The Brain: Universal Invariant Recognition, Part I

Part I, II, III, IV, V, VI, VII, VIII

Abstract

Right now I am totally burned out on parallel programming and COSA. The only remedy is for me to go back to my first love, artificial intelligence. One of the AI problems that fascinate me is invariant pattern recognition. A complete and robust solution to this problem is one of the holy grails of AI research. So far, most of the approaches that I have come across are severely restricted to a narrow problem domain such as scale, rotation or position invariance. In this article, I will argue that a truly universal solution can be realized by taking advantage of the predictability of discrete sensory signals.

The Amazing Brain

I am forever amazed by the brain’s ability to recognize a familiar object independent of its size, distance, orientation or position in the visual field. It may seem trivial on the surface but it is truly remarkable that the brain can figure out that the different views presented by a moving object represent only one object. It is even more remarkable when one discovers that the brain’s invariant recognition prowess is not limited to visual phenomena. For example, most of us have no problem recognizing a familiar musical tune irrespective of its volume, tempo, tonality or scale. Consider also that, even though neurons are notoriously slow processors, the brain can perform its recognition magic almost instantaneously, on the order of tens of milliseconds. This means, obviously, that the brain does not rely on lengthy or complex sequential computations. Given the vastness of the sensory space, there is no question that there is a huge amount of parallel processing taking place. I believe that the brain uses a recognition mechanism that is both simple in principle (it is simple because it must be fast) and universal in scope.

[As an aside, my fascination with the brain is what drives my interest in parallel computing. I am convinced that the majority of computer applications in the future will be autonomous intelligent systems. We will need extremely powerful and scalable parallel computers in order to emulate even a fraction of the brain’s capabilities.]

The Fundamental Language of Cognition

Knowing that the brain uses many elementary parallel processors or performs very few sequential computations is important but it is not enough to reveal the underlying mechanism of invariant recognition. We need more clues. One thing is certain; we know from observing humans and animals that the brain can learn to recognize new patterns. Learning presupposes that sensory information contains non-random patterns. That is to say, sensory signals have temporal correlations that can be discovered because the patterns repeat themselves often enough to stand out above mere noise. Once a correlation is learned by a neural structure, the same structure can subsequently be used to recognize all future occurrences. Part of my intelligence thesis is that temporal correlations comprise the fundamental language that the brain (or any other intelligent system) uses to make sense of its sensory space. In the past, I have argued that the important thing to consider in AI is not what any individual sensory signal represents (e.g., the onset or offset of illumination) but how it is related temporally to other signals.

Fixed Sequential Correlations: Early Sensory Learning

There are two basic forms of temporal correlations: signals can be either concurrent or sequential. In addition, correlations can be either fixed or variable. Signals with fixed correlations always retain a given temporal interval. Sequential correlations are detected and learned early in the sensory cortex of the brain. Creating a network to learn fixed sequential correlations is a simple matter of choosing a fundamental interval and designing the neurons to discover predecessor and successor inputs that satisfy the interval. The network should have a search mechanism that tries as many input pairs as possible. In my experimental spiking neural network, I use a 10 to 1 probability ratio for sensory learning. That is to say, if a predecessor synapse satisfies the interval at least once in every ten tries, the correlation is deemed to be above random noise and the predecessor connection becomes permanent. Subsequently, the neuron will fire every time the two signals arrive in close succession. The system can use massive feedback to recognize a huge number of features over multiple time scales.

My research in invariant recognition has led me to conclude that the sensory cortex only deals with fixed sequential correlations. I believe that concurrent correlations are processed afterwards in the temporal cortex by the same mechanism that handles variable correlations (variable sequence learning). That is the more interesting part of perceptual learning, in my opinion, because both concurrency and variable correlations are essential to invariant recognition. I’ll explain what I mean in Part II of this article.