Would this be a good candidate for the neural Turing Machine architecture? Where the goal is to learn the memory access patterns of an internalised memory bank ...
If the shell commands were vectorised into a formalised encoding and then this encoding was stored in the neural memory bank and the objective function of the neural network was to learn the access patterns (reads and writes) of this memory, then it would have enough contextual awareness (through access patterns) and internal knowledge (through The formalised command vectorisation) to know that if I “cd (somewhere)” run a few commands then a likely command might be “cd (back)” without supervised training data?
All learnt through essentially observation by the network?
And if so, learning the symbolic nature of the directory name, and the relationship to “cd” and how that is accessed would be a rudimentary form of abstract reasoning, no?
I wish someone would build that and force you to use it for the next six months.
NTMs and similar are very slow at learning, and would (e.g.) not pick up on the fact that you've created a directory until you've interacted with it a couple thousand times.
If the shell commands were vectorised into a formalised encoding and then this encoding was stored in the neural memory bank and the objective function of the neural network was to learn the access patterns (reads and writes) of this memory, then it would have enough contextual awareness (through access patterns) and internal knowledge (through The formalised command vectorisation) to know that if I “cd (somewhere)” run a few commands then a likely command might be “cd (back)” without supervised training data?
All learnt through essentially observation by the network?
And if so, learning the symbolic nature of the directory name, and the relationship to “cd” and how that is accessed would be a rudimentary form of abstract reasoning, no?