In neuroscience, we often find ourselves with two different maps of the same territory. On one side, we have elegant mathematical models—Bayesian inference, reinforcement learning, and deep neural networks—that predict human behavior with startling accuracy. On the other, we have the "wetware"—the intricate, messy reality of synapses, dendrites, and neurotransmitters.
The central challenge of this research direction is bridging the gap between these two maps. It is the search for process theories: the specific biological mechanisms that allow a collection of neurons to implement abstract algorithms.
The Gap Between Normative Theory and Physical Reality
Most computational models are "normative"—they describe what a system should do to be optimal. For example, Active Inference as a Computational Framework for Consciousness by Ryszard Auksztulewicz and Lucía Melloni argues that the brain minimizes "surprise." While mathematically sound, these frameworks often lack a "mechanistic sketch" of how biological wiring actually performs the calculation.
This tension appears across every cognitive domain:
- Perception: We know the brain performs Bayesian inference, but how? Aapo Hyvarinen and Patrik O. Hoyer suggest in Interpreting Neural Response Variability as Monte Carlo Sampling of the Posterior that cortical "noise" isn't an error, but the physical mechanism of sampling from a probability distribution.
- Navigation: Beren Millidge shows in Equivalence of Personalized PageRank and Successor Representations that memory retrieval and spatial navigation are mathematically the same algorithm. Yet, we still don't know how the hippocampus performs the specific matrix operations required to run that algorithm.
- Motor Control: When we fail at a task, the brain must perform "structural credit assignment" to find the error. Gregory Dam and Kunlin Wei demonstrate that humans are experts at this, but the underlying dopaminergic circuits remain a "black box" in their Credit Assignment during Movement Reinforcement Learning study.
Biologically Plausible Learning Rules
A major hurdle in this field is the "weight transport" problem. In artificial intelligence, we use backpropagation to train networks, but backpropagation requires neurons to know the exact weights of other neurons—a biological impossibility.
Researchers like Daniel Kunin are searching for Two Routes to Scalable Credit Assignment without Weight Symmetry. The goal is to find learning rules that are "locally available"—meaning a synapse only needs information from its immediate neighbors to learn—while still being powerful enough to solve complex tasks.
Similarly, Shivang Rawat and David Heeger have worked to prove the Unconditional stability of a recurrent neural circuit implementing divisive normalization. By grounding neural models in the physics of oscillators, they provide a blueprint for circuits that are both biologically realistic and mathematically stable.
Functional Alignment as a Clue
How do we know when a computational model has actually captured a biological truth? One approach is "functional alignment."
In Unsupervised deep learning identifies semantic disentanglement in single inferotemporal neurons, Irina Higgins, Doris Y. Tsao, and colleagues found that the latent units in a specific type of AI (a $\beta$-VAE) matched the firing patterns of real neurons in the primate visual system. This suggests that the brain might be using a similar "disentangling" algorithm to understand faces.
However, as James J. DiCarlo notes in A Conversation with James DiCarlo, functional alignment is only the first step. The ultimate goal is an "engineering-grade" model: a predictive map that tells us exactly how pixels on the retina are transformed into concepts in the mind through specific synaptic steps.
From Simple Organisms to Complex Insights
This search for implementation isn't limited to humans. By studying simpler systems, researchers can map the entire pipeline from sensory input to behavioral output:
- Insects: Benjamin Risse and colleagues reverse-engineered the insect polarised light compass, showing how a specific eye structure (the dorsal rim area) feeds into a neural architecture that maintains heading.
- Worms: In Nonlinear Control in the Nematode C. elegans, Megan Morrison and Charles Fieseler use dynamical systems theory to show how a single "control parameter" in the worm's brain can flip its behavior from crawling forward to reversing.
The Path Forward
The recurring theme across these 402 papers is that computation is not enough. A model that predicts behavior but ignores biological constraints is a map of a different country. To truly understand the mind, we must find the "neural implementation"—the specific, cell-type-level circuit diagrams that turn abstract math into biological action. As Christian Leibold suggests in his work on neural kernels, the future of the field lies in reconciling the high-level "what" of memory and learning with the low-level "how" of the synapse.