Multi-HGTM v0.1.0

Multi-HGTM v0.1.0

Multi-HGTM gives every view of an input its own Graph Tsetlin Machine tower and fuses them with a Tsetlin head that learns cross-view conjunctive rules.

Headline result (5 seeds, Tesla V100)

On a cross-view interaction y = (a1 and b1) or (a2 and b2), which is provably not a sum of per-view scores, clause fusion reaches 0.968 +/- 0.040 where additive late fusion reaches 0.727 +/- 0.077 and the best single view 0.690 +/- 0.022. The clause over late gain is +0.241 (95% CI [+0.16, +0.33]), a win on every one of the five seeds (Wilcoxon p = 0.031). On the complementary routing task clause fusion reaches 1.000. On real three-view breast cancer both fusions beat the best single view and clause matches late, exactly the regime where redundant views leave no interaction to exploit.

What is in this release

• MultiViewHGTM: per-view towers, clause and late fusion, missing-view inference, per-view attribution, and readable cross-view rules.
• A vectorised NumPy multi-class Tsetlin Machine that serves as the fusion head and a GPU-free tabular tower.
• Graph Tsetlin Machine towers (CUDA), the five-seed harness with paired Wilcoxon and bootstrap intervals, the paper PDF, docs, four examples, and 25 tests.

Honest limitations

• The synthetic ceilings are 1.0 by construction; they isolate a mechanism rather than rank a benchmark.
• Towers are trained on the label; stacking each tower as a deep machine waits on the depth credit-assignment fix tracked in the HGTM project.
• A Graph Tsetlin Machine over a naive image grid does not localise digit structure under OR-across-nodes voting, so the real-data study uses tabular towers on breast cancer rather than graph towers on digits.