The influence of population size, noise strength and behavioral task on best-encoded stimulus for neurons with unimodal or monotonic tuning curves

Adaptive Tuning Curve Widths Improve Sample Efficient Learning

Natural brains perform miraculously well in learning new tasks from a small number of samples, whereas sample efficient learning is still a major open problem in the field of machine learning. Here, we raise the question, how the neural coding scheme affects sample efficiency, and make first progress on this question by proposing and analyzing a learning algorithm that uses a simple reinforce-type plasticity mechanism and does not require any gradients to learn low dimensional mappings. It harnesses three bio-plausible mechanisms, namely, population codes with bell shaped tuning curves, continous attractor mechanisms and probabilistic synapses, to achieve sample efficient learning. We show both theoretically and by simulations that population codes with broadly tuned neurons lead to high sample efficiency, whereas codes with sharply tuned neurons account for high final precision. Moreover, a dynamic adaptation of the tuning width during learning gives rise to both, high sample efficiency and high final precision. We prove a sample efficiency guarantee for our algorithm that lies within a logarithmic factor from the information theoretical optimum. Our simulations show that for low dimensional mappings, our learning algorithm achieves comparable sample efficiency to multi-layer perceptrons trained by gradient descent, although it does not use any gradients. Furthermore, it achieves competitive sample efficiency in low dimensional reinforcement learning tasks. From a machine learning perspective, these findings may inspire novel approaches to improve sample efficiency. From a neuroscience perspective, these findings suggest sample efficiency as a yet unstudied functional role of adaptive tuning curve width.

The impact of pathological high-frequency oscillations on hippocampal network activity in rats with chronic epilepsy

In epilepsy, brain networks generate pathological high-frequency oscillations (pHFOs) during interictal periods. To understand how pHFOs differ from normal oscillations in overlapping frequency bands and potentially perturb hippocampal processing, we performed high-density single unit and local field potential recordings from hippocampi of behaving rats with and without chronic epilepsy. In epileptic animals, we observed two types of co-occurring fast oscillations, which by comparison to control animals we could classify as ‘ripple-like’ or ‘pHFO’. We compared their spectral characteristics, brain state dependence, and cellular participants. Strikingly, pHFO occurred irrespective of brain state, were associated with interictal spikes, engaged distinct subnetworks of principal neurons compared to ripple-like events, increased the sparsity of network activity, and initiated both general and immediate disruptions in spatial information coding. Taken together, our findings suggest that events that result in pHFOs have an immediate impact on memory processes, corroborating the need for proper classification of pHFOs to facilitate therapeutic interventions that selectively target pathological activity.

A Population Representation of Absolute Light Intensity in the Mammalian Retina

Environmental illumination spans many log units of intensity and is tracked for essential functions that include regulation of the circadian clock, arousal state, and hormone levels. Little is known about the neural representation of light intensity and how it covers the necessary range. This question became accessible with the discovery of mammalian photoreceptors that are required for intensity-driven functions, the M1 ipRGCs. The spike outputs of M1s are thought to uniformly track intensity over a wide range. We provide a different understanding: individual cells operate over a narrow range, but the population covers irradiances from moonlight to full daylight. The range of most M1s is limited by depolarization block, which is generally considered pathological but is produced intrinsically by these cells. The dynamics of block allow the population to code stimulus intensity with flexibility and efficiency. Moreover, although spikes are distorted by block, they are regularized during axonal propagation.

Determine Neuronal Tuning Curves by Exploring Optimum Firing Rate Distribution for Information Efficiency

This paper proposed a new method to determine the neuronal tuning curves for maximum information efficiency by computing the optimum firing rate distribution. Firstly, we proposed a general definition for the information efficiency, which is relevant to mutual information and neuronal energy consumption. The energy consumption is composed of two parts: neuronal basic energy consumption and neuronal spike emission energy consumption. A parameter to model the relative importance of energy consumption is introduced in the definition of the information efficiency. Then, we designed a combination of exponential functions to describe the optimum firing rate distribution based on the analysis of the dependency of the mutual information and the energy consumption on the shape of the functions of the firing rate distributions. Furthermore, we developed a rapid algorithm to search the parameter values of the optimum firing rate distribution function. Finally, we found with the rapid algorithm that a combination of two different exponential functions with two free parameters can describe the optimum firing rate distribution accurately. We also found that if the energy consumption is relatively unimportant (important) compared to the mutual information or the neuronal basic energy consumption is relatively large (small), the curve of the optimum firing rate distribution will be relatively flat (steep), and the corresponding optimum tuning curve exhibits a form of sigmoid if the stimuli distribution is normal.

Matched Behavioral and Neural Adaptations for Low Sound Level Echolocation in a Gleaning Bat, Antrozous pallidus

In active sensing, animals make motor adjustments to match sensory inputs to specialized neural circuitry. Here, we describe an active sensing system for sound level processing. The pallid bat uses downward frequency-modulated (FM) sweeps as echolocation calls for general orientation and obstacle avoidance. The bat's auditory cortex contains a region selective for these FM sweeps (FM sweep-selective region, FMSR). We show that the vast majority of FMSR neurons are sensitive and strongly selective for relatively low levels (30-60 dB SPL). Behavioral testing shows that when a flying bat approaches a target, it reduces output call levels to keep echo levels between ∼30 and 55 dB SPL. Thus, the pallid bat behaviorally matches echo levels to an optimized neural representation of sound levels. FMSR neurons are more selective for sound levels of FM sweeps than tones, suggesting that across-frequency integration enhances level tuning. Level-dependent timing of high-frequency sideband inhibition in the receptive field shapes increased level selectivity for FM sweeps. Together with previous studies, these data indicate that the same receptive field properties shape multiple filters (sweep direction, rate, and level) for FM sweeps, a sound common in multiple vocalizations, including human speech. The matched behavioral and neural adaptations for low-intensity echolocation in the pallid bat will facilitate foraging with reduced probability of acoustic detection by prey.