Binarized Neural Networks for Resource-Efficient Spike Sorting

Deep learning is fastly gaining ground in neuroscience. In the field of implantable brain computer interfaces, a fundamental application of deep learning is to sort action potentials (known as spikes), measured with extracellular electrodes, according to their origin neurons. This enables the generation of precise modulatory patterns of neuronal circuits. Deep learning-based spike sorting algorithms are based on power-intensive dot products, which poses challenges for on-chip processing with resource-constrained devices. In contrast, binarized neural networks offer great potential for on-chip sorting, mainly relying on bitwise operations and accumulations. However, recently published binarized models perform significantly worse than deep full-precision networks and fail on challenging neural data. This work presents a binarized neural network for spike sorting that narrows the performance gap between recently developed binarized models and more accurate full-precision models. The novelty of this work resides in the developed network architecture. In comparison to previous research, this work presents a deep binarized neural network featuring two hidden layers, each containing 256 units to effectively capture the spike characteristics of complex neural data. Before training, spikes were pre-sorted in an unsupervised way to generate pseudo-labels. Subsequently, the deep binarized model and an equally sized full-precision model were trained and evaluated using experimentally obtained and synthetic spike waveforms. The proposed binarized model could achieve results close to more advanced network types, such as convolutional and long short-term memory networks, which is remarkable considering that the binarized model was primarily designed to maintain a balance between resource consumption and accuracy. The equally sized full-precision model could even outperform the aforementioned models, despite its much lighter architecture.