Sign-Constrained Architectures for Biologically Plausible AI

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2224 papers324 conceptsfrozen 2026-10-04

In modern deep learning, a single artificial neuron can exert a positive influence on one neighbor and a negative influence on another. While mathematically convenient, this violates a fundamental rule of neurobiology known as Dale’s Principle: a biological neuron typically releases the same neurotransmitters at all its synaptic connections, meaning it is either strictly excitatory or strictly inhibitory.

Recent research suggests that this "unnatural" flexibility in standard artificial neural networks (ANNs) creates a significant gap between AI and the biological brain. By enforcing Dale’s Principle through sign-constrained weights, researchers are discovering that biological realism does not have to come at the cost of performance. Instead, it provides a window into how the brain stabilizes its own chaotic dynamics and organizes complex information.

The Structural Gap

In a standard Recurrent Neural Network (RNN), the weights in the connectivity matrix can flip between positive and negative values during training. As Jonathan Cornford and Blake Aaron Richards point out in Learning to live with Dale's principle, this lack of sign-consistency prevents these models from accurately representing the structural constraints of biological brains.

When we ignore these constraints, we create models that align well with excitatory neural activity but fail to capture the dynamics of inhibitory populations. Sachi Sanghavi demonstrated in Distinct Roles of Putative Excitatory and Inhibitory Neurons that current AI models primarily optimize for categorization accuracy, leading to a bias that misses the functional diversity of the mammalian cortex.

Enforcing Dale’s Principle

To bridge this gap, researchers have developed methods to bake Dale’s Principle directly into the network architecture. The most common approach, described by Hui Song in Training Excitatory-Inhibitory Recurrent Neural Networks for Cognitive Tasks, involves decomposing the weight matrix into a product of a non-negative matrix and a fixed diagonal sign matrix. This ensures that every output connection from a specific unit maintains a consistent sign throughout training.

While naively enforcing these constraints often leads to unstable learning, new optimization techniques have solved this. For instance, Jonathan Cornford introduced "Dale’s ANNs" (DANNs), which use gradient-scaling based on the Fisher Information matrix to match the accuracy of standard deep learning models on major benchmarks. This proves that the biological requirement for segregated neurotransmitter types does not inherently limit the representational capacity of deep learning.

Taming Chaos through Balance

Why would the brain "limit" itself to single-sign neurons? The answer likely lies in the stability of dynamical systems. In Symmetries constrain the transition to heterogeneous chaos in balanced networks, Andrea K. Barreiro explains that sign-constrained connectivity introduces structural symmetries that regulate the transition to chaos.

By separating neurons into distinct excitatory (E) and inhibitory (I) populations, the network can maintain a "dynamically balanced" state. This balance acts as a stabilizer, preventing the runaway excitation that often plagues unconstrained recurrent networks. As Toshitake Asabuki argues, incorporating these distinct populations allows networks to learn complex tasks—like replaying high-definition movies—by aligning internal predictions with their own chaotic dynamics in a biologically plausible way.

Emergent Biological Motifs

When networks are forced to obey Dale’s Principle, they don't just perform well; they begin to look like real brains. Hui Song observed that enforcing these constraints leads to the emergence of specific connectivity motifs, such as "winner-take-all" inhibitory clusters, which are ubiquitous in biological circuits but absent in standard ANNs.

Furthermore, these constraints help explain the "quirks" of sensory processing. Friedemann Zenke found in Temporal prediction captures key differences between spiking excitatory and inhibitory V1 neurons that when a sign-constrained spiking network is trained to predict the next frame of a movie, the distinct firing rates and temporal dynamics of biological V1 neurons emerge naturally.

Toward a Unified Computational Lens

The shift toward E/I separated architectures suggests that biological constraints are not merely "artifacts of implementation," as Chi-Ning Chou discusses in The Computational Lens. Instead, they may be essential for:

  1. Energy Efficiency: Suppressing predictable information through inhibitory feedback, as seen in the Predictive Coding Light framework.
  2. Continual Learning: Managing interference between tasks. Fahad Sarfraz showed that "Bio-ANNs" with separate E/I populations can combat catastrophic forgetting by sharpening signals and managing interference.
  3. Robustness: Achieving faster task convergence and more efficient saliency prediction, as demonstrated by the midbrain-inspired models of Yash Sawant.

By treating Dale's Principle as a design feature rather than a limitation, researchers are building a new generation of AI that is not only more interpretable to neuroscientists but also more robust and efficient.

Papers in this version

Showing 60 of 224 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 193 papers
  2. Version 1 · 2026-10-04 · 224 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_dale_s_law_compliance_bf16f89d,
  title = {Dale's Law compliance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dale-s-law-compliance/rdr_bf16f89d15e1422cb7c27a510854fdd8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_sign_constrained_architectures_for_biolo_bf16f89d_v1,
  title = {Sign-Constrained Architectures for Biologically Plausible AI},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sign-constrained-architectures-for-biologically-plausible-ai/rdr_bf16f89d15e1422cb7c27a510854fdd8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}