Perceptual Learning (under revision)
My approach to understanding intelligence is guided primarily by the following three principles.
If I had to choose a single word to characterize intelligence, it would have to be timing. Timing is an essential aspect of animal intelligence, from sensory perception to motor control. The ability to anticipate and react to the temporal evolution of various phenomena is essential to survival and adaptability.
Timing is not everything, however. Fundamental to all things is the powerful Principle of Complementarity (PC), otherwise known as the yin-yang principle in Chinese philosophy. It stipulates that all entities, phenomena and concepts come in pairs of complementary opposites. Here are a few examples relevant to temporal intelligence:
While temporality remains the main thrust of this article, I must emphasize that its importance to the brain's operation and organization is secondary to complementarity. The PC imposes a profound symmetry on all aspects of intelligence, as we shall see.
If it's not simple, it's wrong! Over the years I have come to trust in the correctness of this principle. I believe that complexity stems from simplicity and that any attempt to understand intelligence (or any complex whole) will fail unless one's model is simple at the core.
Perception is the identification or classification of sensory signals according to their temporal (or causal) correlations. In this section I present a model of perceptual learning based on the arrival times of signals. Two things to note here are the essential role played by the PC and the fundamental simplicity of it all. Simplicity is necessary to scalability and generality.
I define a signal as as a spike, i.e., the discrete firing of a neuron in response to either an internal or external event or stimulus. Thus a signal is a temporal marker that indicates that something just happened. All signals look exactly the same and carry no information about the nature of their origin. In other words, signals are not symbols to be deciphered or decoded by receiving neurons. The only thing that distinguishes one signal from another is its arrival time.
All sensory events are the result of physical changes in the environment. Examples are the motion of a light dot across the retina, or a change in brightness, color, frequency, temperature, volume, etc… I classify events into two complementary types: positive and negative. A positive event is the onset or positive transition of a phenomenon. A negative event is the outset or negative transition of a phenomenon.
A sensor is a special neuron that reacts to a specific event. Based on the types of events described above, we can design sensors to be either inclusive or exclusive. An exclusive sensor transmits signals as a result of either positive or negative phenomena but not both. An inclusive sensor, on the other hand, fires in response to both positive and negative phenomena. Thus an inclusive sensor can be though of as a change detector. This begs the question: why must there be both inclusive and exclusive sensors? The answer is that, aside from the obvious complementarity aspect, both types are required for a full accounting of causal correlations in the environment.
Note: Although it is known that animals have retinal ganglion cells that respond to both onsets and offsets of visual stimuli, I have found that it is not necessary to have inclusive sensors in a well designed system as the network is able to find inclusive correlations automatically.
There is an implied logic to sensory signals, one that is external to the intelligent agent and is assumed a priori. The logic is so obvious and simple as to be easily overlooked. Sensors react to specific phenomena or changes in the environment. The offset of a given phenomenon is assumed to follow the onset of the same phenomenon. In other words, no phenomenon can start if it has already started or stop if it is already stopped. As an example, if the intensity of a light source increases to a given level, it must first go back below that level in order to increase to that level again. The logical order of events is implicit in all sensory systems. That is to say, it is imposed externally and no special sensory mechanism is needed to enforce it. This may seem trivial but its significance will become clear when I discuss motor coordination in the effector layer. As the Principle of Complementarity would have it, the effector layer is the exact mirror opposite of the sensory layer. That is to say, effector logic is the complementary opposite of sensory logic. I like to say that effecting is sensing in reverse. One must never underestimate the importance and power of the obvious.
How about the sensing of smoothly changing phenomena? Does that require the use of continuously varying signals? The answer is no. Signals are always discrete and they should reflect both positive and negative phenomena. The most effective way to communicate analog stimulus levels is to encode them in the relative arrival times of discrete signals. The first signals to arrive are the most salient ones. Several spiking network researchers, such as Simon Thorpe, Arnaud Delorme and Rufin van Rullen, have done excellent work in this area. I must caution the reader that their work is more in the area of static image analysis whereas temporal intelligence requires a dynamic environment.
Obviously one cannot have an infinite number of sensors for every infinitesimal change. What is important is the sampling resolution of the available sensors. They must react to changes fast enough to accommodate the design goals of the system. The sampling resolution of most biological sensors is about 1 millisecond. Bats have auditory sensors with much higher resolution, apparently on the order of microseconds.
There are only two possible types temporal relationships between two discrete signals: they are either concurrent or sequential. Sequential signals are called predecessors and successors. An event will have a different temporal or causal meaning depending on whether it arrives before (predecessor), after (successor), or at the same time as another event. As we shall see, every signal processing task involves the use of a specific temporal relationship.
It goes without saying that two sequential signals are not identical: one is the predecessor and the other is the successor. But what about concurrent signals? Are they equal in function and purpose? The answer is no. Concurrent signals can be divided into masters and slaves. As I mentioned earlier, signal processing is the guiding of signals through specific pathways. A signal processor or neuron is like a gate keeper that guards the entrance to a pathway. The difference between a master and a slave is that the gate can only be opened if the master has arrived.
There are several corollary issues related to masters and slaves. For examples, a slave cannot serve more than one master; a signal can be either a master or a slave but not both. A problem immediately arises: what determines whether a signal should be a master or a slave? I will address this question when I discuss association neurons in the next section.
In sequence processing, successor signals are the masters and predecessors are the slaves. The gate keeper opens the gate to let the master through only if the master was immediately preceded by its slave. The slave is said to prepare the way for or announce the arrival of its master. Their arrival times are contiguous. Note that there is a one to one relationship between a predecessor and a successor. This means that a sequence detection neuron must be a two-input neuron. In Animal, two-input sequence neurons are used for separating sensory signals from sensory streams into parallel pathways.
There can be as many input connections to a coincidence neuron as needed. The neuron fires if most of the input signals arrive concurrently. However the neuron cannot fire unless the master is present. The idea is that the master signal expects most or all of the slaves to arrive when it does. A slave connection is like an associate. If it does not arrive with its master, the association is strongly weakened.
Temporal perceptual learning consists of finding temporal relationships in sensory signal streams. In a non-random environment, statistically salient temporal correlations can be found by monitoring the arrival times of sensory signals. This is the job of the perceptual network described in the next section.
Next: Perceptual Network
©2004-2006 Louis Savain
Copy and distribute freely