> Our artificial Feynman Machines have several interesting properties which distinguish them from existing Deep Learning and similar systems. In particular, due to the much higher density and locality of processing, a Feynman Machine-based system can perform at least comparably while dramatically reducing the footprint in computational power, training data and fine-tuning. Feynman Machines can be arbitrarily split into modules distributed across clusters and the Internet, and systems running on low power devices such as phones can be dynamically and robustly augmented using low-bandwidth connections to larger networks running in the cloud. Models can be trained on powerful infrastructure and transferred for operation and further custom learning on smaller devices. Importantly, the same architecture can be used as a component in unsupervised, semi-supervised, fully supervised and reinforcement learning contexts. A variant - the Routed Predictive Hierarchy - is described, which allows a Feynman Machine to directly control a traditional Deep Learning network by switching it to use spatiotemporally-selected subnetworks.
This is the abstract of the previous paper that detailed the theoretical basis [0]:
> Reverse engineering the brain is proving difficult, perhaps impossible. While many believe that this is just a matter of time and effort, a different approach might help. Here, we describe a very simple idea which explains the power of the brain as well as its structure, exploiting complex dynamics rather than abstracting it away. Just as a Turing Machine is a Universal Digital Computer operating in a world of symbols, we propose that the brain is a Universal Dynamical Systems Modeller, evolved bottom-up (itself using nested networks of interconnected, self-organised dynamical systems) to prosper in a world of dynamical systems. Recent progress in Applied Mathematics has produced startling evidence of what happens when abstract Dynamical Systems interact. Key latent information describing system A can be extracted by system B from very simple signals, and signals can be used by one system to control and manipulate others. Using these facts, we show how a region of the neocortex uses its dynamics to intrinsically "compute" about the external and internal world. Building on an existing "static" model of cortical computation (Hawkins' Hierarchical Temporal Memory - HTM), we describe how a region of neocortex can be viewed as a network of components which together form a Dynamical Systems modelling module, connected via sensory and motor pathways to the external world, and forming part of a larger dynamical network in the brain. Empirical modelling and simulations of Dynamical HTM are possible with simple extensions and combinations of currently existing open source software. We list a number of relevant projects.
Edit:
Apparently, the project is called Ogma and they have a youtube channel. Here's their video [1] of a toy car learning to self-drive after a few laps of a person driving it with an xbox controller. It runs on a NanoPi Duo. Not sure if that's noteworthy or not.
https://arxiv.org/abs/1609.03971
> Efforts at understanding the computational processes in the brain have met with limited success, despite their importance and potential uses in building intelligent machines. We propose a simple new model which draws on recent findings in Neuroscience and the Applied Mathematics of interacting Dynamical Systems. The Feynman Machine is a Universal Computer for Dynamical Systems, analogous to the Turing Machine for symbolic computing, but with several important differences. We demonstrate that networks and hierarchies of simple interacting Dynamical Systems, each adaptively learning to forecast its evolution, are capable of automatically building sensorimotor models of the external and internal world. We identify such networks in mammalian neocortex, and show how existing theories of cortical computation combine with our model to explain the power and flexibility of mammalian intelligence. These findings lead directly to new architectures for machine intelligence. A suite of software implementations has been built based on these principles, and applied to a number of spatiotemporal learning tasks.
YMMV