Vesper Live

Real-time visualization of a Ternary Gated Network training on CIFAR-10. Each neuron operates in three distinct states: excitatory (+1), neutral (0), or inhibitory (-1), enabling sparse and highly efficient computational routing.

Researcher's Note: "Next up: a five-run experiment series on ImageNet-1K. v18a reproduces the v3 baseline (~67.9% target), then v18b, v18c, and v18d each isolate one asymmetry tweak (TernaryReLU, weight L1, margin loss). v18 final combines whatever helps. About 70 hours total. Email ternarymac@gmail.com with suggestions."
Connecting...
Network Topology: Ternary Neuron States
Excite Neutral Inhibit

Ternary State Distribution

Gate Confidence

Higher gate values mean the block trusts its transformation over the skip connection.

Training Progress

-- Val Acc-- Val Top-5-- Loss┈┈ ResNet-18

Architecture

Model
Vesper
Parameters
-
Mechanism
Ternary Gated Network
Stages
-

Each neuron classifies its activation into one of three distinct states: excitatory, neutral, or inhibitory. Dynamic gates control information flow between skip connections and transformed features, allowing the network to route computation based on input confidence and architectural sparsity.

Auto-refreshes every 30 seconds · Project FARSCAPE