For decades, neuroscientists attempted to build models of the brain by painstakingly mimicking its anatomy—connecting "neurons" in ways that looked like biological circuits. However, a landmark shift occurred when researchers discovered that they could predict brain activity more accurately by ignoring anatomy and focusing instead on function.
This approach, known as goal-driven modeling, suggests that if you train an artificial neural network to solve the same ecological challenges as a biological organism—such as recognizing an object despite changes in position or lighting—the network will "accidentally" develop internal representations that mirror the brain.
The Functional Convergence Hypothesis
The core intuition behind this research is that there are only a few mathematically optimal ways to solve complex sensory tasks. If both a primate brain and a deep neural network (DNN) are optimized to achieve "invariant object recognition," they are likely to converge on the same representational geometry.
In 2014, Charles F. Cadieu, Daniel Yamins, and James J. DiCarlo demonstrated this in Deep Neural Networks Rival the Representation of Primate IT Cortex for Core Visual Object Recognition. They found that DNNs optimized solely for categorization performance on large-scale datasets (like ImageNet) developed internal features that were better predictors of neural firing patterns in the primate inferior temporal (IT) cortex than any model explicitly designed to mimic biology. This suggests that the brain’s visual hierarchy is not just a biological accident, but a computational necessity for solving real-world vision.
From Bottom-Up Anatomy to Top-Down Goals
Traditional "bottom-up" models start with the hardware (neurons and synapses) and try to build up to behavior. Goal-driven modeling flips this. As argued by Katherine R. Storrs and others, we can treat the "objective function"—the goal the system is trying to minimize—as the primary driver of neural organization.
This framework allows researchers to ask "why" a brain region looks the way it does. For example, Aran Nayebi and colleagues showed in Explaining heterogeneity in medial entorhinal cortex with task-driven neural networks that the "messy" and diverse firing patterns of neurons in the medial entorhinal cortex—often ignored in favor of neat "grid cells"—emerge naturally in recurrent networks trained on path integration. The complexity of the neurons is a functional requirement for the task of navigation.
Task Specificity Shapes Regional Identity
If the goal dictates the representation, then different tasks should predict different brain regions. This has been confirmed across multiple sensory modalities:
- Visual Specialization: Kshitij Dwivedi and Gemma Roig demonstrated that models trained on 3D scene parsing better explain navigation-related brain areas, while models trained on object recognition better explain the ventral stream (Task-specific vision models explain task-specific areas of visual cortex).
- Auditory Processing: In the auditory system, Greta Tuckute and Josh H. McDermott found that models optimized for speech recognition and environmental sound classification develop hierarchical stages that correspond to the human auditory pathway (Many but not all deep neural network audio models capture brain responses).
- Proprioception: Even our "sixth sense" of body position follows this rule. Alessandro Marin Vargas and Alexander Mathis showed that networks trained to predict limb movement from muscle signals spontaneously develop the same tuning properties found in the somatosensory cortex (Task-driven neural network models predict neural dynamics of proprioception).
Why It Works: The Normative Framework
The success of these models provides a normative explanation for neurobiology. It suggests that the primate brain is "optimal" in a specific sense: its internal wiring is a solution to the constraints imposed by the environment and the tasks necessary for survival.
When we see a model like ReAlnet-fMRI (Zitong Lu) or CORnet (Daniel Yamins) matching brain data, it isn't just because the model is "smart." It's because the model and the brain are solving the same problem under similar architectural constraints (like hierarchy and recurrence). As Kohitij Kar and James J. DiCarlo argue, these task-optimized networks serve as "digital twins" or "SMART models" that allow us to test hypotheses about the brain in a way that was previously impossible.