Have you ever marveled at how quickly you can process what’s going on in the world around you? It seems as if the incredible complexity of your mammalian brain, even with its billions of units working in parallel, would not be enough to perform these feats of mental prowess. In a recent paper , Gutnisky and Dragoi showed that efficient adaptation can increase the amount of information stored in a network of cells. This experimental confirmation of the ‘efficient coding hypothesis’ begins to elucidate strategies implemented in the neural architecture that allow for such massive information storage.

The experimental setup was beguilingly simple (by today’s standards): show a visually fixating monkey differentially oriented sine-wave gratings while recording from cells in primary visual cortex (V1). It is known that cortical cells will rapidly adapt after being exposed to redundant information for several hundred milliseconds. Thus, the authors recorded responses before and after a 400-ms adaptation to a sine-wave grating of fixed orientation. The test stimulus shown to the monkey after adaptation was a movie of sine-wave gratings presented at 60 Hz.

A pair of cells preferring nearby orientations (less than 30° apart) or very different orientations (more than 60°) exhibited a strong reduction in their correlated activity. In addition—and somewhat confusingly—cell pairs that preferred nearby orientations showed significant decorrelations only when the difference between the preferred orientation and the adapting stimulus was again small or large, i.e. less than 30 or more than 60 degrees. Finally, not only does adaptation reduce the strength of correlations, but it also reduces the variability of correlation.

The authors hypothesized that changes in coding elicited by adaptation would increase the amount of information stored in the network of cells. They estimated the information held in the population from the mean firing rate and covariance matrix of the individual units. This metric, the Fisher information, provides an upper bound that any putative neural decoder could use to extract the stimulus orientation. If changes in the mean and variability of the correlations are taken into account, the post-adaptation discrimination threshold is improved by 40%. Test stimuli that are similar or very different from the adapting stimulus provoked the largest improvement in coding efficiency. Finally, the authors point out that yielding a similar improvement by adjusting firing rate alone would have required an increase of 55%, certainly high enough to be metabolically costly.

This paper highlights one mechanism by which a system’s input-output curve can be modified on the fly. Other evidence along these lines has been accruing for years. Work by Reynolds et al. showed how attention can increase the sensitivity of V4 neurons. Even human psychophysical data, mentioned in the paper, are in agreement with this mechanism. In David McCormick’s lab, work has shown how network activity can influence individual neuron’s responsiveness, ultimately modifying how stimuli are represented (Hasenstaub et al.) Recent work highlighted by Cori Bargmann at the Neuronal Circuits meeting showed multiple ways in which the activity of identified circuits in C. elegans could be modified by the environment. She pointed out that even with knowledge of all the connections in this model organism, we may still understand less than half of how that system functions. These few examples underscore the importance of understanding the dynamics of a system, as Gutnisky and Dragoi have begun to do here.

Discussion Questions:
1. The authors mention that this mechanism is a “metabolically inexpensive solution.” What other advantages does it have?
2. The authors point out one downside of this mechanism is that it could interfere with other information processing. Are there other downsides to decorrelating inputs?
3. Is the proposal that this mechanism may optimize image-discrimination performance in real time feasible? Saccades may take anywhere from 20 to 200 ms, and occur many times per second. 400-ms adaptations are performed here. How many times can this adaptation occur in sequence?
4. As a rough generalization, one could think of this paper as highlighting the importance of circuit dynamics over circuit connectivity. Though the interconnections of the circuits here are not known, much can still be gleaned from how the system stores information (which after all, could be considered the ultimate goal in this instance.) If all-to-all connectivity is true, or if the specificity of connections is not high, then indeed it is not connectivity that matters but in fact dynamics. What impact will this have on the current era of ‘connectomics’?