In classical neuroscience, a neuron is often described as a simple integrator: it receives inputs, reaches a threshold, and fires. This "rate coding" view suggests that the primary way the brain communicates is through the frequency of spikes—a single, one-dimensional volume knob.
However, the brain faces a massive bandwidth problem. It must process high-dimensional sensory data, track internal states, and perform credit assignment, all using the same limited biological hardware. To solve this, neural circuits employ multiplexing: the ability to transmit multiple, independent streams of information through a single communication channel by utilizing different features of the neural signal.
The Two-Channel Spike Train: Rate and Timing
The most fundamental form of multiplexing occurs within the structure of the spike train itself. A single neuron can simultaneously encode fast-changing stimulus features through the precise timing of individual spikes while tracking slower environmental statistics through its overall firing rate.
Adrienne L. Fairhall and William Bialek demonstrated this in the fly visual system, showing that neurons adapt to the variance of a stimulus (the "context") by changing their firing rate, while the specific timing of spikes continues to represent the stimulus itself Multiple Timescales of Adaptation in a Neural Code. This allows the system to remain efficient even as the environment shifts from a dim forest to a bright field.
This dual-coding strategy is not limited to sensory adaptation. In the somatosensory cortex, Richard Naud and colleagues found that neurons decompose signals into an "event rate" (sensory data) and a "burst fraction" (top-down modulatory signals) Fast burst fraction transients convey information independent of the firing rate. By switching between regular spikes and high-frequency bursts, a single cell can transmit a sensory observation and an error signal simultaneously without the two interfering.
Temporal Partitioning: Sending Data in Windows
If you cannot send two signals at the exact same moment, you can send them in rapid succession. This is known as temporal multiplexing. The brain uses specific time windows to "tag" different types of information.
In the auditory cortex, research by Kerry M. M. Walker and Andrew J King shows that neurons encode timbre and pitch in distinct temporal windows Multiplexed and Robust Representations of Sound Features in Auditory Cortex. Timbre information is processed significantly faster than pitch, allowing the brain to identify what a sound is before fully resolving its musical tone.
This sequential encoding also appears in high-level cognition. Liang She discovered that "face patches" in the macaque brain initially encode the visual features of a face using a standard axis code, but then rotate this representation at longer latencies to encode memory and familiarity Temporal multiplexing of perception and memory codes in IT cortex. The same population of neurons serves as both a perceptual sensor and a memory retriever by simply shifting its representational logic over a few hundred milliseconds.
Oscillatory Phase Multiplexing
The brain is a rhythmic organ, and these rhythms (oscillations) provide a "clock" that can be used to partition information. By firing at different phases of an oscillation—such as the 4–8 Hz theta cycle—neural ensembles can pack multiple variables into a single second of activity.
Babak Shahbaba and colleagues found that hippocampal ensembles use theta cycles to multiplex the identity of a stimulus, its position in a sequence, and the animal's performance Hippocampal ensembles represent sequential relationships among an extended sequence of nonspatial events. Similarly, Louis Kang and Taro Toyoizumi proposed that the hippocampus oscillates between phases of high and low inhibition to alternate between processing fine-grained details and abstract concepts Distinguishing examples while building concepts in hippocampal and artificial networks.
Multiplexing as a Solution for Deep Learning
The biological success of multiplexing has inspired new architectures in artificial intelligence. A major hurdle in "biologically plausible" deep learning is the weight transport problem: how does a neuron know its contribution to a global error without a separate "backward" wiring?
Rui Ponte Costa and Will Greedy solved this by multiplexing the forward and backward passes Single-phase deep learning in cortico-cortical networks. In their BurstCCN model, single spikes represent forward inference, while high-frequency bursts carry the error signals needed for learning. This allows the network to learn in a single phase, mimicking the brain's ability to process data and update its internal model simultaneously.
Why Multiplexing Works
Multiplexing works because it exploits the high-dimensional "state space" of a spike train. A stream of pulses isn't just a number; it has a rate, a rhythm, a burstiness, and a relationship to background oscillations. By assigning different meanings to these different facets, the brain achieves a level of informational density that far exceeds what a simple rate-based system could provide. Whether it is a fly adapting to light or a human recognizing a familiar face, multiplexing allows the brain to do more with less.