Leggi in italiano Neuroscientists believe that the human sensory system adapts to the natural world according to an ‘efficient coding’ principle. It maximizes the amount of information processed, while minimizing energy consumption. When it comes to vision, in particular, they have suggested that humans achieve this efficient coding by being sensitive to the statistical properties of images. When we observe a black-and-white texture, for example, we do not actually perceive all the pixels, but rather the correlations among adjacent pixels that create recurring schemes, such as repetitions of horizontal segments or angles. This way we use less computation, and less energy, to interpret the image.
Now, researchers coordinated by Davide Zoccolan, who leads the Visual Neuroscience Lab at SISSA in Trieste, and Vijay Balasubramanian of the University of Pennsylvania, have shown that rats apply the same strategy, a result that paves the way to more controlled experiments. Their results have been published in the journal eLife1.
The researchers adapted to rats an experiment performed in 2014 by Balasubramanian, that showed that humans are more sensitive to the types of statistical patterns that are more variable in natural images. In particular, we prioritize combinations of pixels that create horizontal edges, squares, and L-shapes.
In the new experiment, 42 rats were trained to discriminate between white noise textures (where each pixel has the same probability of being black or white, regardless of the other pixels), and structured textures with recurring patterns. If they got it right, they were rewarded. The training phase took several weeks, during which the rats were first exposed to textures where the patterns were more evident, and then trained to recognize weaker correlations. The rats that reached a sufficient discrimination ability were admitted to the testing phase, where they received no feedback about their choices. The results show that rats have maximum sensitivity for horizontal edges, followed by squares and L-shapes, the same ranking observed in humans in 2014.
“Beside suggesting that efficient coding for vision is a general principle in mammals,” says Zoccolan, “our result opens the possibility of causal manipulation to understand the underlying neural mechanisms”. For example, scientists could rear newborn rats in an artificial visual environment where some correlations are less variable than in nature, and check whether their visual system would adapt to this different statistic.
“This would allow us to understand whether the adaptation of the visual system to natural scenes is wired at birth, or instead takes place later” Balasubramanian says. “Unless you intervene in the early stage of development you cannot tell these two hypotheses apart”. Such experiments cannot be performed with human subjects and are much more difficult in primates.
“To make a bridge between human and rat studies, the investigators needed to make some compromises”, says Jonathan Victor, neuroscientist at the Weill Cornell Medical College who developed the framework to test this formulation of the efficient coding hypothesis in vision. “This raises some interesting questions for future work, regarding the level of specificity at which sensory systems are tuned to their environments and behavioural needs.”