The Physical Limits of Scalable Neural Recording

Adam H. Marblestone, Bradley M. Zamft and seventeen coauthors published Physical Principles for Scalable Neural Recording in Frontiers in Computational Neuroscience, volume 7, article 137, in October 2013 (DOI 10.3389/fncom.2013.00137). The paper asks a question that brain mapping projects usually leave implicit. Is there any recording technology, current or hypothetical, that can measure the activity of every neuron in a mammalian brain at millisecond resolution, and what does physics allow such a technology to look like.

The analysis is quantitative on every point. The mouse brain contains about 75 million neurons in a volume of about 420 cubic millimeters. Capturing one bit per neuron per millisecond requires a data rate of at least 75 billion bits per second, which the authors take as the minimal whole brain data rate. Four recording modalities are evaluated against that target, and each one falls orders of magnitude short in a physically specific way.

The physics behind the bounds

Three constraints run through the analysis. Thermal dissipation is capped near 40 milliwatts for a mouse brain at steady state, because tissue tolerates only about a 2 degree Celsius local increase. Volume displacement is capped near 1 percent, because inserted hardware disrupts vasculature and connectivity. Spatiotemporal throughput is capped by the spike itself, a roughly 2 millisecond event that must be resolved at kilohertz rates with clocks phase-locked across the whole brain. The paper also computes the data transmission problem for embedded devices and finds that radio-frequency transmission fails a power-bandwidth tradeoff, while infrared light or ultrasound could multiplex spatially.

The molecular modality is the paper’s most speculative branch and its most quoted one. A polymerase that records electrical activity into a synthesized DNA strand, the ticker tape idea, sits within the paper’s physical bounds on metabolic load and volume but requires time-stamping mechanisms to approach millisecond resolution. The authors’ stated purpose for deriving these bounds was design guidance, to tell the field which parameters require orders of magnitude improvement before any whole brain recording program becomes feasible.

What the bounds mean for emulation

Whole brain emulation has two distinct data problems. Structural mapping produces the wiring diagram. Activity recording produces the dynamic parameters that make the wiring run. The paper addresses the second problem, and its bounds are one reason the 2008 whole brain emulation roadmap treated simulation parameters as a bottleneck equal to structural acquisition, a structure examined in the roadmap review. Structural connectomics accelerated in the following decade, from the first complete worm diagrams to whole fly brains, but those instruments measure structure, and the millisecond activity record the paper targets requires separate instruments. The distinction between a wiring diagram and a running model is the theme of the MaleCNS comparison.

Comparison to The Consciousness AI

This project grounds its consciousness model in physical computation rather than in software abstraction. The paper’s bounds supply the quantitative floor for that position. Any system that claims to emulate a brain’s dynamics must first be able to observe those dynamics at the relevant resolution, and the paper shows that observing them is constrained by thermodynamics, scattering physics and data rates that no present instrument meets for mammalian brains. The project’s research code, maintained in the tlcdv/the_consciousness_ai repository, keeps its claims to what its own instruments can measure, and the paper’s scaling analysis is the reference class for that discipline. Consciousness as an emergent property of organized physical activity inherits these bounds directly, since whatever generates experience in neural tissue operates at spike resolution in tissue volume.

What the 2013 Bounds Got Right and Left Open

The paper’s bounds have held. Optical scattering still confines widefield access, electrical recording still covers hundreds to thousands of sites rather than millions, and no molecular recorder has been demonstrated in a behaving brain. What the paper left open is the interaction between the modalities, since a hybrid instrument could split the load across physical channels. The recording program it founded continues in embedded-device research and molecular recording proposals. For the emulation question, the paper’s conclusion stands as the quantitative form of a simple statement. A brain can be mapped structurally, mapped dynamically only within orders of magnitude of what its activity demands, and the gap is measured physics. The state of that measurement problem across the field is reviewed in the flagship overview of AI consciousness research, and emulation-specific coverage is collected on the brain emulation page.

Researchers covered here

• AM Adam MarblestoneHarvard University, Wyss InstituteFirst author of Physical Principles for Scalable Neural Recording, brain mapping scaling analysis 1 article
• EB Edward BoydenMassachusetts Institute of Technology, Howard Hughes Medical InstituteExpansion microscopy, optogenetics, connectomics tools for consciousness research 1 article