It would be pretty interesting to use a neural network for AO! It's a good coincidence you linked to SSIM based adaptive sampling, I'm actually using the research you linked to: I implemented the OptiX API for SSIM prediction. So far, we've used it for choosing where to send more rays, but I'm not aware of anyone using it to adapt the max distance for AO rays. That might be fun & fruitful to try!
I'm not certain, but I think in theory that capping the max distance will always bias the AO (to the extent that AO even makes physical sense). You can reduce the bias by mixing lengths, but I'm not sure if you can ever fully compensate for it. More importantly, it's always (theoretically) biasing to separate AO shading from the normal surface shading & lighting calculation. AO is a beautiful simple hack, but I'm not sure it will survive the future, especially as ray and compute budgets increase.
I know I've seen other AI attempts to learn AO & global illumination, just googling a bit:
I'm not certain, but I think in theory that capping the max distance will always bias the AO (to the extent that AO even makes physical sense). You can reduce the bias by mixing lengths, but I'm not sure if you can ever fully compensate for it. More importantly, it's always (theoretically) biasing to separate AO shading from the normal surface shading & lighting calculation. AO is a beautiful simple hack, but I'm not sure it will survive the future, especially as ray and compute budgets increase.
I know I've seen other AI attempts to learn AO & global illumination, just googling a bit:
http://theorangeduck.com/page/neural-network-ambient-occlusi...
https://www.youtube.com/watch?v=z_zmRWxU-PY