Just wanted to post a pointer to a new highly parallel FOSS AI engine for small, distributed/CUDA-enabled linux systems. Apologies for cross-forum posting; this got a bit long and it would have been a daunting task to try and rewrite, but I thought forumgoers might be interesting. It’s also been posted to comp.ai though.
This is a nearly-finished CUDA-enabled source code I wrote mostly for myself for an open source 2D strategy game engine in long-term development; but anyone is welcome to use or help develop! It applies quantum statistical mechanics closely related to density operators and Bayesian dynamics (see e.g. my paper here, also here) to the problem of probabilistic quantum logic, also employing some aspects of quantum computation theory to parse and interpret a formal grammar useful for expressing symmetric relationships in data.
Essentially, its method involves the use of variational calculus to predict the expectation value associated with the possession of a finite set of attributes, given accurate knowledge of the entropy associated with all attributes, where each attribute is represented numerically as a fourier dual of the 2-dimensional phase space. Information transfer is presumed a maximum-entropy Poisson process conforming to a principle of stationary paths; this mirrors known phenomena in neuroscience.
Because Bayesian information theory can effectively be expressed as a probabilistic effect resulting from an inherent ambiguity in the truth of a proposition that corrects itself over long periods of time (see e.g. Jaynes’ dice-testing example in this paper), we allow the phase space to be, over each meaningful chronological interval, the set of propositions held. The attribute is then associated with an eigenstate.
Currently it’s in final stages of debugging stackless python networking, but basically done other than that. Please, if you find this post interesting, feel free to keep following the project as it hopefully continues to grow and expand!