The brain works at more than 10 bits per second

Summary

A recent study claims that human information processing operates with a “speed limit” of approximately 10 bits / s. Sauerbrei & Pruszynski argue that motor control processes, which guide movement in an unpredictable environment, significantly exceed this limit.

Standfirst

A recent article makes a claim with far-reaching implications for neuroscience, technology, and society: that the human brain is subject to an information processing “speed limit” of 10 bits per second. While this speed limit appears to hold for high-level cognitive functions, we argue that unconscious processing for real-time control of movement, which occupies a majority of neurons in the central nervous system and accounts for most of the information throughput of humans, significantly exceeds this limit.

The control of movement in complex, unpredictable, and often dangerous environments is a vital function of the nervous system. Consider an experienced athlete running across a rocky mountain ridge. Every 250 ms, she must swing a leg forward, ensuring her foot clears any protruding rocks, then place the foot on an appropriate support. This requires intense, active visual monitoring of the ground ahead. Because the head is subjected to random, broadband linear and angular motion, the nervous system must sense this motion and rapidly move the eyes to stabilize the scene on the retina. Meanwhile, the athlete’s nervous system must also use sensory signals conveying muscle length and force, touch, and head motion to precisely activate dozens of muscles across the trunk, neck, arms, and legs and maintain balance. Small errors in muscle activation could be catastrophic. Given the unpredictable nature of the environment, the dozens of reflex arcs operating in parallel with a sophisticated voluntary motor program, and the high precision demands in the task, how many bits of information would an observer need to describe the output of the athlete’s nervous system across a single stride?

A recent paper¹ makes an astonishing claim: although the brain collects a vast stream of sensory information, “[t]he information throughput of a human being is about 10 bits/s.” The authors define throughput as the information shared by the environment and a subject’s actions within a one-second window of time. The throughput, in their view, encompasses all motor actions, as well as perceptual and cognitive processing. To compare the flow of sensory input entering the nervous system with the throughput, Zheng and Meister first estimate the maximal information rate at the first stage of visual processing in the retina. By carefully analyzing the dynamic range, noise, and membrane potential kinetics of photoreceptors, they find that sensory information enters the nervous system at a rate exceeding 1 gigabit per second. This sensory information flows through - and is processed and filtered by - immense, distributed networks in the brain, resulting in the firing of motor neurons, the departure of signals from the central nervous system to activate muscle fibers, the development of forces and torques in the musculoskeletal system, motion of the limbs, trunk, and head, and, finally, successful performance of a task. Zheng and Meister next determine the information throughput of a human, not by measuring the uncertainty in the rich signals leaving the nervous system, but rather by calculating the degree of randomness in sequences of abstract symbols (such as letters or numerals) people produce in a range of tasks, including typing, memorizing binary digits, solving Rubik’s cubes, reading, and speech. They find that, despite the diversity of the tasks, the maximal information rate of human behaviour is consistently on the order of 10 bits per second, and conclude that this value is an upper bound on the information throughput of an entire human being. They then suggest the enormous ratio of sensory input to motor output, which they estimate at approximately 10⁸ and refer to as the brain’s “sifting number,” is a key unexplained quantity in neurobiology.

The central claim by Zheng and Meister would imply that the total output of our runner’s nervous system over an entire 250 ms stride could be described with approximately three binary bits. However, the phase, amplitude, and duty cycle for even a single muscle, which the nervous system must actively adjust from stride to stride in response to unpredictable sensory input, cannot be specified with only three bits. What is missing in Zheng and Meister’s argument? We suggest their methodology establishes a clear upper bound on the speed of cognitive processing, which has far-reaching implications for the capacity of declarative memory storage over a human lifetime, the design of brain-computer interfaces, and several other significant issues. Cognitive processing, however, is responsible for only a fraction of the information throughput of the nervous system, which is largely occupied with unconscious, real-time sensorimotor processing for controlling the body. The 10 bits / s speed limit, then, is a lower bound, not an upper bound, on the maximum information throughput of an entire human.

Zheng and Meister attempt to address this issue by dividing the nervous system into an “outer brain,” which is “closely connected to the external world through sensory inputs and motor outputs” and operates at extremely high bit rates, and an “inner brain,” which performs high-level functions related to cognition, goal-setting, and decision-making, and is limited to processing information at around 10 bits / s. This division is problematic for two reasons. First, the outer brain is part of the central nervous system, and the neuromotor outputs that result from its computations therefore contribute to the information throughput of a human. The vestibulo-ocular reflex, for example, maps sensory signals from the inner ear onto the activity of eye muscles to counteract head motion and stabilize the visual scene. This reflex, like many others, operates without the subject’s conscious awareness, and in parallel with any computations the inner brain performs. Put differently, much of the moment-to-moment computation the brain performs is not subject to a cognitive processing bottleneck. Second, while motoneurons and sensory receptors would clearly belong to the outer brain, it is less obvious where the inner brain begins. In the primate brain, many of the cortical neurons implicated in generating voluntary motor commands receive strong and relatively direct inputs from somatosensory receptors.

In fact, it appears likely that most of the neurons in the central nervous system contribute to sensorimotor processing for feedback control. The cerebellum alone contains about half the neurons in the brain, and is dedicated primarily to unconscious processing for balance, posture, and the accurate execution of voluntary movements. Correspondingly, most of the information throughput of the central nervous system likely derives from the computation of adjustments to muscle activity that are unconscious and often small, but indispensable for the execution of everyday tasks. In our view, the nervous system does not simply discard the high-bandwidth sensory information it collects. Nor does it passively transmit it to a receiving system, as an ethernet cable would. Rather, it uses this information to compute what task to perform, what movements are needed to perform the task, and what low-level motor commands are needed to produce the movements in an unpredictable environment. In summary, we propose that fully accounting for sensorimotor processing helps resolve the apparently massive and paradoxical imbalance between the information rates of sensory input and behaviour elaborated by Zheng and Meister.