Traditional theories of hippocampal functioning have posited that the hippocampus stores information about how basic features relate to one another. This allows for an impoverished retrieval cue to evoke an entire memory, because it only takes a few features to be presented for the configural representation to activate all of the feature representations that were activated previously.

There is considerable evidence to support this notion. One of the most suggestive lines of evidence comes from research in which animals are trained on non-linear learning tasks. In such a situation an animal is rewarded for responding to stimulus A. The animal also is also rewarded for responding to stimulus B. However, the animal is not rewarded for responding to stimulus A+B. Thus, the animal must learn to distinguish the conjoint representation A+B from the individual features when they are presented together. Animals with hippocampal lesions are not capable of distinguishing between the three different stimuli, preventing them from learning how to perform this task.

Disconfirming evidence

However, other research suggests that even animals with hippocampal damage can learn to perform at similar tasks. The learning is posited to take place in the neocortex.

Alternative perspective on the hippocampus

O’Reilly and Rudy suggest that what the hippocampus really allows for is the unintentional learning of relational representations within an episode or event. In support of their argument they cite research in which the relationship of stimuli to one another is not required for the task. However, animals that have lesions to the hippocampus present behaviors that would suggest that they did not incidentally learn how things related to one another. While control animals demonstrate behavior suggesting that they are able to recall the previous configuration of stimuli.

Summary of the hippocampus and cortex (see page 317)

Note that the hippocampus is thought to allow for relational information within a single event to be represented inadvertently and very quickly, which the authors suggest is at the heart of episodic memory. And the cortex is thought to allow for generalities between multiple events to be learned. This learning is suggested to be intentional and to be relatively slow. In other words, the cortex allows us to extract knowledge about how the world works over time. This understanding of cortical functioning flows nicely with research demonstrating the importance of neocortex in problem solving tasks.

How does this happen Mr. Wizard? The story of two types of Conjunction learning:

Error-driven back propagating learning and Hebbian-like mechanisms

The Hebbian-like mechanism allows the brain to represent things that co-occur and the back propagating learning allows for the connections between individual representations to be modified through trial and error situations.

Look at Figure 1 for a nice depiction of the difference in resolution between the representation thought to be storedin the hippocampus and the cortex.

Hippocampal Functioning

Pattern separation. Patten separation is the ability of the hippocampus to keep separate highly similar memory events. This is accomplished in theory by having highly simplified representations of memories stored in the hippocampus. Reducing the number of active units within a memory trace reduces the likelihood that similar memories will share features. Those memories that do share features become bound into a single coherent representation.

Pattern completion. Pattern completion is the mechanism that allows for only a subset of retrieval cues to be presented and for the entire memory to be reactivated and retrieved.

Basic Principles of the Model

Learning rate. The hippocampus learns relationships quickly and the cortex learns relatively slowly.

The general scheme for encoding new memories in the hippocampus is that activation comes into the EC from the cortex and then flows to the DG and CA3, forming a pattern-separated representation across a sparse, distributed set of units in these layers. (p. 323)

Retrieval in the hippocampus is still a big fat mystery if you ask me? Retrieval is primarily a product of pattern completion, the presentation of a few features, activations a firing pattern, which accesses a composite represent of an entire episode.

Simulated Data

Nonlinear discrimination problems

Negative patterning. A(reinforced), B(reinforced), A+B(not reinforced)

Ambiguous feature. AC(reinforced), B(reinforced), AB(not reinforced), C(not reinforced)

Biconditional problems.CA(reinforced), CB(not reinforced), DA(not reinforced), DB (reinforced)

Results from the simulations are presented in Figure 6. Intact models were better able to learn the negative patterning problem than HL models. However, there was no difference in the ability of the two models to learn either the ambiguous feature or the biconditional problem.

Do animals really decompose composite item (e.g., AC in the Ambiguous feature problem) into their basic components? (p. 327)

Transverse Patterning

A paired with B = A is the correct choice

B paired with C = B is the correct choice

C paired with A = C is the correct choice

Because the stimuli are presented individually this problem most resembles the NP problem.

Incidental Conjunction Learning

Rats are exposed to a complex novel environment. In the first stage all rats explore the environment as all good natured rats should, because the environment is new and who knows what there might be to eat or to eat them lurking around. In the second phase, the rats are reintroduced to the environment but the objects in the environment have been rearranged. Thus, the environment to some extent is new and being a new environment should evoke exploratory behavior in the rats. However, only the intact rats engage in such behavior, those rats that received hippocampal lesions did not engage in exploratory behavior.

In a simpler design rats were taught a response in a set context. Rats were taught to respond to A in context 1 (C1) and to respond to B in context 2 (C2). Rats with intact brains did not respond to the stimulus when tested in a new context but rats with hippocampal lesions responded to the stimuli regardless of the context.

The results from a simulation study investigating this in intact and HL models is presented in Figure 15.

Contexual Fear Conditioning

Researchers have suggested that animals must first form a representation for the environment in which they will receive the shock prior to being able to form a conditioned fear response. Results from a simulated conditioned fear response experiment are presented in Figure 18.

Trasitivity and Flexibility

Problem space: A versus B = A response

B versus C = B response

C versus D = C response

D versus E = D response

Critical Tests: A versus E = A response (all animals selected A most often)