How would you engineer a novelty detector?

Play cognitive engineer: if you were designing an intelligent system, you'd probably use the same system to detect novelty as that used to detect familiarity. After all, one is simply the inverse of the other - so novelty can be thought of as a below-threshold value for some familiarity signal.

Upon reflection, however, you might notice some subtle problems with your design. First, novelty is informative only insofar as it is engages a change in the behavior of your system: novelty indicates an opportunity for learning which has not been exploited by the system. Critically, this includes situations where there are independently familiar stimuli or responses that have been recombined in some new way. Familiarity detection, on the other hand, is widely thought to be insensitive to such "conjunctive" representations; a separate (by some models) process of recollection is responsible for recovering conjunctive relationships between items.

Second, and more critically, there is not a simple inverse relationship between recollection and novelty: events which cannot be recollected may nonetheless be familiar, which is undiagnostic of novelty: recollection may fail even when items are not associatively or conjunctively novel.

The same thought experiment was conducted in a 2007 paper by Kumaran & Maguire in the journal Hippocampus. They arrived at the following solution, supported by new original neuroimaging data:

• The authors dissociate between three types of novelty: of stimuli per se, of their associations, and of their larger context; spatial & temporal configurations of stimuli are included in the construct of "associative novelty"
• The authors propose that the hippocampus accomplishes detection of associative novelty, whereas surrounding regions of the medial temporal lobe implement stimulus novelty (without regard to configurations; principally the perirhinal cortex)
• This mechanism consists of two parts: a comparator (located in the hippocampus) and a familiarity detector (located in the surrounding perirhinal cortex).
• The familiarity detector: low values on a scalar familiarity signal indicate stimulus novelty, as in many computational models of explicit memory
• The comparator: the CA3 subregion of the hippocampus acts as an autoassociator (allowing for recall in the form of pattern completion), which projects to the CA1 subregion via the "Schaffer collaterals." Critically, CA1 can compare this with sensory input, received via the "perforant path" from entorhinal cortex. Thus this mechanism implements a second type of novelty detection: whether the results of recollection match or mismatch those combinations of stimuli currently encountered in the world.
• The authors report an original fMRI experiment in which associative novelty (the last two items of a previously-observed 4-item sequence were reversed) was directly contrasted with global novelty (all items in the sequence were rearranged). They observed greater hippocampal activity to associative novelty, and greater entorhinal and perirhinal activity to any stimulus recombination (relative to a repeated sequence), consistent with a general familiarity detector.

This and subsequent work by the same authors have compellingly demonstrated that novelty detection relies on at least two distinct mechanisms, owing to the different varieties of novelty which are too often confounded in cognitive science (as discussed yesterday).

On the other hand, this work leaves several questions unanswered:

1. How does this "mismatch detection" relate to the well-known mismatch negativity observed on the scalp following a contextually novel stimulus, and preceeding a p3 complex (described in more detail yesterday)?
2. This paper appears to provide strong evidence in support of a two-system model of recollection and familiarity, although like unitary accounts, familiarity does seed recollection. However, strong evidence for these two-system models often comes from ERPs recorded over the parietal cortex; what is the role of the parietal cortex in the detection of novelty?
3. More generally, prefrontal cortex is known to be involved in novelty detection (e.g., the p3a subcomponent). How would the fMRI results reported above be different if subjects did not have an incidental target detection task?

These remaining questions center on hippocampal-PFC interactions, which appears to be the focus of some of Kumaran's more recent work - particularly with respect to the neurotransmitter dopamine. The function of dopamine is hotly debated, with successful computational models suggesting it has a role in signaling reward mismatch, and other researchers advocating a role for signaling novelty in general.