Saturday, February 09, 2008

How Memory Works

This is a summarization of a paper by Eric R. Kandel on the molecular and synaptic basis for memory, Genes, synapses and memory storage. Kandel won the 2000 Nobel Prize for this work. I was moved to write this after listening to a segment of CBC's Ideas program discussing the nature of learning and memory. At the end of the paper, I draw the inferences from Kandel's work to my own.

The problem of memory has two major parts:
  • The systems component, which concerns "where in the brain memory is stored and how neural circuits work together to create, process, and recall memories. "
  • The molecular component which studies "the mechanisms whereby synapses change and information is stored"
The systems component - a history:
  • 1865 - Pierre-Paul Broca identifies speech production with a specific area of the brain.
  • 1876 - Carl Wernicke identifies language comprehension with a different area of the brain and suggests that complex behaviour requires the interaction of different brain areas.
  • 1929 - Efforts to localize memory fail; Karl Lashley formulates the Law of Mass Action: "the extent of a memory deficit is correlated with the size of a cortical lesion but not with the specific site of that lesion."
  • 1938 - Wilder Penfield localizes specific memories in epileptic patients (this was the subject of a 'Heritage Minute' video in Canada - "I smell toast burning").
  • 1957 - Scoville and Milner localize memory formation in the medial temporal lobe and show there are multiple, functionally specialized memory systems in the brain.
The idea that there are multiple memory systems in the brain has a long history in the philosophy of psychology:
  • early 1800s - French philosopher Maine de Biran argues memory can be subdivided into different systems for ideas, feelings and habits
  • early 1900s - William James divides memory into distinct temporal phases
  • 1913 - Henri Bergson distinguishes between conscious memory and habit
  • 1949 - Gilbert Ryle distinguishes between 'knowing that' and 'knowing how'
  • (1956 - Michael Polanyi - tacit knowledge (this isn't mentioned by Kandel))
  • 1969 - Jerome Bruner describes ‘knowing that’ as a memory with record and ‘knowing how’ as a memory without record
Scoville and Milner's studies of H.R., a patient who had the medial temporal lobe removed, yielded three major findings:
  • There was a short-term memory unaffected by the loss of other memory functions.
  • There was a long-term memory of events prior to the operation.
  • H.R. could form some long-term memories after, but denied doing so.
This established the distinctions postulated by the philosophers. This distinction between types of long-term memory is now characterized using the terms:
  • implicit - corresponding to 'knowing how', is habitual, unarticulated, and not recorded
  • explicit - corresponding to 'knowing that', is cognitive, artticulated and recorded

The molecular component

Kandel started by looking at the hippocampus but decided to focus on the simplest possible case, the marine snail Aplysia.

Why study Aplysia?
  • It is smart (for a snail) - it can create both short-term and long-term memories
  • It is simple - it has only 20,000 neural cells
  • Then neural cells are quite large, and hence easy to study
  • It is possible to map in detail the synaptic connections between cells with each other and with sensory and motor systems.
What they found (this is the key finding):
  • Short-term storage for implicit memory involves functional changes in the strength
    of pre-existing synaptic connections.
  • Long-term storage for implicit memory involves the synthesis of new protein and the
    growth of new connections.
This protein synthesis required to convert from short-term to long-term memory was developed early in evolution and hence preserved through all life forms, and is a general mechanism, responsible for both explicit and implicit memories.

Learning in pre-existing synaptic connections

Let's look at this in detail:

There are two major types of conditioning:
  • habituation - an animal perceives a sensation as innocuous and ignores it
  • sensitization - an animal perceives a sensation as noxious and tries to defend itself or flee
And two forms of learning:
  • non-associative - an animal habituates or sensitizes to a single stimulus
  • associative - an animal habituates or sensitizes to a pair of unrelated stimuli
In order to understand how the animal learns, therefore, "one needs in particular to work out the pathway whereby the sensory stimulus of the reflex leads to a behavioral response."

In the short term, habituation is represented by the weakening of the synaptic connection, and the resulting decrease in the release of glutamate, while sensitization is represented by the strengthening of the synaptic connection, and the corresponding increase in the release of glutamate.

