Efficient coding with biophysical realism

Theoretical framework. Efficient coding (EC), an application of Shannons information theory to neural systems, posits that these systems transform natural stimuli into neural spikes optimally. EC has successfully predicted a number of response properties of sensory periphery from the statistics of natural signals. However, most applications have focused on functional models which ignore essential constraints, such as biochemical/biophysical realism at the cellular or micro-circuit scale, or connectivity patterns at the network scale. Specifically, there is no EC-principle-based derivation of: (A) a mechanistic model of any early vertebrate sensory cascade; (B) optimal top-down and lateral/recurrent connectivity in a multi-layer sensory circuit. Research questions. We hypothesize that the inclusion of biophysical realism will qualitatively improve the predictive power of EC. We will demonstrate this on two systems and across two scales. In Aim A (cellular/micro-circuit scale) we will derive optimal acoustic information encoding in a biophysically-realistic model of the mammalian cochlea. In Aim B (network scale) we will derive optimal visual information encoding in a cortical population with dynamically-adjustable top-down and recurrent connectivity. Approach. In both Aims, we will set up a computational model of information processing that includes key aspects of realism. Specifically, Aim A will capture known aspects of cochlear mechanics, mechanoelectrical transduction, synaptic transmission at inner hair cells, and spike generation in the auditory nerve; Aim B will capture known tuning properties of V1 cells but include the possibility that these cells and their lateral interactions are gained up- or down by feedback from higher-order areas. We will derive EC predictions by large-scale numerical optimization of model parameters. Level of originality. Essential characteristics of neurosensory systems such as biophysical organization underlying spectral decomposition in auditory nerve (Aim A) or a plethora of attentional effects (Aim B) have never been derived directly from EC theory, so it is unclear whether and when they constitute optimal adaptations to natural stimuli. Current theories of EC are furthermore incapable of quantitatively predicting many system-level quantities that can be empirically measured, such as biophysical properties of neurons (Aim A), or patterns of noise correlations among neurons (Aim B). Our proposal addresses these issues for the first time and enables potential applications for improved cochlear implants (Aim A) or signal compression (Aim B). Primary researchers involved. PI Tkacik and two postdoctoral fellows, Gabrielaitis and Mlynarski, have a successful track record of collaboration and advancing EC theory and application to data. Collaboration with two experimental labs will help us develop data analysis techniques alongside the models to permit future rigorous statistical tests of model predictions.

The project's main objective was to explore new implications of the "efficient coding hypothesis," a foundational concept in sensory and computational neuroscience. This hypothesis, originating in the 1960s, posits that neural systems have evolved to encode natural sensory stimuli with maximum efficiency. Over recent decades, it has successfully predicted several quantitative features of sensory organs like the cochlea and retina, and some sensory neuron response properties in the central brain, such as the primary visual cortex. Our hypothesis was that combining the efficient coding hypothesis with known biophysical constraints could yield new predictions. Through various projects, we derived and, in some cases, confirmed these extensions via data analyses, covering a range of nervous systems. In the first paper, Mantas Gabrielaitis developed a new wide-band signal demodulation technique, a key step for analyzing time-varying signals. Inspired by the desire to understand the inner ear's architecture, this new demodulation approach and accompanying fast algorithms significantly advanced the state of the art in sound processing, with possible applications to wireless telephony, ultrasound medical imaging, and beyond. The second paper was a collaborative theoretical-experimental study with the group of Maximillian Jösch on mouse retinal ganglion cells. These cells signal high or low visual contrast through their center-surround receptive fields. We found that cells from different retinal regions exhibit variability in their receptive fields. Efficient coding quantitatively predicted this variability as an adaptation to visual signals whose properties differ across different parts of ecologically relevant visual scenes that mice live in, for example above or below the horizon. The third paper investigated visual processing in the cortex, where cells respond to small, oriented bars of light ("edges"). We explored how feedback from higher brain areas, in addition to the traditionally studied feed-forward input from the retina, modulates neural processing to compress visual signals without compromising behavioral performance. This led to the surprising realization that our coding theory predicts various "top-down attentional phenomena," marking the first theoretical derivation of attentional modulation from first principles. The fourth paper analyzed hippocampal spatial coding data from Jozsef Csicsvari's lab. Efficient coding was suggested to predict not only individual neuron properties but also their interactions. By observing how a rodent's hippocampus learns a new environment's spatial map and developing a new statistical methodology to extract cell-cell interactions from neural recordings, we confirmed our hypothesis. Efficient coding can predict optimal neuron interactions based on individual response properties and system noise levels.