Killer Robots

Black Swan wrote: »

Can we speak comparatively? How does AI compare to random variation, differential reproduction, and evolution? Similarities? Differences? AI now in its single cell stage today, and a primitive cell at that? Will cells combine as they evolve, with many early variations that are not fecund, but an extremely rare variation that becomes a fecund hybrid? Or perhaps these comments are just metaphorical nonsense?

My AI ramble/rant

AI is a buzz word, always in the press, and companies run to commercialise it, even when alternative, established methods exist. A lot of what is termed AI is simple pattern matching, yet we use the word intelligent to describe it.

The principles and techniques have been around for a long time. The model we have of the neuron seems to work well, and is established. The methods we use to train networks of these neurons has settled (genetic algorithms and back propagation). The real progress (besides finally having the CPU power) has been in structuring the networks to give them the opportunity to excel, by creating software that uses the networks to create their own data. Deepmind's alpha zero is a fantastic example of this, a great bit of software surrounding a bunch of standard neuron models, that is able to master any two player game (with perfect information). So whereas we have a good model of the neuron, and how networks of them can be structured to excel at different types of problems, what we dont have is an understanding of how the brain manages networks and builds the "software" around them.

Most of my recent experience has been with genetic evolution simulations and I can honestly say it has helped me (real world) understand peoples behaviour better. It has also helped me understand my wife's, post chemo, symptoms. A lot of the best work in AI has come from studying and trying to understand the biology of the brain, with Biology knowledge as important as computer coding experience.

I often see things in simulations that apply to the real world, traits, afflictions, skills, etc. Always the most relevant results come when natures own rules are modelled as closely as possible.

As for AI killer robots, I do think they will be a major factor in a war soon.

Something else that is often overlooked is that at the moment we implement neural networks on a digital system, computers. Real performance will start to happen when we have mastered the analogue side. All stages/operations in a neural network have an analogue equivalent, which can be implemented in a simple analogue circuit using electronic components. When I first studied higher education we only had limited access to a terminal on the single college computer (early 80s), but there were many analogue computers. These analogue computers were widely used in industry for engineering modelling, and as there was no "clock" they were instant.

It dawned on me some years ago that when you find a neural network solution, building an analogue equivalent allows it to operate instantly. I looked online and discovered that this is an old idea, and already implemented.

A state of the art missile has a digital guidance system which gets it to the target, but in the last few moments (seconds), guidance is handed over to an simple analogue computer/circuit that is able to make changes to course without having to wait and calculate it. Such circuits are high tolerance analogue implementations of a simple neural net developed on a digital platform.

At the moment there is huge investment making what we have faster and more parallel, and students are steered towards code libraries (Tensor Flow etc) from the big IT players. There will be a magnitude leap in AI when hardware manufactures produce analogue neurons on chip, and this is now happening...

EDIT: Sorry the reason this is relevant for killer robots and advanced weaponry instant compute speeds are required for many situations (fire and forget bullets for example)