ConceptMemoryMechanisms
Related: ConceptNeuronLogicality
Because of the size and varied structure within the brain, it is likely that all of these are used.
- Source: Brain Simulator II _ The Guide - Charles Simon
- Chapter: Chapter 4: Applications of Neurons, Four Memory Mechanisms
Let’s consider storing information in the internal charge state of a neuron. For example, a neuron represents a logical 1 if the internal charge is (+0.1) or greater and a logical 0 otherwise.
Using an IF-neuron-model, the storage time is infinite but with a more realistic LIF-neuron-model, the internal charge will decay so the memory must be refreshed by reading periodically. If the memory is not refreshed, it will gradually lose its content. On the other hand, this may be a useful mechanism for short-term memory in the brain. That is, the memory is never reset but only stored and read. Simply waiting some amount of time (perhaps a second) is sufficient to clear the memory by leakage.
This memory has very fast store and retrieval times and uses no energy when it is not being read or modified. If the memory is not refreshed, it will gradually lose its content. On the other hand, this may be a useful mechanism for short-term memory in the brain.
Using neurons to store bits of information by firing continuously.
It’s worth considering a special case where neuron values and synapse weights are restricted to represent digital circuits, to either 1 or 0 (fixed-weight synapses).
Structure: A neuron connected to itself (self-circular). And two synapses connected to it, synapse A with a weight of (+1) and synapse B with a weight of (-1).
Result: It will spike indefinitely if it ever receives a spike from synapse A, and only stop spiking if it receives a spike from synapse B.
The spiking state of this neuron can represent either a 1 (spiking) or a 0 (non-spiking).
Previously, I described how (1) can act as a single bit of memory. This has the advantage of being easy to explain, requires one neuron per bit, and is very fast (for a neuron). But we need to consider that spiking neurons consume energy. This means that storing information in continuous spiking is unlikely to be a widespread approach in the brain.
Observation: As described in (1), this method oversimplifies the problem with a neuron synaptically connected to itself. When such a neuron emits a spike, it will likely be received within the neuron’s ConceptNeuron#RefactoryPeriod and be ignored.
Solution: This can be overcome by increasing the ConceptAxonDelay or by having multiple neurons forming a ring of connections.
A or B depending on whether speed control is needed.
Short-term memory for incoming audible signals might incorporate A, while the process for creating speech might use B so that you can speak at whatever speed you like.
A
Simpler mechanism.
The delay line requires one neuron per bit.
Transfers a spike directly from one neuron to the next (and so is a fixed time delay). (A "delay line" can be thought of as a bucket-brigade where the incoming signal, “In,” is transferred down a chain of neurons.)
The length of the delay-line limits the amount of memory.
Synapses could be added at any intermediate step so, for example, you might think back in your short-term memory to recall the last word, or the last phrase, or more.
B
More complex.
The delay line requires two neurons per bit.
The spike advances by one neuron every time the Step neuron fires and so can provide a variable amount of delay.
(The classic AI mechanism is storing information in the weights of synapses)
A synapse of weight (+1.0) representing a 1 and a synapse of weight (0.0) representing a 0.
Offers the advantages of being able to store much more data as there can be thousands of synapses for each neuron.
Further, this is the only memory mechanism with any degree of permanence. If you want data to be stored for days or years, this is the mechanism for you because the previous mechanisms either use too much energy or decay over time.
The disadvantage is that this mechanism is much slower to change, requiring many spikes to set the synapse weight.