Neural Assembly Pattern Learning

Overview This platform, which is built using Streamlit, is meant to make computational neuroscience principles tangible. It allows you to see how neuronal assemblies, by using only local learning and competition, create stable co-active structures capable of recognizing certain patterns. What Does It Do? The process of creating assemblies involves exposing a simulated neural network to visual patterns during several learning iterations. Neurons firing together create a connection between each other, and thus, distinct assemblies are formed through Hebbian-like plasticity. Some features include: Multidimensional Architecture: Simulated neurons are distributed across three different brain areas (Area 0, Area 1, and Area 2) k-cap Representation: A special mechanism which allows choosing the top-k most active neurons at any point of time Real-Time Visualizations: A set of heat maps representing the intensity of activity of neurons in brain areas User Interaction A user can: Train assemblies using one of three datasets: synthetic patterns, MNIST digits, or Fashion-MNIST items Control parameters such as learning iterations, noise strength, and recognition threshold Use pattern completion to apply occlusion on the pattern and see how it is restored by assemblies Check Distinctiveness through visualized overlap matrices Analyze recognition confidence by looking through the bar charts representing recognition