Abstract:Neurons in the brain have rich and adaptive input-output properties. Features such as heterogeneous f-I curves and spike frequency adaptation are known to place single neurons in optimal coding regimes when facing changing stimuli. Yet, it is still unclear how brain circuits exploit single-neuron flexibility, and how network-level requirements may have shaped such cellular function. To answer this question, a multi-scaled approach is needed where the computations of single neurons and neural circuits must be considered as a complete system. In this work, we use artificial neural networks to systematically investigate single-neuron input-output adaptive mechanisms, optimized in an end-to-end fashion. Throughout the optimization process, each neuron has the liberty to modify its nonlinear activation function parametrized to mimic f-I curves of biological neurons, either by learning an individual static function or via a learned and shared adaptation mechanism to modify activation functions in real-time during a task. We find that such adaptive networks show much-improved robustness to noise and changes in input statistics. Using tools from dynamical systems theory, we analyze the role of these emergent single-neuron properties and argue that neural diversity and adaptation play an active regularization role, enabling neural circuits to optimally propagate information across time. Finally, we outline similarities between these optimized solutions and known coding strategies found in biological neurons, such as gain scaling and fractional order differentiation/integration. Evolution has shaped neural circuits in the brain to support complex tasks and behavior. In doing so, single neurons have developed intriguing coding properties such as heterogeneous and adaptive nonlinear responses, and mechanisms at network levels have emerged to efficiently propagate and process information. Discerning the impact of these distinct scales in shaping biophysical mechanisms is a difficult proposition. Are circuits wired the way they are because of single-neuron properties? Are single neurons behaving the way they do because of circuit requirements? To tackle this question, we turn to neural network models that we optimize with deep learning techniques. Crucially, we endow single neurons in these models with parameters that govern their input-output properties, as well as adaptation dynamics. These mechanisms are then optimized end-to-end together with network parameters to perform various perception-like tasks. While this process could have produced any novel solution, we find that what naturally emerges from optimization bears noteworthy similarity to biophysical mechanisms found in neurons in the brain, such as spike-frequency adaptation and gain scaling. We perform an in-depth analysis of the optimization process that uncovered these strategies and formulate hypotheses relating joint evolutionary pressures of single-neuron properties and circuit-level requirements in the brain.

Importance of Optimizing Neural Activation Function Types

Performance Analysis of Various Activation Functions in Artificial Neural Networks

An Ann Modeling Method With Optimizing Net Structure And Activation Function Types

An ANN model of optimizing activation functions based on constructive algorithm and GP

ON NEURAL NETWORK'S ACTIVATION FUNCTION OPTIMIZED BY GENETIC PROGRAMMING

Normalized Activation Function: Toward Better Convergence

Using Genetic Algorithms to Optimize Artificial Neural Networks

Optimizing neuronal activation function types based on GP in constructive FNN design

A Method on Searching Better Activation Functions

Activation Functions: Dive into an optimal activation function

Evolving activation dynamics in feedforward neural networks

Efficient Activation Function Optimization through Surrogate Modeling

A novel type of activation function in artificial neural networks: Trained activation function

Optimizing Neuron Function Types Based on GP in Functional Network Design

Activation Functions in Artificial Neural Networks: A Systematic Overview

Activation function optimization scheme for image classification

Studies on Optimization Algorithms for Some Artificial Neural Networks Based on Genetic Algorithm (GA)

Dynamic Neural Response Tuning

Neural network with optimal neuron activation functions based on additive Gaussian process regression

Neural networks with optimized single-neuron adaptation uncover biologically plausible regularization

Optimizing Neural Networks through Activation Function Discovery and Automatic Weight Initialization