Memory (under revision)

Note: Due to recent developments, this page is now obsolete. Stay tuned for an update.

Types of Memory

In order to adapt to a changing environment, an intelligent organism must base its decisions on several concurrent sources of information. The purpose of memory is to afford the organism with up to date information about the past, the present and the anticipated future. The organism must be able to quickly process the incoming perceptual stream and, depending on its correlations with past and anticipated events, make appropriate behavioral decisions. The following table lists the required types of events and the corresponding memory subsystems used to generate them:

As seen in Animal's network diagram below, reactive and anticipatory memories are part of a two-layer loop comprised of short-term memory (STM) and long-term memory (LTM). Blue arrows represent master and slave inputs. Red arrows stand for master inputs and green for slave inputs. The most striking aspect of the diagram is its left/right symmetry, a direct consequence of the Principle of Complementarity. Notice how short and long-term memory form two tightly integrated recurrent loops. The two loops are complementary and are each composed of a short and long-term memory module. The left loop is said to be reactive, i.e., it generates signals associated with events already happened. The right loop is said to be anticipatory because it generates signals associated with events that are likely to follow.

Recognition Memory (RM)

Recognition memory is really the perceptual network itself. It consists of the sensory, separation and coincidence layers. It remembers or recognizes learned event patterns whenever its sensors are presented with those patterns. As seen in the diagram, the output of recognition memory is divided into identical parallel pathways that feed directly into short and long-term memory.

Short-Term Memory (STM)

Short-term memory receives afferent master signals from recognition memory and recurrent master signals from long-term memory. It sends its outputs to long-term memory as slave connections. Every STM neuron can make as many slave connections as possible with neurons in long-term memory. As its name implies, STM retains events for a short period after their arrival. This means that STM neurons repeatedly fire for a short time after been triggered. The sustained firing is part of the mechanism that allows the memory system to create event associations over multiple-time scales.

STM is divided into two modules, one for past events (PE) and one for future or anticipated events (FE). A module is essentially a two-dimensional sheet of neurons. For every neuron that reaches maturity in the coincidence layer there is a corresponding target neuron in each of the STM modules. The function of the two modules can only be understood in conjunction with their integration with long-term memory. This is explained in greater detail below.

Long-Term Memory (LTM)

Long-term memory is where the system's causal knowledge and most of its intelligence reside. This is where temporal correlations over multiple time scales are discovered. Its operation is tightly integrated with short-term memory. As with STM, LTM is divided into two complementary modules called reactive memory and anticipatory memory. For every neuron that reaches maturity in the coincidence layer, there is at least one corresponding target neuron in each of the LTM modules. As it stands, a coincidence neuron can only sprout two axonal branches that synapse onto one LTM neuron and one STM neuron.

Each LTM module forms a recurrent loop with its counterpart in STM. This is explained below. LTM neurons receive master connections from the coincidence layer and slave connections from STM. They send their outputs to motor areas and back to STM.

Hierarchical Associations

The purpose of the loop mechanism is to create hierarchical associations over multiple time scales. As soon as an LTM neuron reaches a predetermined maturity level, it synapses onto a new STM neuron in the STM layer. The new STM neuron will in turn synapse onto an LTM neuron in the LTM layer. This mechanism gives rise to complex, temporally-chained, hierarchical associations to be created in the LTM layer. This is what allows us to understand temporally extent phenomena such as movements, words, phrases, sentences, etc...

The Reactive Loop

Reactive Associations

An intelligent system must continually make behavioral decisions based on recent events. This is the function of reactive memory which is depicted on the left side of the diagram below. STM neurons in the PAST-EVENTS module make initially random slave connections with LTM neurons in the reactive module.

Past Events

STM neurons in the PAST-EVENTS module repeatedly fire for a short time after receiving a signal from recognition memory. This affords LTM neurons with a window of opportunity within which they can form temporal associations that span multiple time scales. 

Reactive Neuron

A reactive neuron is like a coincidence neuron in that it fires if its input signals arrive simultaneously. It has one master synapse and one slave. Initially the neuron can have multiple slave input synapses but as soon one of the slaves becomes strong enough, the other slaves are disconnected. Even though a reactive neuron responds to the simultaneous arrival of input signals, its job is not to discover concurrent signals but to find temporally distant correlations between input signals. The reason is that the STM signals arriving at the slave input are retarded and really represent events that occurred some time before the arrival of the master signal.

