Towards the neural population doctrine

Highlights

• New generations of recording and computing technologies have enabled neuroscience at the level of the neural population.

• Landmark scientific findings suggest the arrival of the neural population doctrine, where the population, not the single neuron, is the central object of computation.

• State space models offer a principled framework for population-level analyses.

• Four subfields of neuroscience have benefitted particularly from population analyses; we highlight key findings.

Across neuroscience, large-scale data recording and population-level analysis methods have experienced explosive growth. While the underlying hardware and computational techniques have been well reviewed, we focus here on the novel science that these technologies have enabled. We detail four areas of the field where the joint analysis of neural populations has significantly furthered our understanding of computation in the brain: correlated variability, decoding, neural dynamics, and artificial neural networks. Together, these findings suggest an exciting trend towards a new era where neural populations are understood to be the essential unit of computation in many brain regions, a classic idea that has been given new life.

Introduction

The neuron has been heralded as the structural and functional unit of neural computation for more than a hundred years [1, 2, 3], and this doctrine has driven a vast array of our field's most important discoveries [4]. Increasingly prevalent, however, is the idea that populations of neurons may in fact be the essential unit of computation in many brain regions [4, 5, 6]. Certainly the idea of information being encoded in an ensemble of neurons is not a new one. In fact, the conflict between single neurons and populations of neurons as the perceptual, behavioral, and cognitive unit of computation dates back to the beginning of the 20th century [4,6], with the first concrete theories of networks of neurons introduced in the 1940s [7,8]. The ideas, while being novel, were not testable due to the technological shortcomings of both recording techniques and computational resources. However, we currently find ourselves in the ideal era for scientific discovery, given the astounding progresses in both these enabling technologies.

Electrophysiology recordings have been the hallmark of neuronal recordings over the last 80 years — extracellular recordings of one or multiple electrodes, each capturing up to a few neurons. More recently, multi-electrode arrays and imaging techniques (optical, and more recently voltage) have been used to efficiently capture the simultaneous activity of hundreds and thousands of neurons, with this number steadily growing using tools such as the Neuropixel [9]. In tandem, the increase in computational resources has led to the development of efficient and scalable statistical and machine learning methods; see the methodological reviews [10,11].

As our ability to simultaneously record from large populations of neurons is growing exponentially [9,12], the analysis of the covariation of populations of neurons has provided us with scientific insights in many domains. Here, we highlight several recent findings in four domains of neuroscience where the joint analysis of a population of neurons has been central to scientific discovery that would not be possible using single neurons alone. Firstly, trial-to-trial ‘noise’ correlations have been shown to influence the information carrying capacity of a neural circuit. Secondly, decoding of behavior using correlated populations of neurons can yield levels of accuracy beyond what would be anticipated from single neurons alone. Thirdly, the dynamic analysis of stimulus-driven population recordings over time can be projected into a lower dimensional subspace to reveal computational strategies employed by different brain regions. Lastly, artificial neural networks (ANNs) can aid in simulations that reproduce population structure, as well as directly modeling neuronal activity. We focus on the analysis of a population of N neurons in ‘state space’, where each neuron's activity at any time point is represented as a dot in either the N dimensional observation space or in a lower dimensional subspace.

We pinpoint one or two recent studies in each domain that stand out (indicated using ^* and ^**). Unlike previous reviews on population-level neuroscience [10,11], we focus here not on the data analysis methodologies, but rather the notable scientific findings that have resulted. These scientific findings are, first, reshaping the way the field thinks about computation, and, second, fundamentally population-based. Taken together, these two features point to a future where the central scientific theme is not the neuron doctrine, but the neural population doctrine. We conclude with topics that we think future studies may need to address.