ideas.md

Add an abstract domain tracker to simulator, to determine how good the simulated-hardware-calculated domain actually is

For networks made with only matmul, bias and ReLU, 'abstract' can simply mean a polytope which is relatively easy to keep track of. Open question as to how to communicate this information - perhaps simply the ratio of the volume of the abstract domain to the volume of the hardware-calculated domain

Keep track of a statistical distribution of the domain of the result instead of hard bounds

Use more-abstract-domain bounds (octohedral etc.)

Mitigates problem of non-relational n-cuboid domain growing vastly bigger than the polytope it bounds Suggested by Ka Wing Li

Process fragments as 2-bit exponent + sign

Is there any possible advantage to this?