Relating The Reading To The Lectures

There isn't much in the chapter that is directly related to the lecture, which focused on connectionism. Apart from the last couple of pages, PDP modeling is hardly mentioned in the chapter.

However, with the advent of New Connectionism, modern cognitive science has turned once again to paying a lot more attention to the study of learning, which has largely been neglected because of learning theory being strongly linked to Behaviorism. It turns out that the "lawful nature of learning" that Shanks talks about early in his chapter is mirrored exactly in many of the learning rules that connectionists have developed.

Margin Notes On The Chapter

Introduction

There are many different kinds of memory, ranging from recall of information learned at a previous time, semantic or lexical memory, or memory of procedures for accomplishing tasks. Focus of this chapter is on how learning and memory are possible. By looking at classical conditioning as a prototypical example, it appears that memory obeys a number of fundamental laws.

Classical conditioning -- through pairing, a subject can be trained to elicit a conditioned response (CR) to a conditioned stimulus (CS) as well as to an unconditioned stimulus (US). "Plainly, an animal that produces a CR in the presence of the CS must have learned something about the predictive relationship between the CS and US" (p. 278). What are the main properties of this kind of conditioning?

1. It depends on sequence of events -- CS must precede US
2. Strength of conditioning depends on interval between CS and US. "If the bell precedes food by a few seconds, salivation to the bell will be much greater than if it precedes food by many minutes."
3. CR will be extinguished (will disappear) if CS is presented repeatedly without the US

Instrumental conditioning differs from classical conditioning in several important ways. In this kind of conditioning, reward or punishment is applied after subject makes a particular response. The behavior will increase or decrease in likelihood depending on the outcome (reward or punishment), following Thorndike's "law of effect".

Ebbinghaus was the first to study human memory under controlled conditions. "The terms learning and memory are often presented as different concepts, and the distinction between them is preserved in many ways in psychological research. Yet in reality they are different sides of the same coin." Memories are what learning leaves behind.

The definition of learning and memory is contentious (and Shanks will conclude futile!). Behaviorists equated it with change in behavior. But evidence shows that learning can occur without changes in behavior.

"A more cognitive view, therefore, is that memory is an abstract term that describes mental states which carry information, while learning describes a transition from one mental state to a second, in which the information is in some way different" (p. 281). There are problems with this approach two -- for example, how does it differentiate learning from forgetting? Also some "learning" appears to be noncognitive, but is hard to exclude via this definition. As far as Shanks is concerned, all of this illustrates "the futility of trying to define the terms learning and memory."

Memory is almost always studied using recall procedures or recognition procedures. This is a tradition started by Ebbinghaus. Savings during re-learning of old material can also be used to study memory.

There is a long-standing distinction between short- and long-term memory. Lots of experimental evidence supports this distinction. E.g. evidence showing the limited capacity of the short-term store (STS). Recency and primacy effects in serial position curves from free recall studies also support this distinction. The recency effect is due to STS -- for example, it vanishes with a delay in recall, but with this manipulation the primacy effect is not changed. The size of the STS is constant across domains of items. "This emphasizes the crucial fact that what gets maintained in the short-term store is a set of 'chunks' of information, where a chunk can be a word, a digit, or whatever."

The primacy effect is due to the long-term store (LTS). The first items in a free recall list get lots more rehearsal, which helps produce consolidation of items -- and the primacy effect.

Neuropsychological evidence also supports the STS and LTS distinction. Patients KF and HM demonstrate a double dissociation of memory systems

STS itself appears to be a system of modules. Evidence exists to support Baddeley's decomposition of the STS into a central executive, visuospatial scratchpad, and an articulatory loop.

Another long-standing distinction is between procedural and declarative memory. Semantic and episodic memory are subcomponents of declarative memory.

