Query node High-attention neighbor Low-attention / other edges
⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.

Graph Attention Networks: Learned Attention Visualizer

Graph Neural Networks that use a fixed sum, mean or max to combine neighbor information treat every neighbor as equally important — Graph Attention Networks (GAT) instead learn, per edge, exactly how much weight each neighbor deserves. This simulator computes real GAT attention coefficients on a small 3D graph: pick a query node and watch a live LeakyReLU-scored, softmax-normalized attention distribution light up its incoming edges, switch between independently-learned attention heads, and drag the LeakyReLU slope and softmax sharpness controls to see the distribution sharpen or flatten in real time, backed by a sum-check and entropy readout that confirm the softmax normalization on every recompute.