The mechanisms of reliable coding in mouse visual cortex

Thesis: Ph. D. in Neuroscience, Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2016.Cataloged from PDF version of thesis. Page 262 blank.Includes bibliographical references.As we interact with the environment, our senses are constantly bombarded with information. Neurons in the visual cortex have to transform these complex inputs into robust and parsimonious neural codes that effectively guide behavior. The ability of neurons to efficiently convey information is, however, limited by intrinsic and shared variability. Despite this limitation, neurons in primary visual cortex (V1) are able to respond with high fidelity to relevant stimuli. My thesis proposes that high fidelity encoding can be achieved by dynamically increasing trial-to-trial response reliability. In particular, in this thesis, I use the mouse primary visual cortex (V1) as a model to understand how reliable coding arises, and why it is important for visual perception. Using a combination of novel experimental and computational techniques, my thesis identifies three main factors that can modulate intrinsic variability. My first goal was to understand the extrinsic, stimulus-dependent, factors responsible for reliably coding (Chapter 3). Natural scenes contain unique statistical properties that could be leveraged by the visual cortex for efficient coding. Thus, the first aim is to elucidate how image statistics modulate reliable coding in V1. To this end, I developed a novel noise masking procedure that allowed us to specifically perturb the spectral content of natural movies without altering the edges. Using high-speed twophoton calcium imaging in mice, I discovered that movies with stronger spatial correlations are more reliably processed by V1 neurons than movies lacking these correlations. In particular, perturbing spatial correlations in the movie dynamically altered the structure of interneuronal correlations. Movies with more naturalistic correlations typically recruited large neuronal ensembles that were weakly noise correlated. Using computational modeling, I discovered that these ensembles were able reduce shared noise through divisive normalization. Together, these findings demonstrate that natural scene statistics dynamically recruit neuronal ensembles to ensure reliable coding. Microcircuits of inhibitory interneurons lie at the heart of all cortical computations. It has been proposed that these interneurons are responsible for reliable spiking by controlling the temporal window over which synaptic inputs are integrated. However, no study has yet conclusively investigated the role of different interneuron subtypes. Thus, my second goal was to establish how natural scenes are reliably encoded by dissecting the inhibitory mechanisms underlying reliable coding (Chapter 4). Specifically, I investigated the role of somatostatin-expressing dendrite targeting interneurons (SST) and parvalbumin-expressing soma targeting interneurons (PV), which are known to provide distinct forms of inhibition onto pyramidal neurons. Using a novel combination of dual-color calcium imaging and optogenetic manipulation, I have discovered that the SST->PV inhibitory circuit plays a crucial role in modulating pyramidal cell reliability. In particular, by transiently suppressing PV neurons, SST neurons are able to route inhibition rapidly from the soma to the dendrites. Strong dendritic inhibition allows noisy inputs to be filtered out by the dendrites, while weaker somatic inhibition allows these inputs to be integrated to produce reliable spikes. In agreement with these results, I found that selectively deleting MeCP2 from these interneurons resulted in unreliable visual processing and other circuit-specific deficits, which are commonly observed in Rett Syndrome (Chapter 5). These results underscore the importance of intact inhibitory microcircuits in reliable processing. Finally, my goal was to determine why reliable coding is necessary for visual processing (Chapter 6). To this end, I trained head-fixed mice to perform a natural movie discrimination task. Mice were able to learn how to discriminate between two movies after a short training period. By perturbing the amplitude spectrum of these movies, I discovered that mice used structural information in the phase spectrum to discriminate between the different movies. This suggests that mice also use similar strategies as higher mammals for scene recognition. Inspired by this result, we trained mice on a harder target categorization task, where mice had to identify the movies from an ensemble that were more similar to the target movie to gain a water reward. We developed this movie ensemble by blending together the phase spectrum of a target and non-target movie at different fractions. Optically activating SST neurons in V1 improved the ability of mice to correctly identify "target-like" movies. This increase in behavioral performance correlated well with an increase in V1 coding reliability. Thus, reliable codes are a prerequisite for accurate visual perception. Taken together, this work bridges the gap between cells, circuits and behavior, and provides mechanistic insight into how complex visual stimuli are encoded with high fidelity in the visual cortex.by Rajeev V. Rikhye.Ph. D. in Neuroscienc