Discovering Correlations

The way a temporal correlation is discovered is as follows: every time a master signal arrives, concurrent slave synapses are strengthened and all others are weakened. The problem is that slave synapses are constantly firing and there may be several synapses firing concurrently. The question is, which synapse is the right one? The answer is, and this is crucial, the synapse that was the last to begin firing is strengthened more than the others. In other words, the longer the elapsed time between the arrival of the master signal and the arrival of a slave signal, the weaker the synapse becomes relative to the other synapses. Conversely, the shorter the elapsed time, the stronger the synapse will be relative to the others. Eventually, the strongest correlated synapse will be selected. The neuron fires every time a master signal arrives at the same time as the retarded slave signal. The following table shows the main characteristics of a reactive neuron.

Causal Meaning

It is beneficial to think of the function of reactive neurons from an outside observer's point of view. The purpose of a reactive neuron is to attach a causal meaning to incoming perceptual signals based on their predecessors. For example, signal B arriving after signal A has a different causal meaning than B arriving after C. The overall function of the reactive mechanism is to sort incoming signals according to their causes. The recurrent loop makes it possible to discover extremely complex causal correlations the number of which is limited only by the dynamic complexity of the system's environment.

The Anticipatory Loop (under revision)

Anticipatory Associations

To anticipate means to participate in advance. An anticipated event is an event that is generated before its time on the basis of a precursor event. Anticipated events are just as important as past events to adaptation. An intelligent system must be able to predict the evolution of events in its environment and act accordingly. This is the function of anticipatory memory (shown on the right side of the above diagram).

Future Events

Predicting future events can only accomplished from past experience. As with reactive memory, STM neurons in the FUTURE-EVENTS module repeatedly fire after being triggered by incoming signals from recognition memory. The neurons make initially random slave connections with LTM neurons in the anticipatory module. The slave connections are chosen in the same manner as in the reactive loop.

Anticipatory Neuron

To get a good understanding of the function of the anticipatory loop, it is necessary to see it as the complementary opposite of the reactive loop. As seen in the table below, an anticipatory neuron is the opposite of a reactive neuron:

In other words, the neuron fires in anticipation of the arrival of the master signal.

Anticipatory Meaning

From an observer's point of view, the main difference between a reactive neuron and an anticipatory neuron is in the meaning of the signals they generate. Whereas the function of a reactive neuron is to attach causal meaning to a perceptual signal depending on the arrival of a predecessor signal, the function of an anticipatory neuron is to attach anticipatory meaning to a signal based on its successor.

Attention

Focusing

At any given time, there is a huge number of signals streaming into short and long term memory. These parallel streams are causally sorted by the reactive loop and become potential actions. The problem is, unless there is a way to select which stream is dominant, the system cannot focus on a given goal. It tries to do too many things at one time and becomes confused. In modern medical parlance, it is said to be suffering from an acute attention deficit disorder.

Two Attentional Mechanisms

The anticipatory loop plays a decisive role in determining the attentional goals of the system. Actions are created by the reactive loop based on recent events but are selected based on anticipated events. There are two mechanisms that control attention and the selection of actions. The first, which is explained below, is dictated mostly by chance and the system's environment. The second is pre-wired and is driven by pain and pleasure perception. The latter mechanism is explained in greater detail in the motivation page. Both mechanisms depend on the anticipatory loop.

The Magic Number Seven

Psychology teaches us that the maximum number of items that can be stored in short-term memory is seven. How is that possible since, at any given time, there can be any number of signals arriving at the STM layer? One can only conclude that the brain has a way of suppressing all but seven of the most salient signal streams coming in from the perceptual network. The mechanism of attention must not only be able to identify sequential signals as belonging to a given stream, it must decide which stream is salient and use this information somehow to directly suppress individual neurons in STM. On top of it all, it must choose only seven streams or less to focus on.

To be continued...

Next: Motivation

©2004-2006 Louis Savain