Kandel doesn't include the diagram from Mann at right in the paper, but it nicely illustrates the process. The little blue ots represent the release of glutamate.

Kandel's description of this process (pp. 34-35) provides the chemical basis for Hebbian (associative) learning:

"Two events need to happen simultaneously: glutamate needs to bind to the postsynaptic nmda receptor, and the postsynaptic membrane needs to be depolarized substantially... This coincident activation of the nmda receptor and postsynaptic depolarization only occur when the weak siphon stimulus (cs) and the strong tail shock (us), are paired together."

Three major lessons are drawn from this work:
  • learning can lead to changes in the strength of connections (synaptic strength)
  • a single connection can participate in several types of learning
  • each of the three simple types of learning - habituation, sensitization and classic conditioning - gives rise to both short-term and long-term memory, depending on the number of repetitions
The growth of new connections

History of the distinction between short-term and long-term memory:
  • 1885 - Herman Ebbinghaus identifies two phases while learning nonsense syllables
  • 1941 - Zubin and Barrera, 1941 note the distinction in people hit in the head
  • 1960s - Louis Flexner and his colleagues identify a biochemical difference between them; long-term memory requires the synthesis of new protein during the consolidation phase
What's important is that there is a genetic basis for both the synthesis of the protein and for the consolidation phase.

Kandel notes, "Aplysia and Drosophila [a type of fruit fly] share some of the same genes and proteins for converting short- to long-term memory... creb has a role in learning in Drosophila that is similar or identical to its role in Aplysia, demonstrating striking evolutionary conservation. "

The mechanisms through which the proteins - CREB-1 and CREB-2 (aka ATF-2) - interact with the nucleus are complex and diagrammed (from Mann) at right.

In combination with other factors (such as, in the fruit fly, the the loss of a cell adhesion molecule), the interaction with the nucleus stimulates genes that results in the production of new synaptic connections.

Explicit memory storage

Explicit memory is more complex because:
  • it involves conscious participation in the memory recall
  • it doesn't depend on a simple stimulus; it usually depends on several sensory cues
Based on studies of mice, the hippocampus appears to play a major role in explicit memory. The hippocampus is basically a set of interconnected neural cell fields. It acts as a clearing-house for sensory input. Plasticity (the growth of new connections) has been discovered at all levels of the hippocampus. And the creb proteins appear once again to be implicated in the production of new connections.

Other work has demonstrated the plasticity of sensory systems. For example, experiments in kittens have demonstrated plasticity in the visual system. Cortical plasticity has also been demonstrated in adult monkeys. "These several studies suggest that long-term memory storage lead to anatomical changes in the mammalian and even the human brain much as it does in Aplysia."

A good example is the work done correlating the growth of synaptic connections and place memory. There are cells in the hippocampus, called pyramid cells, that are place cells - they fire when we occupy a certain place in our environment. So these cells form a cognitive 'map' of the environment. Various manipulations can lead to remapping, in which all the place cells change.


a. Learning and Memory

At this point we reach the end of Kandel's paper. What are we to make of these discoveries? What lessons should we draw?

For me, it requires a clarification of a comment that I have made on several occasions recently: learning is not memory. Kandel does draw a distinction (p. 31): "Learning refers to the acquisition of new information about the world and memory refers to the retention of that information over time." But what does that mean?
  • Learning is a semantic process. It is about things. It has meaning.
  • Memory is a syntactic process. It is a set of mechanisms. It may or may not have meaning.
This is a difficult distinction because the two are so frequently found in the same location. Pyramid cells, for example, that contain a 'map' of the environment, are created through a process of remembering, as a result of the changes of synaptic connections in the hippocampus, but also represent (via the sensory impressions that cause those changes) distinct places in the environment.

Nonetheless, the two are not the same. It should be clear from this work that it is possible to create memories that have no semantic content. It should be clear that through manipulations of the physical process we can create meaningless memories.

