Partnering with Researchers at UCF, Harvard, and Carnegie Mellon to Develop New Forms of Memory Storage and Recall
The way in which the brain stores and retrieves memory and information nonlocally across the tissues differs significantly from the way in which memory is stored and retrieved digitally. Developing and studying new forms of memory storage and recall are the goals of Project LATCH and Project TOGGLE.
I have had brief email discussions with Dr. Tononi many years ago, and I suppose his "Global Workspace" idea makes sense in that you need a mechanism to broadcast (and bind) information and memory nonlocally across the tissue to perform the equivalent of backpropagation and perceptual binding. The one issue I have with Tononi's idea is that he asserts this theory but there is little empirical investigation as far as I'm aware, and that's something that I would like to change. The Blue Brain project has mapped the neural networks in tissue to a high dimensional lattice and so achieving perceptual binding could be akin to the NP-hard shortest vector problem (SVP) or learning with errors problem (LWE). This was estimated by researchers like Dr. Tsotsos and some have proposed possible classical explanations for how the tissue can achieve this (like predictive coding), but I find them to be unconvincing. My view is that regardless of how you configure a classical neural network, the physics is fundamentally different and incomplete.
The problem with the idea that you can scale an arbitrary neural network up enough to reach some complexity threshold beyond which you achieve "consciousness," or that consciousness itself is a gradient of complexity is that in my view consciousness and perceptual binding is an NP-hard problem, and no classical or even quantum computer in themselves are known to be able to break lattice problems in a tractable amount of time. As you scale a conventional neural network it reaches thermodynamic limits of scale beyond which there is a point of diminishing returns on scaling. The brain operates on only 20 watts while these AI chatbots require their own nuclear power plants for the data centers and GPU clusters for training, and ultimately require constant interpretation and training by humans-in-the-loop as when they are trained on their own outputs quality decays rapidly.
The idea I have is that the way in which information is bound in the brain or "broadcasted" across the tissue to achieve the learning updates and perform the equivalent of backpropagation is that (possibly topologically protected) spin entanglements evolve (or are "orchestrated" as Penrose puts it) to a point of saturation defined by entropic limits (related to gravity through the Ryu-Takayanagi formula) beyond which there is a cascade of superradiant biophotons which have a property called orbital angular momentum which broadcasts the collective spin information across the tissue. This mechanism is what we wish to explore with projects BLUEBIRD and SIDECHANNEL.
In a sense, this is related to the idea that you reach a threshold of complexity beyond which a system can become "conscious," but the difference is that in a classical and conventional neural network there is no perceptual binding - the mechanism to actually broadcast the error signal and also to bind the information together to give the perception of time is missing or brute forced. You also do not necessarily need a very complex neural network - in fact there is research that shows single celled paramecium display complex learning behaviors and memory storage and recall.
Comments