Anwar Debes

Watch a Tsetlin Machine learn

Tiny automata vote on which inputs matter. Nobody tells them the rule is XOR. They find it, and they can tell you why.

Included input Included negated input Excluded The glass plane is the include threshold

Speed, Label noise in the training data, Clauses

Progress

Epoch0

Test accuracy on clean labels50.0%

What it learned, in plain logic

It found XOR on its own: x₁⊕x₂ as two clauses for and two against, and it ignores x₃ to x₆ completely.

Explain one prediction

What you are looking at

Inputs x₁ to x₆ are random bits. The label is x₁ XOR x₂, the other four inputs are pure noise, and a share of the training labels is flipped on purpose. Each bar is one Tsetlin automaton deciding whether one input, or its negation, belongs in one clause. Bars above the glass plane are included. Clauses marked + vote for 1, clauses marked − vote for 0, and the majority wins.

No GPUs were harmed. This is a real Tsetlin Machine running in plain JavaScript on your CPU. Anwar Debes, research map.