Sorry, my fault: I don't mean that PAC-Learnability means that neural nets memorise their training instances. That's more my interpretation of their observed be

Sorry, my fault: I don't mean that PAC-Learnability means that neural nets memorise their training instances. That's more my interpretation of their observed behaviour, if you like. What I meant was that PAC-Learnability doesn't assume any ability like reasoning, and really no other ability than er, PAC-Learnability. There's a debate, of course. I like to point to Domingos' paper: Every Model Learned by Gradient Descent Is Approximately a Kernel Machine https://arxiv.org/abs/2012.00152 https://arxiv.org/abs/2012.00152 With the full understanding that it's just one paper.