Program synthesis in neuro-symbolic differentiable interpreters.

Differential interpreters in which the weights of the machine can be interpreted as a program. I think this is different in spirit from ∂4, which corresponds to hard sketches here with holes that are whole neural networks.

Baby steps (see the log):

1. inc_stop
2. dup_add
3. reg_jmp
4. sub

• Notes
• Run logs

• JAX
• Haiku
• Optax

• Chapter 3 (∂4: A Differentiable Forth Interpreter) of Matko Bošnjak's Ph.D thesis, On Differentiable Interpreters, UCL, 2021.

Many thanks to Matko Bošnjak and Rob Zinkov for discussions, insights, suggestions, and pointers.