Time robustness of deep learning models for real-time neural decoding of arm movement.

In the past years, it has been proven that the use of AI algorithms is useful to improve the control on Brain-Machine Interfaces (BMI). Through the search for patterns in the neural signal by Machine Learning (ML) algorithms, it was possible to derive information about the brain activity, leading to a greatly improved accuracy in neural decoding. BMIs were proposed that allowed people to compose sentences by selecting letters on a virtual keyboard, others that allowed people to control chronic pain through stimulation of the brain in a virtual environment, and others that allowed control of a robotic arm. However, it has been proven that neural signals change over time, mainly due to neuroplasticity, but also due to inflammatory processes and the displacement of electrode microarrays in the case of intracortical signal retrieval (ECoG). Unfortunately, the change in signals inevitably leads to a deterioration in decoding capabilities and performance and it is still an open research problem. The aim of this Master Thesis work is to find an optimal Transfer Learning (TL) strategy that maintain high decoding performance and increase robustness without requiring a large number of recording sessions to re-train the algorithm. This work fits in the context of a larger project aiming building a BMI to let a Non-Human Primate (NHP) control a robotic arm. The first step was to find a public dataset that provided several recordings of an NHP performing a task involving an arm or hand movement. In particular, it was essential that the neural signals were ECoG retrieved in the primary motor cortex (M1) and that several recording sessions were provided over a longer or shorter period. Four different datasets were analysed and the Zenodo dataset was chosen, which provided 37 sessions of recording an NHP while reaching a random target in the transversal plan, allowing for the analysis of coherent neural data over a long time span. Subsequently, it was necessary to apply pre-processing techniques to derive multi-unit activity (MUA) signals, spike counts and hand velocity. Next, three neural networks (Bidirectional Recurrent Neural Network, Bidirectional Long Short-Time Memory, and Bidirectional Gated Recurrent Unit) were trained and optimized on the signals of the first recording session, in terms of root-mean squared error (RMSE) and Pearson’s Correlation Coefficient (CC). This model was used to predict the hand velocity signal for all the following sessions; therefore, with no adaptation to the variability of the signal in time, the result represented the lower bound for the decoding performance of the model through time. After that, the upper bound of the decoding performance was obtained by training a dedicated model on each single session, maintaining the weights of the previous training session. At this point, several TL strategies were evaluated to explore the spectrum in between the previous approaches. Most strategies involved re-training the model on a small amount of data less frequently that at each session, to reduce the need for labelled data. Eventually, a self-supervised approach was proposed to evaluate if a solution that did not require kinematic signals as labels could achieve comparable performance to supervised learning approaches.