Active Structure vs Machine Learning

Here are some quotes from Sun Tzu (500 BC):

“In the midst of chaos, there is also opportunity”

“Engage people with what they expect; it is what they are able to discern and confirms their projections.
It settles them into predictable patterns of response, occupying their minds while you wait for the extraordinary moment — that which they cannot anticipate.”

“Rouse him, and learn the principle of his activity or inactivity. Force him to reveal himself, so as to find out his vulnerable spots.”

“If quick, I survive.
If not quick, I am lost.
This is "death.”

“Therefore, just as water retains no constant shape, so in warfare there are no constant conditions.”

The advice, from a brilliant military strategist 2500 years ago, is “Learn to rapidly adapt, understand your enemy, appear to be predictable, while not being so”.

Active Structure AGI is an undirected structure that can modify itself based on activity within the structure, and extend itself, if necessary, in minutes. It is by no means a copy of a real neural network, but it can do the things a real neural network can do, and it has an English language interface. Its innate undirectedness gives it properties that only come from many layers of real neurons.

A Simple Example 

At rest, a plane’s fuselage rests on its undercarriage, and the fuselage supports the wings. On take-off, the wings support the fuselage (forces flow in different directions, and through different objects), and the undercarriage is folded up and stowed away – it effectively ceases to exist. Active Structure handles reasoning about physical things, like an undercarriage,  and abstract things, like “opportunity” or “recklessness”, or “loyalty”.

Don’t Fight the Last War

Machine Learning uses an Artificial Neural Network (ANN). It is in fact a “trained” directed resistor network with inputs and outputs – the epitome of what Sun Tzu warns against. It has no notion of momentum, of inertia, of capacity, of self-repair, of a new thought expressed in language.

An artificial neuron sums numerical values (weights) together, and provides an output, which may be to another layer. What an artificial neuron doesn’t do is all the other things a real neuron can do – it can excite itself, it has memory, it can obviously model everything you can think about. It can handle concepts like loyalty and safety, and in large enough numbers it can provide intelligence – rapid adaptation to changing circumstances, if you prefer.

Let’s be frank – an ANN is earthworm intelligence. It may be suitable for tracking the number of green cars you make each month. It is not remotely suitable for battle, where circumstances can rapidly change from what was envisaged. Hardwiring a direction is the wrong thing to do - the connections should be initially undirected, so the value of their contribution can be assessed dynamically, or removed completely if that bit just got blown up.

The vast chasm between a real neural network and an ANN can be seen in the two neurons that keep you alive. The two neurons excite each other, with the level of stress in your body changing the firing rate from fifty beats a minute at rest, to one hundred and eighty beats a minute at maximum stress. Such a simple device is easy with real neurons, but is impossible if all you have are resistors (yes, they are buffered, but they are still only resistors, with their connections changing the weightings to modify the value of an output, not to switch it or amplify it or delay it). Imagine you are tasked with building a mobile phone only with resistors, and the problem should be clear. 

How do you do all the other stuff? 

An ANN has proved to be a very poor base on which to build an autonomous vehicle driving on a highway (or a car park – who would expect a business jet to be in a car park, and have a belly just below the height of a car?) – imagine how much poorer it will be when someone is deliberately trying to smash into you, or force you off the road.

While directed, real neurons can excite themselves, and give the appearance of undirectedness, leading to the ability to turn things around in your head, to seize an opportunity when it arises, to patch around failing equipment, to survive.

The ANN is the worst possible thing to take into battle. It has been “trained” by someone, possibly years before, who was tasked with determining possible scenarios that could be crystallised into weightings. This is very much like fighting the last war, or reasonable projections of that war. But you are fighting an intelligent adversary, who will understand the experiences that have shaped your thinking, and will likely do something you hadn’t thought of. 

An ANN sort of works as a way of finding a cat in a picture (but not a dog, they are too variable), but not for more serious uses. The many useful properties of real neurons have been lost. When a bird sees that the angle between it and another moving object remains stationary when it flies, it will accelerate or turn away to avoid a collision. That simple act relied on memory - what was happening a few moments ago. All the ANN could report would be the data "Death of bird".- it doesn't do prediction (which you imagine would be sort of necessary for autonomous vehicles).

It was a serious error to call a directed resistor network an Artificial Neural Network – people have grown up believing the label on the can, and not understanding how far it is from reality. It is doubly insidious, in that much of a human’s thinking is done unconsciously, because of the Four Pieces Limit, so we are unaware of the huge amount of processing we do this way – processing that is surfaced in Active Structure, making all of it modifiable by events.

Part of what Active Structure is about is a means of breaking the Four Pieces Limit.

One obvious use is analysing a complex specification, which may have contributions from a dozen different specialists, each of whom can barely understand the technical language of the others. This ability to integrate spills over into the procedures that define a particular role, where the machine has some insight into the roles around it, and permits coordination under stress.

 

Active Structure©