This, in turn, tells us a lot about the reliability of synaptic networks, and hence, of networks in general. In reliable networks, the mechanisms that cause the creation of connections between neurons are meaning-preserving, that is, they represent memories, and not merely manipulations of the process.

(I am being areful about how I state this, because there will be different accounts of what constitutes 'meaning-preserving').

This suggests:
  • Approaches to testing that test for learning, and not merely memory: such testing will be individual-centric (like the environment maps in the hippocampus) and not standardized (which is more likely to reflect syntactic manipulations).
  • Approaches to teaching which are based on creating semantic connections with the world, through the production of meaningful experiences, rather than syntactic manipulations of memory, such as memorization and rote
But this needs to be studied further. What constitutes meaning-preservation? It is not (as I'll show below) truth-preservation. But what is it? How do we measure for meaning, and not just syntactic compliance? Can knowing how learn help us determine what we learn?

b. Practice and Reflection

Again, as noted previously, learning is the result of repeated experiences of the same (or similar) type; the neural connections required for long-term memory will not be created without this repetition.

Learning is therefore not simply the presentation of information to an individual. It is not simply the transfer of a fact from one person to another. At best, this process could create only a short-term memory. In order to activate the neural connections necessary we need to stimulate the production of creb proteins, which happens only through repetition.

Advertisers, of course, know this, which is why they repeat brand names, jingles and phone numbers over and over. Seasoned politicians also know this, which is why the best oratos employ catching phrases that will be repeated over and over, as in the video Yes We Can (maybe one of the best political advertisements ever).

As I have said before, learning is not content. Learning is something over and above the pressentation of semantically meaningful information to a person. To learn, one does not simply 'acquire' content, one grows. To learn is a physical act, not a merely mental act.

Again, though, we want to look at this more closely. For example, what constitutes a repetition?

For example: the need for repetition would seem to suggest that a lecture would be a poor form of teaching, since it does not produce repetition. But:
  • Can we style lectures such that the repetition is contained in the lecture?
  • Can people listening to lectures create repetition through the use of different modalities, such as taking notes, live-blogging or summarizing?
  • Can we create repetitions through the conduct of lecture-related tasks, such as projects or problem-solving based on the contents of lectures?
  • Does learning for ourselves stimulate the production of the repetitions required for memory?
  • Is there a connection between semantic content and repetition - does learning in authentic contexts increase the probability of remembering?
I would suggest that the answer to each of these questions is 'yes'. But they are the sorts of things that bear further investigation.

c. The nature of knowledge and inference

There is a persistent school of thought in both the philosophy of psychology and also in educational theory that suggests that cognition is based on logical and linguistic rules, that there is a logical syntax that governs learning and cognition.

Examples of this range from the postulation of Chomsky's deep grammar to Fodor's language of thought to Hempel's H-D model of the sciences. The proposition is essentially that meaning-preservation is tantamount to truth-preservation, where truth-preservation is as is well understood from logic and mathematics.

But what we learn here is that learning is associative, not propositional. That the mechanisms that govern this process are not expressions of truth-preservations, but are - at best - expressions of meaning preservation, where meaning has to do with sensory perceptions and states of affairs in the environment rather than abstract principles of logic and mathematics.

I have expressed this in the past as follows:

Our old understanding of logic, inference and discovery is based on universals:
– rules, such as rules of inference, or natural laws
– categories, such as classifications and taxonomies

Our new understanding, through, is based on patterns recognition:
– patterns, such as the activations of similar sets of neurons
– similarities, such as the perception of similar properties in nature

That is not to say that these universal principles play no role in our understanding. It means, rather, that we need to see them in a new light:
  • These principles represent 'convenient fictions', not underlying principles of nature
  • These principles are learned - they are not innate
There's a lot more work to be done here. The nature of inference based on patterns and similarities is poorly understood. It is one thing to say things like 'an understanding of learning based on simple causation is mistaken' and quite another to describe the complex mechanisms that actually occur.

We need to dig into the logic of similarity, following the work of people like Tversky and Varela, to conjoin this with our understandings of social network theory and graph theory.