The brain constantly processes a deluge of sensory information from multiple distinct streams—such as visual and auditory inputs, or feedforward sensory data and feedback internal states. To make sense of the world, neural circuits must identify and extract the shared underlying signals hidden across these diverse sources.
In data science, the standard tool for finding shared low-dimensional signals between two datasets is Canonical Correlation Analysis (CCA). Recently, a powerful research direction has emerged that bridges data analysis and neurobiology, proposing that individual biological neurons and local circuits are mathematically optimized to perform CCA and related source-separation tasks.
CCA as a Tool for Neural Discovery
Before modeling neurons as correlation analyzers, researchers used CCA as a statistical lens to decode population-level activity. Traditional neuroscience often focused on single-neuron "spike-counting" methods, like spike-triggered averages. However, Jakob H. Macke et al. demonstrated that by using CCA, researchers could simultaneously find a stimulus filter and a corresponding population response filter. This shifted the analytical focus from isolated neurons to the shared information channel between the environment and an entire neural population.
CCA also proved that seemingly noisy brain signals carry structured information. Evan Archer et al. applied CCA to recordings from the macaque visual cortex, revealing a tightly co-varying, low-dimensional signal shared between discrete spike trains and continuous local field potentials (LFPs). This discovery suggested that LFPs—often dismissed as background biophysical byproducts—actually reflect population-level information processing.
If CCA is such an effective tool for uncovering the latent structure of neural activity, an intriguing question arises: Could the brain's architecture be physically optimized to perform CCA natively?
The Normative Leap: Similarity Matching
To understand how a biological network might compute correlations, researchers use a normative framework. They start with a global mathematical objective function and mathematically derive a network structure and learning rules that can solve it.
A key breakthrough in this framework is the concept of similarity matching. The core intuition is simple: if two inputs are similar, their corresponding neural outputs should be similarly correlated. Cengiz Pehlevan et al. showed that complex tasks like Nonnegative Independent Component Analysis (NICA)—the problem of separating mixed, independent sensory sources—can be framed as a similarity matching problem.
By optimizing this objective, the authors derived a biologically plausible two-layer network. The first layer performs "prewhitening" (decorrelating and normalizing the inputs), while the second layer enforces nonnegativity and similarity matching to cleanly isolate the original sources. Crucially, because the objective relies on matching correlations, the resulting synaptic update rules are entirely local—meaning a synapse only needs to "know" the activity of its immediate pre- and post-synaptic neurons, matching how biological plastic changes occur.
Why Multi-Compartmental Neurons Emerge
While a two-layer network can separate signals, real brains feature highly complex individual cells, such as pyramidal neurons. These neurons are not simple point-like integrators; they possess distinct dendritic compartments (e.g., apical and basal dendrites) that receive different input streams.
Xinyuan Zhao et al. demonstrated that if you model a single neuron as an optimizer seeking to maximize the canonical correlation between two distinct input streams, a multi-compartmental structure is mathematically required. To compute the correlation between stream $A$ and stream $B$ without mixing them prematurely, the streams must be fed into separate dendritic compartments. The multi-channel CCA objective function ensures that the neuron can learn to align these streams using local synaptic updates, without needing to explicitly communicate internal dendritic states across the entire network.
Input Stream A ───► [ Dendritic Compartment A ] ──┐
▼
[ Soma / Output ]
▲
Input Stream B ───► [ Dendritic Compartment B ] ──┘
Building on this single-neuron model, David Lipshutz et al. extended the architecture to a single-layer neural network capable of multichannel CCA. By utilizing multi-compartmental neurons and local, non-Hebbian learning rules, this framework provides a concrete explanation for how biological networks integrate diverse sensory modalities in real time.
Unifying Biology and Optimization
Why does the mathematics of correlation maximization map so cleanly onto the strange anatomy of biological neurons?
David Lipshutz et al. provided a unified normative framework showing that a broad class of generalized eigenvalue problems—the mathematical family to which CCA belongs—naturally generates multi-compartmental structures and non-Hebbian plasticity when solved in an online, streaming fashion. The physical structure of the neuron is quite literally a reflection of the geometry of the optimization problem it is solving.
To make these models even more realistic, researchers have focused on resource efficiency. While early similarity matching models required two distinct layers of neurons to separate signals, David Lipshutz et al. derived single-layer network architectures that can perform NICA. These networks operate under strict biological constraints: they process data online, use purely local learning rules, and produce strictly non-negative outputs (representing realistic neural firing rates).
By viewing the neuron not as a passive filter, but as an active canonical correlation analyzer, this research direction provides a beautiful, principled bridge between the statistical demands of unsupervised learning and the physical realities of biological neural circuits.
Key Foundational Work
Canonical correlations reveal co-variability between spike trains and local field potentials in area MT) (2015) — contributed 1 concept elements to this direction.
Receptive Fields without Spike-Triggering (2007) — contributed 1 concept elements to this direction.