Loading field map
Systems neuroscience is the branch of neuroscience that studies how networks of neurons—from small local circuits to whole-brain systems—produce behavior, perception, cognition, and movement. Where molecular neuroscience asks what molecules make neurons work, and cellular neuroscience asks how individual neurons integrate signals, systems neuroscience asks a different question: how do the interactions of many neurons give rise to the brain's functions? Its central premise is that the explanatory key to brain function lies not in any single neuron but in the patterns of activity distributed across populations of cells and in the rules that govern their communication.
The field is organized around a small set of enduring questions. The first is neural coding: how is information about the world represented in the activity of neurons? A classic formulation asks whether individual neurons act as feature detectors—each cell responding selectively to a specific stimulus, such as a particular face or direction of motion—or whether information is distributed across populations, with any given stimulus represented by the pattern of activity across many cells. Both answers contain truth, and much of the field's history has been an attempt to specify the conditions under which each holds.
The second question concerns transformation: how does neural activity change as it passes through successive processing stages? The visual system provides the canonical example. Light hits the retina, and signals travel through the lateral geniculate nucleus of the thalamus to the primary visual cortex, then onward through a hierarchy of cortical areas. At each stage, the representation changes: neurons respond to more complex features, their receptive fields grow larger, and their responses become more invariant to irrelevant stimulus properties. Understanding the computational logic of these transformations is a core goal.
The third question is integration and coordination: how do distributed brain regions work together to produce unified behavior? A visual scene, a memory, and a motor command must be bound together into a coherent experience and action. This requires understanding not just what each region does, but how regions communicate, synchronize, and influence one another. This question has grown in importance as techniques for recording from many brain areas simultaneously have matured.
The fourth question is state and modulation: how do the brain's internal states—arousal, attention, motivation, sleep—shape its processing? The same sensory input can be processed differently depending on whether an animal is alert or drowsy, attending to the stimulus or distracted. This introduces a layer of complexity beyond the feedforward processing described above, because it means the brain's output is not a simple function of its input.
Systems neuroscience emerged as a distinct enterprise in the mid-twentieth century, but its roots lie in earlier traditions. In the late nineteenth and early twentieth centuries, anatomists and physiologists mapped the brain's gross structure and identified the major pathways connecting sensory organs to cortical areas. The Spanish anatomist Santiago Ramón y Cajal's demonstration that neurons are discrete cells, communicating at synapses, provided the cellular foundation on which all later systems work would build. In the early twentieth century, the British physiologist Charles Sherrington studied reflexes and coined the term "synapse," establishing the idea that behavior could be understood as the coordinated action of neural circuits.
The field's modern identity crystallized around a set of mid-century discoveries. In the 1950s and 1960s, David Hubel and Torsten Wiesel, working at Harvard, recorded from single neurons in the visual cortex of anesthetized cats and monkeys. They found that neurons responded selectively to oriented edges and bars of light, and that these neurons were organized into columns of similar selectivity. This work established the single-neuron recording approach as the dominant method for studying sensory processing and provided a concrete example of how a complex perceptual function—vision—could be decomposed into the responses of identifiable cells. It earned them the Nobel Prize in Physiology or Medicine in 1981.
Around the same period, the study of the motor system took shape through the work of Edward Evarts, who recorded from neurons in the motor cortex of awake monkeys trained to perform precise movements. This demonstrated that it was possible to study the neural basis of voluntary action, not just reflexes, and opened the door to studying cognition—attention, decision-making, memory—at the level of single neurons in behaving animals.
A parallel tradition developed from computational and theoretical work. In 1943, Warren McCulloch and Walter Pitts showed that networks of simple threshold units could compute logical functions, suggesting that the brain might be understood as a kind of computing machine. In the 1980s, the connectionist movement, associated with David Rumelhart and James McClelland, revived interest in neural network models that could learn from experience. These theoretical efforts provided a vocabulary for thinking about what neural circuits might be doing, even when the experimental techniques to test those ideas directly did not yet exist.
Systems neuroscience is not organized into sharply separated schools, but it does contain distinct traditions that differ in their methods, assumptions, and explanatory goals. These approaches coexist and increasingly combine, but each has its own logic.
The oldest and most established approach is the recording of action potentials from individual neurons, either one at a time or, more recently, from many neurons simultaneously using multi-electrode arrays. The core assumption is that the action potential—the brief electrical pulse that travels along a neuron's axon—is the fundamental unit of neural communication, and that understanding how these pulses encode information is the key to understanding brain function.
This approach has been extraordinarily productive. It has yielded detailed maps of sensory cortices, revealed the selectivity of neurons for faces, places, and objects, and shown how neurons in the motor cortex represent the direction of movement. Its strength is its directness: when a neuron fires, you know it fired, and you can correlate that firing with stimuli and behavior. Its limitation is that it captures only a tiny fraction of the neurons in any circuit. The brain contains billions of neurons; recording from even a hundred simultaneously samples a minuscule fraction. This raises the question of whether the properties observed in recorded neurons are representative, and whether the code is best understood at the level of individual cells or populations.
Partly in response to the limitations of single-neuron recording, a newer approach emphasizes the simultaneous activity of many neurons and seeks to understand the structure of that population activity. The central idea is that the relevant variable is not the firing of any single neuron but the pattern of activity across the population, which can be thought of as a point in a high-dimensional space. Because the activity of neurons is correlated, the population activity typically occupies a lower-dimensional subspace—a smaller set of meaningful dimensions along which the brain's computations unfold.
