Sunday, May 1, 2016

Audio Perception and Oscillation

Brains are reality modeling machines, which isolate surprising events for our protection and delectation. Does music have to be perpetually surprising, to be heard?

Imagine the most boring thing imaginable. Is it sensory deprivation? More likely it will something more active, like a droning lecturer, a chattering relative, or driving in jammed traffic. Meditation can actually be very exciting, (just think of Proust!), and sensory deprivation generates fascinating thought patterns and ideas. LSD and similar drugs heighten such internal experiences to the point that they can become life-altering. Which indicates an interesting thing about the nature of attention- that it is a precious resource that feels abused not when it is let loose, but when it is confined to some task we are not interested in, and particularly, that we are learning nothing from.

Music exists, obviously, not to bore us but to engage us on many levels, from the physical to the meditative and profound. Yet it is fundamentally based on the beat, which would seem a potentially boring structure. Beats alone can be music, hypnotically engaging, but typically the real business of music is to weave around the beat fascinating patterns whose charm lies in a tension between surprise and musical sense, such as orderly key shifts and coherent melody.

Why is all this attractive? Our brains are always looking ahead, forecasting what comes next. Their first rule is ... be prepared! Perception is a blend of getting new data from the environment and fitting it into models of what should be there. This has the virtues of providing understanding, since only by mapping to structured models of reality are new data understandable. Secondly, it reduces the amount of data processing, since only changes need to be attended to. And thirdly, it focuses effort on changing or potentially changing data, which are naturally what we need to be paying attention to anyhow ... the stuff about the world that is not boring.

"Predictive coding is a popular account of perception, in which internal representations generate predictions about upcoming sensory input, characterized by their mean and precision (inverse variance). Sensory information is processed hierarchically, with backward connections conveying predictions, and forward connections conveying violations of these predictions, namely prediction errors." 
"It is thus hypothesised that superficial cell populations calculate prediction errors, manifest as gamma-band oscillations (>30 Hz), and pass these to higher brain areas, while deep cell populations [of cortical columns] encode predictions, which manifest as beta band oscillations (12–30 Hz) and pass these to lower brain areas." 
"In the present study, we sought to dissociate and expose the neural signatures of four key variables in predictive coding and other generative accounts of perception, namely surprise, prediction error, prediction change and prediction precision. Here, prediction error refers to absolute deviation of a sensory event from the mean of the prior prediction (which does not take into account the precision of the prediction). We hypothesised that surprise (over and above prediction error) would correlate with gamma oscillations, and prediction change with beta oscillations."

A recent paper (and review) looked at how the brain perceives sound, particularly how it computes the novelty of a sound relative to an internal prediction. Prediction in the brain is known to resemble a Bayesian process where new information is constantly added to adjust an evolving model.

The researchers circumvented the problems of low-resolution fMRI imaging by using volunteers undergoing brain surgery for epilepsy, who allowed these researchers to study separate parts of their brains- the auditiory cortex- for purposes completely unrelated to their medical needs. They also allowed the researchers to only record from the surfaces of their brains, but to stick electrodes into their auditory cortexes to sample the cortical layers at various depths. It is well-known that the large sheet of the cortex does significantly different things in its different layers.

Frequencies of tones (dots) given to experimental subjects, over time.

The three subjects were played a series of tones at different frequencies, and had to do nothing in return- no task at all. The experiment was merely to record the brain's own responses at different positions and levels of the auditory cortex, paying attention to the various frequencies of oscillating electrical activity. The point of the study was to compare the data coming out with statistical models that they generated separately from the same stimuli- ideal models of Bayesian inference for what one would expect to hear next, given the sequence so far.

Electrode positions within the auditory areas of the subject's brains.

Unfortunately, their stimulus was not quite musical, but followed a rather dull algorithm: "For each successive segment, there is a 7/8 chance that that segment’s f [frequency] value will be randomly drawn from the present population, and a 1/8 chance that the present population will be replaced, with new μ [mean frequency] and σ [standard deviation of the frequency] values drawn from uniform distributions."

Correlations were calculated out between the observed and predicted signals, giving data like the following:

Prediction error and surprise are closely correlated, but the experimenters claim that surprise is a better correlated to the gamma band brain waves observed (B).

The difference between observation and prediction, and between surprise and prediction error. Surprise apparently takes into account the spread of the data, i.e. if uncertainty has changed as well as the mean predicted value.

What they found was that, as others have observed, the highest frequency oscillations in the brain correlate with novelty- surprise about how perceptions are lining up with expectations. The experimenter's surprise (S) measurement and prediction error (Xi) are very closely related, so they both correlate with each other and with the gamma wave signal. The surprise measure is slightly better correlated, however.

On the other hand, they observed that beta oscillations (~20 Hz) were correlated with changes in the predicted values. They hypothesized that beta oscillations are directed downward in the processing system, to shape and update the predictions being used at the prior levels.

Lastly, they find that the ~10 Hz alpha oscillations (and related bands) correlate with the uncertainty or precision of the predicted values. And theta oscillations at ~6 Hz were entrained to the sound stimulus itself, hitting when the next sound was expected, rather than encoding a derived form of the stimulus.

It is all a bit neat, and the conclusions are dredged out of very small signals, as far as is shown. But the idea that key variables of cognition and data processing are separated into different oscillatory bands in the auditory cortex is very attractive, has quite a bit of precedent, and is certainly an hypothesis that can and should be pursued by others in greater depth. The computational apparatus of the brain is very slowly coming clear.
"These are exciting times for researchers working on neural oscillations because a framework that describes their specific contributions to perception is finally emerging. In short, the idea is that comparatively slow neural oscillations, known as “alpha” and “beta” oscillations, encode the predictions made by the nervous system. Therefore, alpha and beta oscillations do not communicate sensory information per se; rather, they modulate the sensory information that is relayed to the brain. Faster “gamma” oscillations, on the other hand, are thought to convey the degree of surprise triggered by a given sound."

  • Bill Mitchell on the Juncker regime.
  • Who exactly is corrupt in Brazil, and how much?
  • There are too many people.
  • But not enough debt.
  • The fiscal "multiplier" is not constant.
  • Population has outstripped our willingness to build and develop.
  • What's going on in the doctor's strike?
  • Schiller on lying in business, Gresham's dynamics, and marketing.
  • Lying in religion.
  • Stiglitz on economics: "The strange thing about the economics profession over the last 35 year is that there has been two strands: One very strongly focusing on the limitations of the market, and then another saying how wonderful markets were."
  • Should banks be public institutions?
  • Does democratic socialism have a future in Russia?
  • A Sandersian / Keynesian stimulus is only effective if the Fed plays along.
  • Science yearns to be free.
  • Trump's brush with bankruptcy and friends in high places.

No comments:

Post a Comment