Search Papers | Poster Sessions | All Posters
Poster A100 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Benchmarking Deep Learning Architectures for Predicting Visual Stimuli Given Single Neuron Spike Patterns
Raymond Carpenter1 (), Cole Vita1, Vladas Pipiras1; 1University of North Carolina at Chapel Hill
Drawing insights from neuronal processes is integral for understanding the neural mechanisms underlying cognitive processes, providing a higher definition recording for brain-computer interfaces, and helping develop advanced neurorehabilitation strategies. Our study sought to survey and identify machine-learning models and deep-learning architectures capable of predicting visual stimuli based on the spike patterns of single neurons. We worked with Neuropixel data from the Allen Brain Observatory [1,2] consisting of the firing rates from single neurons, also called spike trains, from several male mice's visual cortex, thalamus, and hippocampus. Each recording involved around 2,000 separate units. The mice were shown 118 different natural images of predators, foliage, and other scenes from their natural habitat at random in repetition and for 250 ms each. The firing rates of the separate units were then used as predictors for the shown images. A Long Short-Term Memory (LSTM) network provided the highest accuracy, with up to 96.6% accuracy. Other architectures, such as Transformer networks and Graph Attention Networks, had prediction accuracies of over 90%.
Keywords: Function Prediction Visual Processing Deep Learning Spike Train Data