In the past, I have conducted a number of projects related to population coding and noise correlations in visual cortex and using statistical modeling and machine learning for neural data analysis.
- Representation of natural scenes in V1: The response properties of neurons in the visual cortex are thought to be adapted to the statistics of natural stimuli. I collaborated with the lab of Andreas Tolias, to study how correlations in the visual stimulus are transformed into correlations between neurons using two-photon imaging in mice. We found that input with natural higher order correlations increased population sparseness in the neural responses. This increase changed the population representation such that different frames from the natural movie were easier to discriminate, demonstrating a functioncal benefit of population sparseness (Nature Neuroscience, 2014). [more]
- Noise correlations in V1: Neurons in sensory areas are noisy and a certain fraction of this noise is correlated among nearby neurons. These noise correlations can greatly affect the coding accuracy of neural populations. It was widely believed that noise correlations are quite strong (~0.1-0.2), based on measurements in anesthetized animals. We measured noise correlations in awake animals using chronically implanted tetrodes and found them to be surprisingly weak (Science, 2010). More recently, we performed additional measurements in anesthetized animals. We used a latent variable model of the population activity to show that anesthesia induces a comodulation common to all neurons, increasing estimates of noise correlations (Neuron, 2014).
- Decoding large neural populations: We used logisitic regression to decode orientation in a classification task from the activity of up to 20 neurons in V1 to study how neural populations represent sensory information. We found that the full information in the population could be decoded fast (within ~50 ms) with a decoder that was the same across contrasts and ignored correlations (Journal of Neuroscience, 2012). More recently, we extended this line of work to large populations (~300 neurons) recorded from mouse V1 using 3D AOD two-photon imaging. We found that informations scales linearly with population size and noise correlations do not matter for decoding up to this population size (paper currently under review).
- Theoretical analysis of population coding in V1: We analyzed neural population codes from a theoretical perspective. In one project, we studied the validity of Fisher information as a tool for infering optimal population codes, e.g. optimizing the tuning width of a population of V1 neurons. We found that Fisher information fails in particular at short integration times or when SNR is low – ideal observer analysis yields much wider tuning functions than Fisher information does (PNAS, 2011). In additon, we studied how heterogenous populations of neurons (e.g. with variable tuning width) encode orientation and found that here information does not saturate in the presence of noise correlations in contrast to homogenous populations (Journal of Neuroscience, 2011).