For decades, systems neuroscience operated primarily from the bottom up. Researchers meticulously mapped the anatomical wiring of the brain and hand-tuned mathematical models to fit the firing patterns of specific neurons. While this approach cataloged the "what" of neural circuits, it struggled to explain the "why"—the underlying computational pressures that force these specific structures to emerge.
A profound shift has occurred through the framework of neuroconnectionism, as formalized by Adrien Doerig et al. in The neuroconnectionist research programme. Instead of forcing artificial models to directly mimic neural data, researchers optimize generic artificial neural networks (ANNs) purely to solve high-level behavioral or ethological tasks. Remarkably, when these networks learn to solve complex environmental challenges, they spontaneously develop internal representations, geometries, and structural motifs that closely mirror biological brains.
This phenomenon—functional convergence—suggests that the functional architecture of the brain is not an arbitrary biological accident, but a mathematically optimized solution driven by the computational requirements of the tasks an organism must perform.
The Geometry of Task Constraints
Why should a network trained to classify images on a computer develop internal units that match a monkey's visual cortex? The answer lies in the geometry of the task itself.
To achieve invariant object recognition, a system must transform highly variable raw sensory inputs (pixels or retinal activations) into stable semantic categories. It must discard irrelevant latent variables like lighting, pose, and background while preserving identity. The physical world imposes strict mathematical constraints on how this transformation can occur.
[Raw Input Space] ---> (Hierarchical Invariance Processing) ---> [Stable Semantic Categories]
^ ^
| |
Driven by Environmental Statistics Enforced by Task Objective
In their landmark work, Daniel Yamins et al. demonstrated in Hierarchical Modular Optimization of Convolutional Networks Achieves Representations Similar to Macaque IT and Human Ventral Stream that optimizing deep convolutional networks solely for high-level object recognition performance yields internal features that naturally align with the primate ventral stream. As Charles F. Cadieu et al. further showed in Deep Neural Networks Rival the Representation of Primate IT Cortex for Core Visual Object Recognition, these task-optimized representations rival the actual neural firing patterns of the primate inferior temporal (IT) cortex.
As Daniel A. Butts notes in Data-Driven Approaches to Understanding Visual Neuron Activity, these goal-driven models often achieve higher predictive accuracy regarding biological brain activity than models engineered explicitly to fit neural data. Because both biological evolution and gradient descent optimize for the same functional goals under identical environmental statistics, they converge on the same high-dimensional representational geometry.
Universality Across Sensory and Motor Modalities
This convergence is not unique to vision; it represents a universal principle spanning diverse neural subsystems:
- Spatial Navigation: In the medial entorhinal cortex (MEC), traditional models focused heavily on highly symmetric "celebrity" neurons like grid cells. However, Aran Nayebi et al. demonstrated in Explaining heterogeneity in medial entorhinal cortex with task-driven neural networks that training recurrent neural networks (RNNs) on path integration tasks naturally reproduces the complex, heterogeneous response patterns of the entire biological MEC population, providing a functional rationale for the "messy" majority of neurons.
- Proprioception: By training thousands of neural networks on varied behavioral tasks, Alessandro Marin Vargas et al. found in Task-driven neural network models predict neural dynamics of proprioception a direct linear relationship between a model's task performance and its ability to predict real neural firing patterns of body position and velocity in primates.
- Audition and Speech: In Brains on Beats, Umut Güçlü et al. showed that networks optimized to extract semantic information from music replicate the hierarchical complexity gradient observed in the human auditory cortex. Similarly, SUBBA REDDY OOTA et al. established in Speech Taskonomy: Which Speech Tasks are the most Predictive of fMRI Brain Activity? that high-level language regions are best predicted by models optimized for automatic speech recognition, pinning down the specific computational task that mirrors human linguistic processing.
- Foraging and Decision Making: When deep reinforcement learning agents are trained to solve resource-management problems, Nathan J. Wispinski et al. showed in Adaptive patch foraging in deep reinforcement learning agents that their internal LSTM dynamics spontaneously mirror the neural activity patterns seen in foraging primates, converging on identical mathematical solutions for optimal timing.
The Emergence of Structural Wiring Rules
The pressure of task optimization does not stop at representational alignment; it also dictates the physical and structural wiring rules of neural circuits.
For instance, the primary visual cortex (V1) exhibits intricate lateral connectivity patterns. Rather than being an arbitrary biological artifact, Sebastian Klavinskis-Whiting et al. proved in Prediction of future input explains lateral connectivity in primary visual cortex that these precise wiring rules—such as direction selectivity and co-axial alignment—emerge naturally when a recurrent network is optimized for a predictive coding objective (anticipating the next frame of a natural movie).
[Task: Predict Next Video Frame]
│
▼ (Gradient Descent Optimization)
[Emergent Recurrent Wiring Rules] ───> Matches V1 Lateral Connectivity
Similarly, Christopher J. Cueva demonstrated in Emergence of functional and structural properties of the head direction system by optimization of recurrent neural networks that optimizing generic RNNs for angular velocity integration recovers the exact mechanistic ring-attractor structures found in biological head-direction systems. Even the fundamental division of labor between cell types follows this logic: Li Ji-An showed in Understanding the functional and structural differences across excitatory and inhibitory neurons that the functional asymmetries between excitatory and inhibitory neurons (such as high projection selectivity and sparse connectivity) emerge as an optimal solution to categorization tasks under physical constraints.
Boundaries of the Alignment Paradigm
While task optimization is a powerful lens for understanding neural organization, the alignment is bounded by specific constraints. As Pamela K. Douglas emphasizes in Cognitive computational neuroscience, the goal is to discover the intersection of task performance, biological plausibility, and architectural constraints.
Two major caveats prevent a naive equivalence between AI performance and biological truth:
- The Target Dependency: In No Free Lunch from Deep Learning in Neuroscience: A Case Study through Models of the Entorhinal-Hippocampal Circuit, Rylan Schaeffer et al. demonstrated that the emergence of specific periodic representations (like grid cells) is highly sensitive to the mathematical structure of the supervision target or readout function. If the target itself implicitly contains the structural bias, the network is merely inheriting the representation rather than discovering it autonomously.
- The Engineering Divergence: Maximizing raw task performance using modern machine learning techniques does not guarantee better biological models. Drew Linsley et al. showed in Better artificial intelligence does not mean better models of biology that as deep networks are scaled up to achieve superhuman accuracy on engineering benchmarks, their internal representations can actually diverge from primate visual processing strategies.
This divergence underscores that the brain is not optimized for raw task performance in a vacuum. Instead, biological neural representations are the result of optimizing for functional tasks under strict evolutionary, metabolic, and architectural constraints. The core power of the task-driven paradigm lies in identifying exactly which combinations of objectives and constraints force a generic computational system to arrive at the biological solution.