This approach has been particularly influential in motor neuroscience. Work by Krishna Shenoy, Mark Churchland, and others has shown that the activity of populations of motor cortical neurons during movement follows stereotyped trajectories through a low-dimensional space, and that these trajectories can be understood as reflecting the dynamics of the local circuit rather than simply a feedforward transformation of sensory input. This has led to a shift in how the motor cortex is conceptualized: not as a passive map of movement parameters, but as a dynamical system whose internal state evolves over time to generate the pattern of muscle activations.
The population approach has also transformed sensory neuroscience. Instead of asking what each neuron responds to, researchers ask how the pattern of activity across many neurons discriminates between stimuli. This has led to the development of sophisticated decoding methods—techniques for reading out what stimulus was presented or what decision was made from the pattern of neural activity. These methods have shown that information is often present in populations even when no single neuron is highly selective, and that the brain may use population codes that are more robust and efficient than single-neuron codes.
A third approach operates at a larger scale, using functional magnetic resonance imaging (fMRI), electroencephalography (EEG), and magnetoencephalography (MEG) to study the activity of whole brain regions and their interactions. This tradition, which emerged in the 1990s and grew rapidly with the spread of fMRI, asks how cognitive functions—attention, memory, decision-making, emotion—are implemented in the coordinated activity of distributed networks.
This approach has identified large-scale networks that are consistently engaged across tasks and individuals, such as the default mode network (active when the mind is at rest and wandering) and the frontoparietal attention network. It has also developed methods for studying functional connectivity—the statistical dependencies between the activity of different regions—which can reveal how regions communicate even when they are not directly connected anatomically.
The strength of this approach is its ability to study the human brain non-invasively and to capture the whole-brain context in which local circuits operate. Its limitation is spatial and temporal resolution: fMRI measures changes in blood flow that are indirect proxies for neural activity and unfold over seconds, while the brain's computations happen in milliseconds. This means that fMRI can tell you which regions are involved in a task, but not precisely what those regions are computing.
A fourth tradition is explicitly theoretical. Its practitioners build mathematical models of neural circuits and use them to explore what computations are possible, what mechanisms could implement them, and what predictions follow from different hypotheses. This tradition includes models at many levels: biophysically detailed models of individual neurons, simplified models of neural networks, and abstract models of computation.
The relationship between theory and experiment in systems neuroscience is complex. Some theories are tightly coupled to specific experimental systems, such as models of how the retina encodes visual scenes or how the hippocampus forms memories. Others are more abstract, exploring general principles of neural computation, such as predictive coding—the idea that the brain constantly generates predictions about sensory input and processes only the error between prediction and input—or attractor networks—the idea that persistent patterns of activity, such as those underlying working memory, correspond to stable states of a dynamical system.
Theoretical work has been influential in shaping how experimentalists interpret their data, but it has also faced challenges. Many models are underconstrained: they can fit the data but do not make sharp, testable predictions. And the gap between model and experiment is often large, with models simplifying away much of the biological complexity that experimentalists know to be important.
These approaches are not rivals in the sense of competing for the same explanatory territory. They operate at different scales and answer different questions. The single-neuron and population approaches are both concerned with how information is represented in neural activity, but they differ in whether the relevant unit of analysis is the individual cell or the population. The large-scale imaging approach is concerned with the organization of the whole brain, which is invisible to electrode recordings. The theoretical approach is not tied to any particular scale; it can be applied to single neurons, local circuits, or whole-brain networks.
The relationships are better described as complementary and increasingly integrated. Modern experiments often combine multiple techniques: recording from many neurons while simultaneously imaging the activity of large brain regions, or using optogenetics—a technique for controlling the activity of specific neurons with light—to test causal hypotheses suggested by correlational observations. The development of large-scale recording technologies, such as Neuropixels probes that can record from hundreds of neurons across multiple brain regions simultaneously, is blurring the boundary between the single-neuron and population approaches. And the growth of machine learning has provided new tools for analyzing high-dimensional neural data and new theoretical frameworks for understanding how complex computations might be implemented in neural circuits.
The current landscape of systems neuroscience is characterized by several converging trends. The first is the rapid growth of large-scale recording and imaging technologies. It is now possible to record from thousands of neurons simultaneously, to image the activity of entire brain regions in behaving animals, and to track the activity of the same neurons over days or weeks. This has shifted the field's focus from characterizing individual neurons to characterizing the structure of population activity.
The second trend is the increasing importance of causal manipulation. Recording tells you what neurons do, but not what they are for. Techniques such as optogenetics and chemogenetics allow researchers to turn specific populations of neurons on or off and observe the behavioral consequences. This has made it possible to test hypotheses about the causal role of specific circuits in behavior, moving beyond correlation to intervention.
The third trend is the growing integration of theory and experiment. The availability of large datasets has made it possible to test theoretical models more rigorously, and the development of new theoretical tools, particularly from machine learning, has provided new ways of thinking about what neural circuits might be computing. The relationship between artificial neural networks and biological neural networks has become a major topic of discussion, with researchers asking whether the principles that make artificial networks work—such as hierarchical feature extraction and predictive coding—are also at work in the brain.
The fourth trend is the expansion of the field's scope to include naturalistic behavior. Early systems neuroscience relied heavily on tightly controlled laboratory tasks, often involving anesthetized animals or highly constrained movements. There is now a push to study the brain in more natural conditions: freely moving animals, social interactions, complex natural stimuli. This has been enabled by new technologies for recording from behaving animals and by new computational methods for analyzing the resulting data.
The field's enduring challenge remains the gap between its levels of analysis. We have detailed knowledge of the biophysics of individual neurons, and we have increasingly rich descriptions of the activity of neural populations, but the link between these levels—how the properties of individual cells and synapses give rise to the emergent dynamics of populations, and how those dynamics produce behavior—remains incompletely understood. Systems neuroscience is, in large part, the ongoing attempt to bridge that gap.