High precision coding: How the visual cortex processes information about the world
By Dominic Burrows
Dr. Carsen Stringer, of Howard Hughes Medical Institute, gave a seminar at the Centre for Developmental Neurobiology as part of the 2020-2021 “NEUReka!” seminar series. Dr. Stringer’s research focuses on understanding how the visual cortex encodes information about the world, and how such encodings influence behaviour. During the seminar Dr Stringer presented fascinating research, which she conducted as a post-doctoral researcher in the lab of Marius Pachitariu (Stringer et al., 2021).
Individual neurons provide a noisy and inaccurate representation of the outside world. Sensory neurons, such as those found in the visual cortex, are known to respond with high variability to repeated presentation of identical stimuli. How then does the brain make sense of the world around us? If each neuron independently carried information about a different object in the visual scene, then our perception of the same image would be different each time. One theory suggests that information is carried in large populations of neurons, most of which fire together in response to a given stimulus, to enable robust encoding of stimuli in the presence of single neuron noise. However, our understanding of how single neuron noise affects the reliability of population codes and ultimately behaviour, remains an open question. In the last decade, the development of two-photon microscopy has enabled the recording of tens of thousands of neurons deep within the cortex, allowing researchers to finally study the encoding of stimuli across large populations of neurons in the brain.
Dr Stringer set out to understand how single neuron variability might affect the ability of neural populations to reliably encode information about a stimulus. To do this, the Pachitariu lab recorded the activity of tens of thousands of neurons in the mouse primary visual cortex, whilst mice were shown gratings of different orientations over many repeats. Dr Stringer first found that individual neuron responses are highly noisy - each neuron unreliably fired in response to multiple presentations of their preferred stimulus orientation. In order to understand the fidelity of the representation of visual stimuli across the population rather than in single neurons, she used a decoder. This technique is a method of classification which trains a model to learn the mapping of different inputs (neural responses) onto a range of outputs (grating orientation). In this way it can inform us how much information about the stimulus is stored within the neural responses. Remarkably, she found that using neural population activity, the decoder could distinguish between angles that were up to 0.34 degrees different. This lies in stark contrast to the ability of mice to behaviourally distinguish between different orientations on a learning task (about 30 degrees). From these results, Dr Stringer suggests that neural populations encode visual information with high fidelity, even in the presence of single neuron noise.
Given this striking ability of the visual cortex to encode stimuli, why then is the animal much worse at behaviourally distinguishing them? To answer this, Dr Stringer re-analysed data from a previous study, which recorded from multiple cortical regions while mice performed a similar visual discrimination task. This time a decoder was trained to discriminate between visual stimuli, using neural activity on both correct and incorrect trials. This enabled the comparison of the decoders performance with the correct/incorrect choice of the animal. Interestingly, Dr Stringer found that in all brain regions except for primary sensory areas, as the animal began to respond to the stimulus, the decoder’s performance correlated more with choice than the stimulus itself. This suggests that as the signal leaves the visual cortex, noise from the internal state of the animal may interfere with this high fidelity signal.
Overall, Dr Stringer’s works helps to uncover the link between sensory encoding of information and behavioural responses. While population activity faithfully represents visual information in the visual cortex, this signal becomes disrupted by internal noise from other cortical regions. This may suggest a sub-optimal learning strategy for animals which is sufficient in real world environments.
Carsen Stringer, Michalis Michaelos, Dmitri Tsyboulski, Sarah E. Lindo, Marius Pachitariu, High-precision coding in visual cortex, Cell, 2021.
Banner image credit: Dchordpdx, CC BY 4.0