Fgpa Implementation Of Artificial Neural Networks For Brain Computer Interface Applications

Author : Jonathan Mishler
Publisher :
ISBN 13 :
Total Pages : 0 pages
Book Rating : 4.:/5 (139 download)

DOWNLOAD NOW!

Download or read book Implementing an Integrate-and-fire Neural Network on a Bidirectional Brain-computer Interface written by Jonathan Mishler and published by . This book was released on 2022 with total page 0 pages. Available in PDF, EPUB and Kindle. Book excerpt: There is a growing interest in using intracortical microstimulation (ICMS) as a means of neurorehabilitation, from using it to rewire synaptic connections in the brain, to providing a means of providing artificial sensations to provide feedback to patients controlling external devices such as robotic limbs via decoded neural activity. Towards the goal of neurorehabilitation, there have been recent efforts to integrate artificial neural networks (ANNs) with the brain and train them to manipulate neuronal activity in a context-specific manner with the goal of developing brain-computer interfaces (BCIs) that can restore function to damaged neural circuitry. In this thesis, I present my work whose aim was to interface artificial spiking neurons with biological neurons in primary motor cortex to create hybrid biological/artificial neural networks that altered the firing dynamics of the biological neurons. In these hybrid networks, spikes that are detected from the biological neurons send artificial postsynaptic potentials (PSPs) to the artificial neurons, whose magnitudes and polarities are defined by weights that characterize the strengths of their connections. When the membrane potentials of the artificial neurons exceed a predetermined threshold, they spike, and subsequently trigger ICMS to manipulate the activity of the biological neurons. We first characterize the effects of ICMS on neural activity in primary motor cortex (M1) of pigtail macaques, and show how it elicits a brief excitatory response followed by a longer inhibitory response. We then show how the probability of evoking single action potentials has at least three dependences – the stimulation amplitude, time delay between the neuron’s previous spike and stimulation onset, and its firing rate. Finally, we show how repetitive stimulation can increase or decrease the probability over time, likely due to mechanisms of short-term plasticity. I used these results to explore how various properties of the hybrid biological/artificial neural networks shape the closed-loop dynamics between the biological and artificial neurons. To do so, I measure changes in the auto-, and cross-correlograms of the biological neurons between their spontaneous and closed-loop dynamics, and show how features within the correlograms are related to the size, connectivity, number of hidden layers, magnitude of inhibition, and stimulation delays of the network. I then show how the closed-loop dynamics can be simulated by recording the spontaneous activity as well as the open-loop responses of the biological neurons to ICMS, and use the model to further characterize how the features of the hybrid network influence the closed-loop dynamics free from experimental constraints such as the non-stationarity of the firing rates and evoked spike probabilities of the biological neurons over time. I also use the model to explore how stimulation artifacts impacted the closed-loop operation of the hybrid neural network by obstructing the detection of spikes. To do so, I characterize the number of obstructed spikes as a function of both the network size and connection strength between the biological and artificial neurons. I found that compared to the size of the network, the strength of the connections between the biological and artificial neurons was the greater determinant of how many spikes were blanked. Lastly, I discuss preliminary work of the development of a computational model whose goal is to demonstrate that ANNs can be interfaced with the brain to restore lost motor function. I first train a small recurrent ANN, which simulates a motor circuit within the brain, to perform a 1D motor task with a reinforcement learning algorithm. I then lesion the network by deleting a subset of the neurons and their related connections, and show how even with retraining, the network is incapable of relearning the task. Finally, I discuss future steps in which a second ANN, whose outputs use ICMS to activate neurons within the original ANN, can be used to restore network function and task performance.