Representational Alignment of Artificial and Biological Neural Networks

This direction is split into other directions: linear representational mapping. Its last version stays readable and citable here.

Version 1 of 1301 papers309 conceptsfrozen 2026-10-04

For decades, neuroscience and artificial intelligence operated as parallel tracks: one attempting to reverse-engineer the biological brain, the other trying to engineer intelligent behavior from scratch. Recently, these tracks have converged through a technique known as neural representational alignment.

The core intuition is simple: if an artificial neural network (ANN) and a biological brain are both tasked with solving the same problem—such as recognizing an object or predicting the next word in a sentence—they may converge on similar internal representational strategies. By "mapping" the hidden layers of AI models to the firing patterns of biological neurons, researchers can treat AI architectures as explicit, testable hypotheses for how the brain processes information.

The Linear Mapping Hypothesis

The primary tool for measuring this convergence is the linear encoding model. To test if an ANN "thinks" like a brain, researchers present the same stimulus (like a photograph or a story) to both a computer model and a human subject (whose brain activity is recorded via fMRI, EEG, or ECoG).

They then attempt to find a simple linear transformation—usually a ridge regression—that can translate the high-dimensional activation vectors of the AI into the recorded biological signals. As Martin Schrimpf and Greta Tuckute demonstrate in The neural architecture of language, the success of this linear mapping is a rigorous metric. If a simple linear map works, it suggests the two systems share a fundamental geometric structure. If a complex, non-linear decoder were required, it would imply the "intelligence" resides in the decoder rather than the model itself Papale et al..

Hierarchical and Functional Correspondence

One of the most striking findings in this field is the spatial and temporal "mirroring" between AI layers and brain regions.

In computer vision, K. Seeliger and Umut Güçlü showed that the early layers of a CNN align with the primary visual cortex (V1) at roughly 75ms, while deeper layers align with higher-order object recognition areas at 200ms CNN-based Encoding and Decoding. Similarly, Chengxu Zhuang and James J. DiCarlo found that models trained via unsupervised contrastive learning—which mimics how infants learn without labels—develop a ventral stream hierarchy that matches primate neural data Unsupervised Neural Network Models of the Ventral Visual Stream.

In language, the alignment is even more specific. Badr AlKhamissi and colleagues identified a sparse "language network" within Large Language Models (LLMs) that is uniquely responsible for linguistic processing and maps directly to the human brain's own language regions The LLM Language Network.

Why Alignment Happens: Predictive Processing

Why do these silicon-based models align so well with carbon-based brains? The evidence suggests it is the objective function—the goal the model is trained to achieve.

Schrimpf et al. found that models optimized for next-word prediction (like GPT-2) are significantly more brain-aligned than those trained on other tasks. This supports the "predictive coding" theory in neuroscience: the idea that the human brain is essentially a prediction engine, constantly anticipating the next sensory input to minimize surprise. When we add feedback loops to AI to mimic these predictive dynamics, their similarity to human brain activity increases further Guo et al..

The Performance Paradox

Intuitively, one might assume that as AI becomes more accurate at a task, it becomes more "brain-like." However, recent work has uncovered a "performance paradox."

Drew Linsley and Thomas Fel observed that as models reach "superhuman" performance on benchmarks like ImageNet, they actually begin to diverge from biological neural patterns Performance-optimized deep neural networks are evolving into worse models. Similarly, Yuko Nakagi identified a "detachment phase" during LLM training where models temporarily move away from brain-like representations to master complex logic Triple Phase Transitions.

This suggests that while the brain is a highly efficient solution for survival, it is not the only solution for high-performance computation.

Brain-Tuning: Closing the Loop

If alignment is a measure of biological plausibility, can we use it to build better AI? Researchers are now exploring brain-tuning, where models are fine-tuned directly on neural data.

Dietrich Klakow and Mariya Toneva demonstrated that forcing a speech model to predict fMRI signals from human listeners improves the model's semantic understanding on downstream tasks Improving semantic understanding in speech language models via brain-tuning. By injecting biological inductive biases into artificial systems, we move toward AI that not only performs like a human but "understands" the world through a similar representational lens.

Papers in this version

Showing 60 of 301 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_representational_alignment_of_artificial_405d55cf,
  title = {Representational Alignment of Artificial and Biological Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/representational-alignment-of-artificial-and-biological-neural-networks/rdr_405d55cf8fcb417198c32147a92236f5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_representational_alignment_of_artificial_405d55cf_v1,
  title = {Representational Alignment of Artificial and Biological Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/representational-alignment-of-artificial-and-biological-neural-networks/rdr_405d55cf8fcb417198c32147a92236f5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}