In his much cited work Marr stressed the idea that psychological phenomena can be understood at multiple levels of analysis [Marr, 1982]. He separated what is computed, how it is computed and the level of implementation from each other. Although this distinction clarified the way of thinking in some domains, it did not help too much in others. Frequently, the hardware level determines the most efficient way to compute something, and what an organism does is limited by what it can compute efficiently. Thus these three layers cannot be separated productively from each other. Cognitive neuroscience in particular tries to bridge the gap between these three levels wherever possible. Keeping this spirit, I found it appropriate to start the discussion of memory at the level of neurons.
Hebb combined the associational-learning theory of Clark Hull and his contemporaries in the 1930s and 1940s with what was known of nervous system activity at his time, to describe the basis of learning and memory. He separated short term memory from long term memory. He described short term memory as an active process of a short duration, whereas long term memory involved structural change in the nervous system.
Hebb hypothesized that if one neuron frequently takes part in exciting another, some growth process or metabolic change takes place in one or both cells and the strength of their connection increases. Thus the neurons become linked functionally. In his view, the cell assemblies organized to process perceptual information were capable of continuing their activity after the stimulation has ceased. This repeated activation after the initial sensory input was necessary to produce the structural changes for the long term memory. These repeated reverberations formed the short-term memory. Furthermore, for the structural synaptic changes to occur, there must be a period in which the cell assembly is left undisturbed. Hebb called this period the consolidation, which might take from 15 min to an hour. The amnesia of events just prior to a concussion are the evidence for this process. Finally, Hebb hypothesized that any cell assembly could be excited by others. So particular thoughts could occur in the absence of the original event they correspond to. This was the basis for thought and ideation.
The work by Bliss, Gardner-Medwin and Lø mo demonstrated the phenomenon of long term potentiation [Bliss and Lomo, 1973]. They electrically stimulated perforant path fibers to the dentate area of the hippocampal formation of rabbits. They measured the magnitude of the response of cells that are known to receive projections from the stimulation area using microelectrodes. Their basic finding was that increasing the frequency of stimulation for a short period of time resulted in potentiation. Potentiation can be defined as increased sensitivity of the receiving area. Immediately after the high frequency stimulation they observed a gradually declining increase in the response of the recipient cells. This is a transient effect and is known as posttetanic potentiation. More importantly, even though this high magnitude response decreases gradually, it does not necessarily return to the baseline and remains at an elevated level. This is known as long term potentiation, or LTP for short. Although the phenomena of posttetanic potentiation and LTP have strong appeals as explanations of short term and long term memory, they are low level mechanisms, and they do not explain architecture level phenomena such as differential effects of lesions in different parts of the brain.
The molecular basis of LTP is beginning to become clear. Although the trigger for LTP is generally agreed to occur in the postsynaptic cell, the site at which it is expressed is still disputed. Some authors have suggested that the mechanism of LTP is enhanced neurotransmitter release, others that it is increased postsynaptic sensitivity to transmitter, and still others believe that both maybe the case. Bekkers and Stevens report a molecular mechanism that behaves similarly via a pre-synaptic mechanism [Bekkers and Stevens, 1990]. Their analysis of the statistical properties of synaptic transmission, before and after the induction of long term potentiation, suggest that expression of LTP largely arises in a presynaptic mechanism - an increased probability of transmitter release.
It is also possible that structural changes in neurons underlie some forms of memory and learning. In an extensive set of experiments by Greenough and his colleagues it has been shown that when animals are trained for specific tasks or exposed to specific environments there are changes in the dendrites of neurons [Greenough and Chang, 1985]. Greenough has shown that an increase in the number of dendrites is accompanied by an increase in the number of synapses, which might account for functional changes. In addition, the quality of the synapses, new or old, are known to change in long term potentiation. Although the exact mechanisms for these changes are still being discovered, it seems clear that behavioral change stems from morphological changes in neurons.
Studies on simple animals to discover the neural underpinnings of learning has made significant contributions to our understanding of the mechanisms of memory. Hawkins and Kandel suggested that higher forms of learning may be based on the mechanisms of simple forms of habituation, sensitization and conditioning [Hawkins and Kandel, 1984]. Studies on the invertebrate Aplysia showed that the siphon withdrawal reflex in this animal is an extension of the mechanism underlying sensitization. Similarly, Hawkins and Kandel showed how several higher order features of classical conditioning, including generalization, extinction, second-order conditioning, blocking and the effect of contingency can be accounted for by combinations of the cellular processes that underlie habituation, sensitization, and classical conditioning. The Hebb synapse is only one possible way in which neural networks could learn. Hawkins and Kandel described a different mechanism that does not depend on simultaneous events in pre-synaptic and post-synaptic neurons, but rather rely on neurochemical events within the sending neuron.
Gluck and Thompson studied the phenomena in Alypsia using a computational model [Gluck and Thompson, 1987]. They developed a computational model of the neural substrates of elementary associative learning, using the neural circuits known to govern classical conditioning of the withdrawal response of Alypsia. This study proved the importance of actually building computer models that allow the researchers to see the dynamic interactions. Gluck and Thompson discovered several shortcomings of the model created by Hawkins and Kandel and suggested new directions of research. This work demonstrates clearly that the human brain is very bad at simulating complex systems reliably. When any non trivial model of a complex system passes the level of vagueness, and becomes a precise specification of the mechanisms, it is important to start using computer models to observe counter-intuitive results that emerge out of the complexity.
Ambros-Ingerson and colleagues worked on a simulation of the olfactory learning in rabbits using the same principles as Gluck and Thompson [Ambros-Ingerson et al., 1990]. They modeled layers I and II of the olfactory paleocortex, as connected to its primary input structure, olfactory bulb. They demonstrated that long term potentiation by means of repetitive sampling of inputs caused the simulation to organize encodings of learned cues into a hierarchical memory that uncovered statistical relationships in the cue environment, corresponding to the performance of hierarchical clustering by the biological network. These findings suggest that similar principles of learning may govern how networks store information in different neural systems. Different systems have different physical organization of the neurons and different input and output relationships, but the same learning principles may apply to them in general.
Barto and Jordan proposed a novel learning algorithm for neural networks that might be biologically more plausible than the standard back propagation [Barto and Jordan, 1987]. The standard algorithm for learning in artificial neural networks involves comparing the output of the network with the desired output, and propagating the gradient of the error backwards through the connections to make the network gradually approach the desired setting. This method performs well for using neural networks to solve artificial problems of classification and recognition, however it is unreasonable to expect the existence of a ``desired output'' oracle in nature. The Associative Reward-Penalty algorithm proposed by Barto and Jordan relies on an approximation of the gradient by individual units independently, rather than an exact computation that has to be propagated through the network.
It is generally accepted then, that learning and memory are based on morphological change in the synaptic structure of the neurons. It seems plausible that these changes occur in the systems that process the original information. Kolb discusses three questions that need to be answered if this theory happens to be true [Kolb and Whishaw, 1990a].
The first concern is that neurons in the lower levels of processing, such as primary visual cortex, should not change much, otherwise the information sent to higher areas would be radically different. I think this problem has a natural solution, because low level areas, by definition are exposed to the raw stimuli from the world. As processing progresses to higher levels, these stimuli gradually get recognized and categorized and finally labeled. It seems then the number of different stimuli a higher level has to deal is much more restricted than a lower level. Consider recognizing faces as an example. Even the different perspectives, lighting conditions and distances a single face can be seen from make it highly unlikely for the primary visual cortex to get the exact same impression of a face twice in a life time. However, at some higher level of the visual system, the face is processed, recognized, and labeled as ``my mother''. At that level of the system, the different inputs one can get is limited by the different number of people one can recognize. Because long term potentiation requires repetition of the input stimuli, it is natural that changes due to learning do not easily occur in a low level system, but memories are usually formed in higher level systems.
Second concern is how we remember ideas and thoughts, if sensory experience is what changes the sensory systems and forms memories. This would not be a concern, if ideas and thoughts are just a complicated mosaic of sensory imaginations, presumably at higher perceptual fields and association fields. There is evidence from the visual imagery work that thinking and perceiving may be using the exact same machinery. If we understand and think in terms of our lower level sensory primitives, we do not need any extra machinery to have memories about our thoughts.
The final question is how we find specific memories, if they are widely distributed in large cell assemblies. This would really be an impossible problem if brain was organized like a computer where processing and memory are in two different locations, connected to each other with a narrow bottleneck. However it does not seem very likely that the same organization exists in the brain. Processing and memory are probably handled by the same underlying hardware. If this is the case, then just as a system can recognize an external stimuli and activate its related parts, it can also recognize an internal stimuli, generated by the activation of another system in the brain. Remembering can be like a chain reaction. A small activation in a single system, maybe caused by an external stimuli, or a previous piece of thought, first activates the related parts of this system. Then this activation grows like a snowball, and connected systems start activating their own relevant memories. It is wrong to restrict oneself to think that something searches inside the head and finds the memories that we want. It is probable that the memories sense when they become relevant and jump into our attention. Given that we hold millions of bits of information at a given time in our long term memory, the fact that we can find anything relevant when we want to think about a particular thing, can only be explained by a parallel processing system of this sort.