Transfer Learning strategies for time robust neural decoding in a Brain-machine interface

Recent and continuous advancements in neuroengineering and Machine Learning demonstrates the huge potential of Brain-machine interface in the field of neuroprosthetics. This rapidly evolving technology aims to provide innovative solutions to people affected by disabilities, in order to restore motor, sensory and cognitive functions. This is the goal of B-Cratos project, whose purpose is the development of a closed-loop neural interface for controlling a robotic hand prosthesis also capable of providing sensory feedback to the patient. Neural decoding is possible thanks to Deep Learning models trained using high-performance computing resources on datasets acquired from the German Primate Center (Deutsches Primatenzentrum, DPZ), that can classify signals recorded via implanted microelectrode arrays. DPZ researchers recorded the neural activity of two macaque monkeys trained to perform a grasping task with a series of objects of different shapes and sizes. These signals were pre-processed and used in previous works for the development of a classifier. For this purpose, a Bidirectional Recurrent Neural Network was trained to successfully identify the objects grasped by the monkeys, simulating a real-time decoding. In this work, different transfer learning strategies were implemented in order to exploit the knowledge acquired by the pre-trained classifier in a model that can be used on new recording sessions, subsequent to the first one. The possibility to effectively transfer the information learned from a pre-trained model would represent a significant advantage in the use of BMIs, considering the high variability of neural signals and the need to recalibrate the device to maintain high performance over time. Before implementing the transfer learning, the two datasets used were appropriately reduced, in order to present a common dimensionality: the different dimensionality was due to the application of offline spike-sorting algorithms independently on the two sessions. Feature extraction was performed through different models and their performances were evaluated in terms of final accuracy achieved after the fine-tuning. Furthermore great importance was attributed to a convergence analysis: this analysis was conducted to evaluate the classifier’s ability to quickly learn the neural patterns relevant for object classification. Among all the models developed for the reduction and consequent classification, the Partial Reduction model showed the best results, consisting in a single dense layer, with a latent space of the same dimension of the data from the first training session, preceding the layers responsible for classification whose weights can be setting as trainable or non-trainable parameters. This network showed to outperform a classifier trained from scratch in the classification task involving 37 objects with different shapes and sizes, both for the accuracy achieved and for the convergence speed. Specifically, the variant with frozen layers reported an accuracy of 45% and reached the 30% accuracy threshold in only 9 training epochs, compared to the reference that achieved a 37% final accuracy and required 45 epochs in order to get to the threshold. Despite the limitations presented by the datatset, this work showed that an appropriate weights initialisation can contribute to better and faster re-training, providing promising results for the implementation of transfer learning strategies in neural decoding models.