Edmund T. Rolls, Oxford Centre for Computational Neuroscience
Oxford University Press (2012) P.
Summary of p. 27-65 re: learning P. Keywords: CA3, CA1, hippocampus
Learning is based on increases in synaptic strength occurring as a result of the firing of pre-synaptic neurons and the activation of post-synaptic neurons. This is the basis of long-term synaptic potentiation (LTP). The pre and post synaptic activity needs to be nearly simultaneous. NMDA receptors appear to be necessary for this process. Only strongly activated neurons become modified, and this has the effect of excluding weak signals and noise. Thus information is stored where there is strong co-activity between neurons.
The number of memories that can be stored in a local network such as a cortical column is limited to about 10,000. Such a network may have only about 100,000 neurons. The position of the network in the cortex defines what its firing refers to, for instance whether it is visual or auditory.
A stimulus pattern is a set of neuron firings. One form of learning is for a stimulus (an uconditional stimulus) to become associated through learning with a conditioned stimulus. Another type of learning involves strengthening the synaptic connections that relate to a scene such as eating breakfast yesterday. Each memory is represented by a pattern of neural activity.
The excitatory connection between pyramidal neurons is seen as a key aspect of the cortex. In memory recall, a stored pattern of connections fires, while other neurons in the network are inhibited by inhibitory neurons. Once an attractor state such as this starts, positive feedback between the neurons keeps it firing, and this accounts for the maintenance of short-term memories. The number of connections on each neuron determines the maximum number of memories. The connections between neurons are viewed as a variety of ‘basins of attraction’ corresponding to learned patterns. Strongly activated neurons tend to inhibit other neurons, and the connections are reinforced so as to give a greater chance of activation later on. Other neurons learn to respond to other patterns, and in that way categorisation patterns are built.
This works in a heirarchical system with inputs moving from a small region in one stage to a small region in the next stage. Thus feature combinations might be formed at quite an early stage in the visual cortex. The system requires inhibitory neurons that receive inputs from pyramidal neurons to inhibit the output of other pyramidal neurons.
The dorsolateral prefrontal is associated with planning and reasoning, both of which require several items to be held in the short-term memory. Each item is seen as being held in a separate network. The short-term memory is seen as being closely related to the hippocampus. There is evidence for the hippocampus being involved with the association between objects and time. Recent studies show firing rates in hippocampal neurons indicating which parts of a task are current, and sequences of neurons can be activated at successive times. A large number of hippocampal neurons fire in relation to a sequence of individual events.
The hippocampus receives inputs from many areas of the non-primary cortex, and information, such as what object has been seen, and where it has been seen, constitute what is remembered. The CA3 network in the hippocmpus holds all these associations, and can recall these memories from the higher sensory cortex. Forward connections to the hippocampus from the association cortex run via the parahippocampal gyrus, the perirhinal cortex and the entorhinal cortex. Back projections to the cortex via run via the CAI region of the hippocampus. There is a convergence down onto CA3, and divergence on the way back up from CA1 to the cortex. It has been suggested that activation of only a small part of the memory in CA3 is sufficient to produce recall in the cortex. The dense interconnection of the CA3 neurons making them effectively a single network, and is suggested to allow associations between inputs from different parts of the cortex. Later recall of just one association may through the CA3 may allow recall of other associations. Mossy fibre inputs to CA3 come from approximately 48 randomly selected neurons in the dentate gyrus, and it is suggested that they activate random CA3 neurons as a way of keeping individual memories separate. Memory storage is impaired by damage to a dentate mossy fibre system.
The CA1 cells in the hippocampus lie at the start of the return path to the cortex. While CA3 may hold memories in such a way as to facilitate free association of ideas, in the transfer between CA3 and CA1 the memory is consolidated to a single feature. From CA1 back projections run via the entorhinal cortex to the parts of the cortex from which the inputs originally came. Information from the cortex converges ultimately on CA3, an area which has the smallest number of neurons in the entire memory circuit. Associative processing in CA3 is suggested to modify the back projections to the cortex. P. The primate hippocampus contains spatial-view neurons that increase their firing when a part of the spatial environment is being viewed. The allocentric (world-based) representation of space is seen as important for memorising the position of objects.
Forgetting is seen as an important aspect of the memory system. The storage capacity of a network is stated to be of the same order as the number of synapses, and the number of memories have to be restricted to this. Weakening of synapses can accomplish some of this forgetting. If a pre-synaptic synapse is not active the post-synaptic synapse will gradually weaken in a process known as long-term synaptic depression. This process also serves to limit the number of synapses on a neuron.
Attention: This subject is discussed in terms of the biased-competition hypothesis. There is suggested to be top-down biasing of the competitive interaction between stimuli in the visual field. The cells representing the intended stimulus suppress other cells representing distracting stimuli. The top-down aspect is suggested to derive from the short-term memory system, which may encourage the brain to track particular aspects of the environment at any given time.