Hybrid WNNs + LLM Embedding Recommender
A context aware, compute efficient recommender trained on individual user data.
v0.1.0 MNIST Benchmark
788 neurons: 1-to-1 channels <> 784-4 input-output neurons
A demo of training efficiency, getting >60% on ~120 training samples (instead of 60k samples). Also useful to get a feel for how the connections between neurons changes performance.
ARC-AGI Using WNNs
A unique attempt using layers of 30x30x4 neurons
The ARC challenge represents a new dataset for true general intelligence; ARC tests the ability of an agent to abstract and learn an implicit rule given a few examples. In this way, ARC combines linearities within ARC tasks with non-linearities across tasks, since each has a different implicit rule to be abstracted.
Useful to get a feel for how the connetioeen neurons changes performance.
Netbox Device Discovery: context-aware autocomplete for relational data
40 neurons: 3-to-1 channels <> 30-10 input-output neurons
Hitting 85%+ accuracy with predicting roles of new devices given as little as 70 examples of previous devices.
