If you're an animal wondering around trying to figure out what to eat and not die, you are going to want to remember to avoid any and all substances that made you feel the slightest bit ill. This is why conditioned taste aversion is a prime example of single-trial learning. This post isn't about conditioned taste aversion though. It's about how you know which taste you should be avoiding. The answer is novelty-detection. You automatically note whether you recognize each thing you encounter throughout the day. That way, when something important happens that hasn't happened before, you know what stimulus to attribute it to.

Some people would probably think, "Hmm... how does the brain detect novelty?," but Schuman and colleagues were like "Pshaw! How 'bout... how do individual neurons in the human brain detect novelty?" Sometimes, in the treatment of epilepsy, deep brain electrodes are implanted, presumably to detect the epileptic focus for more accurate removal. This single neuron recording technique was possible because somebody figured out that as long as you are sticking this huge electrode down into the hippocampus, you might as well stick a bunch of leetle teensy wires down through the middle of the fat electrode.

This isn't a technique you will see that often, since there simply aren't that many people with epilepsy in need of surgical treatment. However, when you do get to see it, it provides quite an advantage over something like fMRI because the temporal resolution is excellent while still allowing one to definitely implicate this structure or that. I'm not as familiar with the human literature as I probably should be. In the animal literature, there is a sharp distinction between the functions of the amygdala and hippocampus, but in this study they catch a little of both and don't seem too surprised to get largely overlapping data for the two.

The paradigm is extremely simple and I'm going to simplify it further. Recordings were taken while people were looking at something they had seen before (familiar) or something novel (novel). In the six seconds following stimulus onset some neurons increased their firing when presented with novel stimuli and some some with familiar stimuli. Familiarity-detection appears to be the readout of some sort of memory process, because familiarity-detectors were apparent even 24 hours after the first encounter with some stimulus. Oh yeah, and it only took one encounter to do the trick. The change in single neuron rate-coding was rapid and automatic.

Novelty and familiarity coding neurons were found in both the amygdala and the hippocampus, but there were about 2.5X more in the hippocampus. One thing that is maybe worth noting at this point is that it is not clear that the process of novelty detection is occurring in these structures. In the discussion, they seem to argue that the only other novelty encoding brain systems are modulatory (read: slow) so they can't explain the effect on these neurons. It seems more likely to me though that novelty detection will occur at each step in stimulus processing. It would be interesting to break this down into different sensory categories instead of complex visual stimuli. An extremely heavy orange apple that squeaks when you bite it and tastes like a coconut will probably add up to being more novel than if just one sensory modality is affected. I guess that might be a good question, can there be such a thing as a novelty gradient or is it all-or-none?

More than one neuron per stimulus registered as a novel or familiar detector. This leads to questions about how these neurons are coordinated. In the paper, they produce a 'population decoder' computational model that has access to all the firing data. After the decoder was trained up, they fed it info from different numbers of neurons. With one neuron it could tell whether the subject was looking at a novel or familiar object 67% of the time. Give it six neurons and it scores a 93%. Naht bad. Shades of lie-detection. Lucky for the liars we probably still aren't legally allowed to stick electrodes in peoples heads on a hunch.

The most interesting finding comes when they compare the information in the firing rates to the behavioral response of the patients. When the population decoder has access to all the firing info, it can actually guess whether the stimulus is novel or familiar better than the patient. For the study, this eliminates a strong alternative hypothesis that the neurons are merely representing the patient's intentions or motor sequencing. The neurons can say one thing and the patient can say another. Take that determinists! We have free will after all! It really makes you wonder though which region is sending the wrong signal and tripping them up.

One last note. In the discussion, they mention that it has been very difficult to find these types of neurons in other primates. I should think that novelty-detection would be conserved way far back. The explanation in the paper is that the paradigms you have to use with primates require lots of trials before they are up to speed. Since we can do weird tasks right off the bat for no reason it is easier to correlate behavior with firing. Your homework is to design a novelty-detection task for primates that they can learn very rapidly in which we can measure their error rate while recording (which generally means they have to sit still).