Accuracy at every stage, from float ANN to the Akida simulator (3 seeds)

Akida's events: non-zero inputs per layer, and the work that remains
8-bit Akida4-bit Akidadense (every connection)

Energy per recording, counted with NeuroBench (same constants as Counting Energy)

DVS Gesture: what is in the dataset

recordings per gesture

lighting conditions

recording length

training set (subjects 1–23)test set (subjects 24–29)

The 11 gestures as the event camera sees them (one test recording each, 1.8 s, looping)

Blue: ON events (a pixel got brighter), orange: OFF events (darker); brightness = number of events in the 60 ms frame (log scale). Shown at 64 × 64 for clarity; the network gets 32 × 32. Under each: what the 4-bit Akida model (seed 0) predicts after averaging the 30 frames. The panel further down steps through any of them frame by frame.

One recording, frame by frame: camera, network input, and Akida's answer

the camera (64 × 64 shown)

what the network gets: 32 × 32, uint8 counts

ON

OFF

ON + OFF