Artificial Intelligence

AI is an attempt to abstract out and formalize as a well-understood, formally defined process (an algorithm) some behavioral patterns found in Nature.

The main themes are search, classification, planing (map and policy building), model building and probabilistic inference.

The problem with current state-of-the-art AI is that it creates and operates models way over-complicated, disconnected from reality models. The best example is a image classification problems is Machine Learning, where features are based on pixels, not high-level shapes - the way brain process information.

For the brain, human or animal, there are certain lets call it meta-features or shortcuts, which makes pattern recognition much quicker and lest costly process. For example, small children will tell you that a car has eyes. This is because the brain has a structure of recognizing faces and track the direction of a gaze and, in higher animals, evaluating facial expressions. Even dogs has distinct facial expressions and related hard-wired cues.

The big idea here is that way to general computational models, based on automatic feature extraction out of raw pixels will never match an animal performance with the comparable computation costs. Brains are using hard-wired heuristics and cues, evolved according to certain aspects of the shared environment. Brain does not work at a single neuron (or a pixel) level. It, presumably, deals with cues and shapes.

Human made neural networks also must have a structure which reflects certain aspects of physical reality. It is structure that defines functionality of specialized brain centers, not properties of individual neurons.

The specialized structure which does pattern recognition has evolved in a recursive bottom-up process, so it "perfectly" matches certain aspects of reality which shaped these specialized structures. Natural selection made them highly specialized, not universal. This is why we are so bad at imagining what we have never seen or what contradicts the laws of mechanics of our physical environment.

So, building a highly specialized pattern recognition software which uses evolved hard-wired structure (procedures) and trained from observations (experience) models is an example of what AI is.

Beware of abstract bullshit

Animals definitely have an intelligence related to its habitat and social formation. There are no animals who would bump into walls, grossly miscalculated their jumps or mistake an enemy for a friend. What they lack is an abstract language and hence abstract reasoning, which, it seems, important part yet only a part of overall intelligence. Non-verbal knowledge dominates.

It is important to realize that, despite of what idiots claim, most of the knowledge of the world is obtained by trial-and-error, by going-into-unknown and seeing what happens, by remembering the past experience and avoiding what gives a negative feedback. Evolution, at a level of populations and genes, selects what fits the best, and implicitly selects adequate representations of the environment over inadequate ones (this is why we are afraid of snakes and spiders, and we should be).

Most of human machinery, however fine and sophisticated, also has been built by repetitive trials and errors with a very few occasional optimizations form "pure intelligence" which is also primarily based on the very same trial-and-error approach.

So, do try to build an omnipotent god-like absolute intelligence - it does not exist. There is no gods, there is no omnipotence, there is no general intelligence. Everything is bound by environment and environmental and social conditioning. Specialized pattern recognition and adequate representation of environmental constraints and features, and ability to refine and to use this representation is good-enough.

In other words, try to build an animal first.

The key ideas

The key ideas, it seems, are structure and pattern matching. We build a structures inside our brain and the pattern-match or "search" against them.

In the process of a supervised learning (someone must distinguish "right" from "wrong", signal from noise) built-in instincts and feeling provide that "supervision", as well as the structure of built-in "machinery" which reflect the constraints and major characteristics of the environment in which it has been evolved. (day/night changes, temperature ranges, etc.)

Supervised learning is a generalization of a process by which life adapts to the changing (but governed by the universal laws) environments and evolves.

The representation-building and searching sub-systems are semi-independent, although they must agree on the underlying representation. It seems that a "weighted graph" is an appropriate human abstraction. Axons and dendrites form the graph, electrical charges and signal frequencies implement "weights".

There is definitely a process of building-up inner maps or models - representations of the outside world - every higher organism builds and maintains such representations since the moment of its birth. The machinery (execution engine) and hard-wired heuristics (the code) are inherited.

It is important to understand that there is no divine spirit or any similar crap anywhere - the mind is what the body does, and the body in a mechanical machine. The machine has been made out of layers of evolved building blocks, evolved according to the properties, constraints and limitations of the physical environment.

For example, the behavior a living cell exhibits is mostly due to the mechanical *3D-structure* of its proteins and enzimes. The constraints are that there must be water inside a cell, otherwise nothing works. Another constraint is that a cell cannot exist in vacuum, there must be some environment around it which provides energy sources, etc.

The more complex behavior, again, comes from a physical structure. This time it is a structure made out of cells (neurons). It is structure that defines the behavior. So, it seems, that information must have a structure. It is then used as both code and data by higher-level processes.

We are applying state of the art Machine Learning techniques to design and implement autonomous intelligent agents capable of improving their knowledge without being reprogrammed.

The goal is to replace unreliable and error-prone people in day-to-day business processes with programs that outperform them. Real-time electronic stock trading is the obvious example.

This is already mainstream approach in the fields of Medical Diagnosis, where programs, which are maintaining and improving their knowledge bases and sets of rules (trained on the past events and its own actual experiences) are outperforming an average doctor.

We are designing and implementing such autonomous intelligent agents for various kinds of businesses.

Last modified 3 years ago Last modified on Feb 27, 2018, 11:44:48 AM
Note: See TracWiki for help on using the wiki.