Summary: The lessons learned from developing artificial neural networks offer a practical framework for studying the brain as a computational system rather than as a disconnected collection of individual cells. Applying insights from AI can accelerate our understanding of human cognition, behavior and neurodegenerative disease.
Source: The Conversation
Despite decades of research and large investments, how the human brain computes remains largely mysterious. At the same time, artificial neural networks—computational models that roughly mimic certain properties of biological neurons—have advanced rapidly. We have gained important insights into how complex computation can arise from network structure, learning objectives and update rules. These lessons can and should be applied back to biological brains to guide neuroscience toward a more explanatory, mechanistic understanding.
Neurological and neurodegenerative conditions are increasing in prevalence worldwide, making it urgent to decode how brain computation can go awry. Modern artificial neural networks solve challenging tasks by optimizing toward specific goals; translating that way of thinking to neuroscience can help identify the objectives and learning rules that shape brain circuitry across evolution and individual development.
Thoughts, perceptions and behaviours all arise from computations carried out by neural circuits. To develop more effective treatments for disorders that alter thought and behavior—such as schizophrenia or depression—we need models that explain how those computations are performed and how they can fail.
Recording activity from the brain often yields complex, high-dimensional signals that are difficult to interpret. Rather than focusing exclusively on assigning a single, fixed function to each recorded neuron, a computational perspective emphasizes the network architecture, the optimization objectives the system serves, and the rules by which connections and activity change over time. In a paper published in Nature Neuroscience, the authors argue for framing the brain in these terms: network structure, learning goals and update rules.
Brain network models
Artificial neural networks are engineered to capture key features of real neurons—integration of inputs, nonlinear activation and layered connectivity—while remaining flexible and tractable. Building an artificial network typically involves three steps: (1) specify the architecture, that is how units are connected and organized; (2) define the learning objective, such as predicting the next sensory input or maximizing reward; and (3) choose the optimization or learning rule that adjusts the network to meet its objective.
Importantly, designers do not handcraft the function of every single unit. Instead, units acquire specialized roles through the interaction of architecture, objective and learning dynamics. The brain’s development and lifetime learning may follow a similar pattern—evolution and experience shaping circuits toward particular computational goals without needing to predefine the role of each neuron.
Assigning neuron roles
This view challenges the traditional search for simple, unit-level descriptions of neurons. In many artificial networks, individual units do not admit compact verbal or mathematical descriptions; their functions emerge from the broader system. The same may be true in biological brains. To understand neural computation, we should prioritize characterizing the network’s wiring diagram, the objectives that evolution and behavior impose, and the plasticity mechanisms that update connections across developmental and evolutionary timescales.
Rather than attempting to label single cells with simple functions, researchers can achieve more explanatory power by describing how populations of cells and their connectivity implement computations relevant to behavioral goals. This shift reduces misleading intuition about isolated neuron “functions” and highlights how computation is distributed and emergent.
Optimizing frameworks
One clear success of an optimization perspective is the understanding of dopamine-related signals. Dopaminergic neurons in the brain appear to encode reward prediction errors—signals that report the difference between expected and received rewards. Reward prediction errors are a core training signal in many artificial reinforcement-learning systems. Framing dopamine in this objective-driven way links an abstract learning rule to observable neural activity and behavior, offering a unifying explanation for how reward-based learning is implemented biologically.
Similarly, when artificial networks are trained to treat game score or similar metrics as a reward signal, they learn to perform complex tasks by adjusting internal representations and actions to maximize that reward. Even when we cannot fully interpret each unit, understanding the network’s objective and learning dynamics reveals the roles of populations and pathways in service of that goal.
Progress in systems neuroscience therefore requires both bottom-up descriptive efforts—mapping connectivity, cell types and gene expression—and top-down theoretical work that proposes candidate objectives and learning rules. Using artificial neural networks as models, neuroscientists can generate hypotheses about what goals the brain might be optimizing and how its plasticity mechanisms implement those optimizations.
As machine learning continues to advance, a framework for neuroscience informed by these advances may illuminate how the brain computes, adapts and sometimes malfunctions. This approach promises to complement traditional experimental neuroscience and to provide actionable models for understanding and treating neurological disease.
Funding: Blake Richards receives funding from the Natural Sciences and Engineering Research Council of Canada, the Canadian Institute for Advanced Research, and Healthy Brains, Healthy Lives.
Source:
The Conversation
Media Contacts:
Blake Richards – The Conversation
Image Source:
The image is adapted from The Conversation news release.