Repository files navigation

A neural network library for learning boolean-valued, discrete functions on GPUs with gradient descent.

The library is implemented in Python using theFlax andJAX frameworks.

Questions? Ask @Z80coder

Lossless hardening with ∂𝔹 nets. I. Wright. In"Differentiable Almost Everything: Differentiable Relaxations, Algorithms, Operators, and Simulators", ICML 2023 Workshop, Honolulu, 2023.

Draft paper:"∂B nets: learning discrete functions by gradient descent" (April 2023).

Neural network research with the Wolfram language (30 mins).

∂B nets quick overview (30 mins).

∂B nets overview (1 hour).

The working prototype was implemented in Wolfram. The demos below were snapshots of work-in-progress.

• Neural logic nets (15m)

• The Soft-NOT operator (10m)
• The Soft-AND operator (10m)
• The differentiable Hard-AND operator (17m)
• The differentiable Hard-OR operator (5m)
• The differentiable Hard-MAJORITY operator (13m)
• The hardening layer (11m)
• The hardening operation (19m)
• A classifier architecture (20m)
• Neural logic nets (15m)
• Learning XOR (parity) (10m)
• Numerical regression (23m)
• If-Then-Else neuron (23m)
• Neural conditions and actions (24m)
• Neural decision lists (15m)
• Boolean logic nets and MNIST (18m)
• Neural logic nets for differentiable QL (30m)