The distinction is between procedural and declarative memory has its roots in cognitive neuropsychology. For instance, HM's problem is with facts -- he can learn new procedures. "It is now known that amnesics who otherwise have acute difficulty learning and remembering new information can perform very well on a whole range of procedural learning tasks" (p. 291). What is the distinction between tasks that can be learned by amnesics, and tasks that cannot? Such patients fail on declarative tasks. "Declarative memory refers to memory for facts  as well as memory for past events  that is to say, any memory that can be stated as a proposition. In contrast, procedural memory is non-factual: as we have seen, Pavlovian conditioning and the ability to ride a bicycle would be good examples" (pp. 291-292). Different brain areas are assumed to be involved in these two types of memories.

There are still problems with the declarative/procedural distinction. "Probably the main worry concerns data suggesting that amnesics are more impaired in some forms of declarative memory than others." E.g., episodic memory is most impaired. But not just episodic memory is impaired.

With respect to normal, "a natural prediction of the theory is that it should be possible to observe cases of procedural learning in the absence of accompanying declarative knowledge, yet it has proven very difficult to find convincing examples of this." It appears that procedural and declarative memories "march in step" -- there is a close correspondence between them.

Paradox -- some evidence supports the procedural/declarative distinction, but other evidence does not. As a result, some researchers resist the decomposition of memory systems. "An alternative view is that memory is simply a record of the mental operations carried out on stimuli. What sort of operations are there? The most obvious distinction is probably between data-driven and conceptually-driven operations" (p. 295). For example, from the depth of processing perspective, memory is affected by the type of operation carried out on stimuli. Encoding operations are also important -- Tulving has shown that cues present during learning are very important for producing proper retrieval, leading to his encoding specificity principle: "a retrieval cue will facilitate recall if and only if its relationship with the to-be-remembered item was processed at encoding."

All of this has lead to the transfer-appropriate processing theory of memory. "This theory states that performance on some memory tests benefits from a prior learning episode to the extent that the mental operations needed to complete the test overlap with those required during the learning episode."

"The theory of transfer-appropriate processing is a way of accounting for memory phenomena in processing rather than in modular terms. Instead of explaining dissociations between declarative and procedural tests in terms of different underlying brain structures, such dissociations are viewed from within a single memory system in which a variety of different operations may be performed on a stimulus. Memory is simply the record of these operations" (p. 299).

What is the explanation of forgetting? To start, we must realize that the term "forgetting" is ambiguous -- can describe behavior, or the "deeper" notion of information being eradicated from a memory system. (Shanks argues that this latter kind of forgetting never occurs.)

The traditional view is that all forgetting is caused by retroactive interference. Such interference might overwrite old information, or it might make the original association hard to retrieve. Current evidence suggests that forgetting is not due to the decay of memory traces. The amount of RI can decline as the retention interval is increased -- which supports the view of retrieval failure as the source of forgetting.

A common view is that categories are represented by mental prototypes. "The concept CHAIR, for example, might be represented by a typical chair that has been mentally abstracted from our experience of a large number of actual chairs. On this account, responding to a new stimulus is a function of its similarity to the prototype" (p. 305). Ease of classification depends on distance from prototype, as many studies have shown.

"Perhaps the most compelling reason to believe that abstraction of the prototype underlies categorization is the abundant evidence that the prototype stimulus itself will be classified accurately and rapidly, even when it has never been presented in the training stage of an experiment." Prototypes, once abstracted, tend not to be forgotten.

But prototype theory is not completely supported. Specific exemplars can affect learning.

Another form of representation is parallel distributed processing (PDP). "After sufficient training, a network can be extremely good at classifying diverse objects into their correct categories, with its conceptual knowledge being distributed across a large number of parallel weights between highly interconnected units. Networks like this are exciting because they store conceptual knowledge by appealing only to processes known to operate in the brain, namely the transmission of excitatory and inhibitory signals between units" (p. 307).

But all of these theories of conceptual representation are incomplete. First, they are all feature based. Therefore they all ignore functional attributes. (NB: This strikes me as a pretty lame criticism -- why can't we treat functional attributes as features?). Second, they are all similarity based. "Although simialrity often does play a crucial role in category decisions, people also appear to be able to use concepts in a far more flexible way that is largely unconstrained by similarity."