Energy efficiency as a normative account for predictive coding

Summary

In this issue of Patterns, Ali et al. demonstrate that predictive coding emerges in an artificial neural network optimized to be energy efficient. The results offer an explanation for why brains may implement predictive coding.

In this issue of Patterns, Ali et al. demonstrate that predictive coding emerges in an artificial neural network optimized to be energy efficient. The results offer an explanation for why brains may implement predictive coding.

Main text

Understanding what the brain does has occupied physiologists and neuroscientists for more than a century. Our current understanding of the brain gives us a set of functions that can be attributed to different parts of the brain and how they may underlie different behaviors¹— though brain scientists have long wished that one single function could be associated with the entire organ, much as is done with other body organs (e.g., heart).² The majority of ideas with that ambitious goal, however, have been disregarded as being incorrect or overly simplistic. The “prediction machine,” a century-old theory of brain function,³ has been an exception.⁴

According to this theory, the brain relies on an internal model of the world to predict what will happen next given its past sensory inputs as well as its own planned and executed actions. A specific variant of this theory, known as predictive coding,⁵ has become one of the most influential theories in neuroscience and encouraged more theoretically driven investigations into the brain. In this issue of Patterns, Ali et al.⁶ demonstrated that predictive coding is an emergent property of a neural network that functions with limited energy budget: a link that strengthens predictive coding as a unified theory of brain function.

Over decades of work, a body of empirical observations has been accumulated highlighting physiological and anatomical features of the brain that support its function as a prediction machine. Among the strongest evidence are the distinct dynamics of brain activity in predictable versus unpredictable environments.^7,8 If prediction is, at least in part, what the brain does, the next crucial question is how the brain performs prediction. Predictive coding^5,9 is a theory that offers an algorithmic model for how prediction could potentially be performed by the brain. It suggests a very specific algorithm and neural architecture for updating the brain’s internal model of the world and using that internal model for prediction. Predictive coding asserts that information flows in a particular way between brain areas: each brain area contains two distinct populations of prediction and error neurons. The prediction neurons in a brain area (e.g., the visual area V2) send a prediction of what should happen next to its lower area (e.g., the visual area V1) while the error neurons of the lower area pass an error of that prediction back to the higher area. Through this hierarchical message passing, brains predict their sensory inputs and use the error signals to modify their generative model of the world. The specific architecture that predictive coding requires (i.e., top-down prediction and distinct prediction and error neurons) beg the question of why the brain is evolved to implement predictive coding.

Ali et al.⁶ offered an intriguing answer to this question that involves a fundamental constraint of the brain: its limited energy budget. From the energy consumption perspective, brains are expensive organs. The human brain is approximately 2% of the body’s mass but consumes about 20% of the body’s energy production. A large part of this energy is consumed to maintain communication between neurons (synaptic transmission). Therefore, neuronal structures that maintain the required function while keeping the energy consumption at minimum have survived throughout evolution.¹⁰ The idea of minimizing the synaptic energy consumption is at the core of the computational model proposed by Ali et al.⁶ The authors fed a recurrent artificial neural network (RNN) with a fully predictable input stream (a predictable sequence of images). Energy minimization was the only objective function used for optimizing the RNN. The connection weights of the RNN were all tuned to minimize the total energy consumption of the network. The authors used the artificial neurons’ preactivation (the sum of the synaptic inputs to a neuron) as a measure of the RNN’s energy consumption. They discovered that by minimizing the energy consumption of the RNN, the model learned to predict the stream of its sensory input (i.e., predict the next image in a sequence). In other words, in a predictable environment, the RNN became a “prediction machine” in order to minimize its energy consumption. Models that were trained using alternative measures of energy consumption (such as postactivation or neuron outputs) did not learn to predict the sensory stream, making the use of preactivation as a measure of energy expenditure essential.

Most interestingly, further examinations revealed that the energy-efficient RNN was, in fact, an implementation of predictive coding with emerging top-down prediction and distinct prediction and error neurons. The specialized prediction and error units were identified in the RNN based on their expected functional roles: the prediction units provide sensory predictions while the error units report deviation from the prediction. After identifying the two types of neurons, the authors also did an extensive analysis of the prediction and error neurons, their circuitries, and the role that they play in performing prediction in the RNN. Lesioning the specialized neurons also confirmed their proposed role in the function of the RNN. None of these characteristics were imposed by design in the model but rather emerged in a network that had to limit its energy consumption while functioning in a predictable environment. These results offer a simple yet elegant account of why brains might implement predictive coding.

Despite the limited energy budget, brains should still be able to integrate external and internal information to guide behavior. As recent studies have shown, at least some functional properties of brain areas (e.g., response properties of visual areas) can be best understood as emergent properties of a neural network optimized for ecologically relevant behaviors (e.g., object categorization, or self-motion estimation).^11,12 An exciting direction for future research would be to investigate how constraining the energy budget affects performance in ecologically relevant tasks. Does energy limitation deteriorate behavioral performance, or is it a bottleneck that can act as an implicit regularization for discovering more intelligent, generalizable solutions? In addition to their intriguing discovery, Ali et al.⁶ provide a concrete framework for combining the notion of energy efficiency with other ecologically relevant constraints for modeling the brian.