In traditional neuroscience, we often view the brain as a collection of individual "encoders." In this view, a single neuron in the motor cortex might represent a specific direction of arm movement. However, if you record from that same neuron tomorrow, its "tuning" might change, or the electrode might lose the signal entirely. This phenomenon, known as representational drift, presents a massive hurdle for Brain-Machine Interfaces (BMIs), which typically require tedious daily recalibration to account for these shifting signals.
Recent research suggests a solution to this instability: stop looking at individual neurons and start looking at the neural manifold.
The Geometry of Population Activity
While a recording might capture the activity of hundreds of neurons (a high-dimensional space), the actual patterns of activity are surprisingly constrained. Neurons do not fire in every possible combination; instead, their collective activity resides on a low-dimensional "surface" or manifold within that high-dimensional space.
Chethan Pandarinath and Eva L. Dyer describe this as a paradigm shift in Latent Factors and Dynamics in Motor Cortex and Their Application to Brain–Machine Interfaces. They argue that the "latent factors" governing these manifolds are the true drivers of movement. Crucially, while individual neurons may drift or disappear, the geometric structure of the population activity remains remarkably stable over months.
Latent Dynamics as an Invariant Anchor
If the manifold is the "stage" where neural activity happens, the latent dynamics are the "rules" of the play. These dynamics describe how the neural state evolves from one moment to the next.
Jonathan C. Kao and Krishna V. Shenoy demonstrated in Leveraging neural dynamics to extend functional lifetime of brain-machine interfaces that these dynamics are an invariant property of the motor cortex. By treating the rules of state evolution as fixed, they developed "decoder hysteresis"—a method that allows a BMI to remain functional even after losing 60% of its recording channels. The system uses its knowledge of how the population should behave to fill in the gaps left by missing neurons.
Aligning Shifting Manifolds
Even if the manifold's shape is stable, our view of it shifts because the specific neurons we record change from day to day. To maintain a stable BMI, we must "align" the manifold recorded today with the one recorded yesterday.
Tao Fang and Yu Qi proposed modeling these signals as "Semantic-Dynamic" (SD) features. In Extracting Semantic-Dynamic Features for Long-Term Stable Brain Computer Interface, they show that by anchoring decoders to these stable manifolds, systems can autonomously re-align themselves without needing a human to manually recalibrate the device every morning.
This alignment isn't just possible within one person over time; it may even work across different individuals. Ayesha Vermani and Il Memming Park have leveraged generative models to perform unsupervised alignment of neural time series data, suggesting that the underlying dynamical rules of a task are consistent enough to serve as a universal anchor across subjects.
Why Drift Might Be a Feature
If the brain is capable of such stable population-level representation, why do individual neurons drift at all? Paul Masset and Shanshan Qin argue in Drifting neuronal representations: Bug or feature? that drift might be a byproduct of a robust learning system.
Qin suggests that drift is a "coordinated random walk." While individual neurons change their tuning, they do so in a way that preserves the representational geometry—the relative distances between different stimulus representations. As long as the relationship between "moving left" and "moving right" stays the same in the manifold, the downstream areas of the brain (or a well-designed BMI) can still understand the message.
Robustness to Context and Noise
This manifold-based perspective also explains how we remain flexible. Matthew Mender found that when monkeys face new physical demands (like added resistance), their neural population structure shifts predictably rather than collapsing.
By focusing on these aggregate, low-dimensional structures, researchers like Collin Dunlap are building more resilient neuroprosthetics. These systems use features like mean wavelet power across the manifold to maintain control even when the fine-grained details of individual spikes are lost to noise or electrode degradation.
In summary, the "noise" of individual neurons is often just a different path across the same stable landscape. By decoding the landscape (the manifold) rather than the path, we can build interfaces that last a lifetime.