temporal information, we believe it may unlock improved ways of understanding and interpreting arbitrary time-series datasets, including those derived using EEG, fMRI, or electrophysiology techniques.

We anticipate that this challenge will provide an opportunity for collaboration between engineers with machine learning skills and neuroscientists with an interest in applying new data analysis techniques to research problems with medical applications. ABR will provide Tensorflow/Keras/Nengo code and advice for using the LMU, and offer technical help to participants as needed. An ideal outcome for the project would be a prototype medical application of the LMU with clear prospects for future commercialization. The LMU and related technologies are the intellectual property of Applied Brain Research and must be licensed for commercial usage.

Learn more about Applied Brain Research (ABR) at: https://appliedbrainresearch.com/