Scaling Laws for Neural and Physiological Signals

44 concepts39 papers1 proposal

For decades, decoding biological signals—whether the electrical oscillations of the brain (EEG, fMRI) or the micro-movements of muscles (sEMG)—felt like trying to listen to a whisper inside a crowded stadium. The data is notoriously noisy, sparse, non-stationary, and highly variable from one individual to the next. The traditional consensus in computational neuroscience and medicine was that biological data possessed an inherent "noise ceiling" that no algorithm could breach, forcing researchers to rely on highly specialized, hand-crafted features for tiny, isolated cohorts.

But a profound shift is occurring. By treating physiological and physical signals not as isolated streams of noise, but as high-dimensional continuous languages, researchers are discovering that these systems obey the same scaling laws that govern large language models. As compute, parameters, and data volume scale up, performance improves with startling, power-law predictability. The bottleneck wasn't the signal; it was the scale.

Overcoming Noise Through Model Capacity

Why does making a model bigger help it read messy muscle or brain signals? Intuitively, a larger parameter space allows a neural network to construct a more robust, invariant geometric manifold. Instead of overfitting to the specific sensor noise or individual quirks of a single recording session, a scaled-up architecture learns to project disparate, noisy inputs into a shared latent space where the underlying semantic meaning is preserved.

For instance, Nicholas Mehlman et al. demonstrated that increasing the parameter count of standard Transformer architectures leads to log-linear improvements in decoding accuracy for surface electromyography (sEMG). Crucially, these larger models exhibit superior generalization to entirely unseen users, proving that model capacity is a primary driver for overcoming biological variability.

This phenomenon extends directly to non-invasive brain-computer interfaces (BCIs). Niklas Smedemark-Margulies et al. argued that BCI performance is primarily limited by the capacity of the decoding model rather than the inherent quality of the electroencephalography (EEG) signal. When model capacity scales, communication speed and accuracy rise accordingly, allowing discriminative deep models to extract intent from messy signals far more effectively than traditional generative baselines.

The Power-Law of Data Volume and Diversity

Scaling parameters is only half the equation; the volume and diversity of training data act as the essential fuel that prevents large architectures from overfitting. In the past, neural datasets were confined to small, single-subject experiments. Today, foundation models are trained on massive, aggregated repositories.

Consider speech decoding from brain activity. Dulhan Jayalath et al. showed that performance in neural signal decoding scales log-linearly with the volume of unlabelled pre-training data, showing no immediate plateau. By scaling up to hundreds of hours of magnetoencephalography (MEG) data using self-supervised learning, non-invasive recordings began to match the accuracy of invasive surgical implants. Similarly, Xiao Ding et al. demonstrated that deep learning performance continues to improve monotonically as per-subject data scales to hundreds of minutes, confirming that high-volume, longitudinal data can systematically overcome signal noise.

To ingest this data, researchers are building generalist architectures that span thousands of individuals. Models like those introduced by Yassine El Ouahidi et al. (REVE) pretrain on tens of thousands of hours of EEG data across thousands of subjects, showing that downstream task accuracy scales predictably with both model size and data volume.

However, simply adding parameters is not always the optimal path. Danny Dongyeop Han et al. discovered that in electrophysiology foundation models, data diversity and training duration are far more critical to downstream performance than parameter count alone. This aligns with evolutionary perspectives: Naoki Hiratani et al. noted that the scaling of intermediate processing layers in biological brains is fundamentally constrained by the total volume of sensory experience an organism encounters over its lifespan, as increased volume mitigates the risk of overfitting.

Brain Alignment and the Limits of Scale

If scaling models allows us to decode the brain better, does it also mean these models are becoming more like the brain? This is the core question of neural encoding—predicting biological brain activity from external stimuli like text or images.

The evidence suggests a striking mathematical convergence. Richard Antonello et al. evaluated language models across multiple orders of magnitude and found that their ability to predict human fMRI responses scales logarithmically with parameter count, improving by roughly 4.4% for every order of magnitude increase in size. Laurent Bonnasse-Gahot similarly demonstrated that neural encoding accuracy scales linearly with the logarithm of a model's parameter count, recovering complex biological features like left lateralization in language processing. As a comprehensive survey by SUBBA REDDY OOTA et al. highlights, this predictable scaling indicates that artificial and biological networks may be converging on similar representational strategies as they solve complex tasks.

Yet, this alignment has boundaries. Abdulkadir Gokce et al. fitted power-law functions to misalignment scores between artificial networks and the primate visual ventral stream. They discovered a provocative decoupling: while scaling compute and data consistently improves an AI's behavioral accuracy, its internal neural alignment with biological brains eventually hits a theoretical ceiling where further scaling yields no biological convergence.

Beyond the Brain: Wearables and Physical Operators

The realization that continuous, noisy signals scale predictably has rapidly spread beyond neuroscience into physiology and scientific computing.

In healthcare, wearable sensors track messy, multi-modal streams of heart rates, motion, and temperature. Girish Narayanswamy et al. demonstrated that physiological sensor models follow identical power-law scaling relationships, allowing researchers to precisely forecast the compute and data required to hit specific clinical accuracy targets. By leveraging these scaling laws, Yuwei Zhang et al. scaled pretraining to 60 million hours of sensor data, enabling zero-shot activity recognition and natural language health descriptions.

Even the laws of physics are being mapped through this lens. Neural operators designed to solve partial differential equations (PDEs) exhibit the exact same scaling characteristics as language models. The authors of Unisolver showed that transformer-based PDE solvers improve predictably as training data and parameter counts increase. Similarly, Zhongkai Hao et al. and Emmanuel de Bézenac et al. demonstrated that large-scale pretraining on physical simulations yields predictable power-law improvements in downstream accuracy and zero-shot generalization.

Whether modeling a partial differential equation, a wearable sensor stream, or the firing of a billion cortical neurons, the underlying mechanism remains consistent. Scale provides the capacity to filter out stochastic noise and uncover the low-dimensional, deterministic manifolds governing complex systems.

Key Foundational Work

End-to-End Deep Image Reconstruction From Human Brain Activity (2019) — contributed 1 concept elements to this direction.
Developmental and evolutionary constraints on olfactory circuit selection (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_neural_scaling_laws_add715d4,
  title = {neural scaling laws},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-scaling-laws/rdr_add715d4f91c4061ac741ca07bcd3fec},
  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_neural_scaling_laws_add715d4_v1,
  title = {neural scaling laws},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-scaling-laws/rdr_add715d4f91c4061ac741ca07bcd3fec/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake