Several theories propose that single cortical neurons compute canonical correlation analysis (CCA), a classical statistical method that finds the directions in which two data streams vary together most strongly. In these models, a pyramidal neuron receives one stream on its basal dendrites and another on its apical dendrites, and local synaptic rules slowly rotate its weights toward the shared signal. The theory is clean. The evidence that these circuits work on realistic data is thin. The models have been run on Gaussian toys, MNIST cut into pieces, and one video-text dataset, and they are judged by convergence plots rather than by how much correlation they actually capture. The most general version, the unified generalized-eigenvalue framework, has never been run on any dataset.
We propose a controlled test of one specific claim. These local, online circuits can recover the offline CCA solution on real sensory data, but only within a sample and tuning budget set by how ill-conditioned the inputs are, meaning how unevenly their variance is spread across directions. Real inputs such as speech spectra, articulator traces, and neighbouring pixels are highly correlated within each stream, so we expect conditioning, rather than the dendritic architecture, to explain why these networks look fine on toys and may fall short on real data. The project measures this dependence, isolates it with a prewhitening ablation, and reports the networks against the offline linear CCA ceiling they are mathematically trying to reach.
Core Claim
The biologically plausible CCA networks of Neurons as Canonical Correlation Analyzers, A Biologically Plausible Neural Network for Multichannel Canonical Correlation Analysis, and Normative Framework for Deriving Neural Networks with Multicompartmental Neurons and Non-Hebbian Plasticity all solve a saddle-point (minimax) problem. In this problem, the interneuron or lateral variables effectively whiten each input stream while the pyramidal weights search for correlated directions. The normative framework paper itself reports that its minimax objective is nonconvex-concave and diverges unless the learning-rate ratio $\tau$ is small.
Our hypothesis is that the whitening sub-problem sets the pace. Under a step size small enough to keep the fastest direction stable, the slowest direction relaxes at a rate proportional to the smallest covariance eigenvalue. Steps to convergence should therefore grow roughly linearly with the condition number $\kappa = \lambda_{\max}/\lambda_{\min}$ of the within-stream covariances, the ratio of the largest to the smallest variance direction. This is a hypothesis, not a theorem; the experiments measure the exponent.
If the hypothesis holds, three things follow. First, the networks' apparent success on synthetic data reflects benign conditioning. Second, feeding them prewhitened inputs should close most of their gap to offline CCA. Third, the deficit of single-layer circuits reported in Biologically plausible single-layer networks for nonnegative independent component analysis has a common explanation: those circuits must whiten and rotate in one layer, and they are outperformed by a two-layer network that whitens first.
The motivation comes from the empirical-validation literature. Algorithms with clean convergence rates often hide condition-number dependence or need parameters that cannot be known in advance. Linear Convergence with Condition Number Independent Access of Full Gradients is a direct example: it removes $\kappa$ from full-gradient counts but pays $\kappa^2$ in stochastic steps and needs $\kappa$ to set its schedule. We bring that kind of finite-sample scrutiny to normative neural circuits.
Research Program
We will implement four local online circuits in a common codebase with identical data streaming. The first is two-compartment online CCA from Neurons as Canonical Correlation Analyzers. The second and third are Bio-CCA and Adaptive Bio-CCA with output whitening from A Biologically Plausible Neural Network for Multichannel Canonical Correlation Analysis. The fourth is the CCA instance of the general eigenvalue-problem algorithm from Normative Framework for Deriving Neural Networks with Multicompartmental Neurons and Non-Hebbian Plasticity. Running the fourth is itself new, since that paper reports no datasets.
Every circuit is scored against exact regularized linear CCA computed offline on the same training split. The primary metric is the fraction of the offline held-out total correlation (summed over the top $k$ components) that the circuit captures. We also record samples-to-90%-of-ceiling, wall-clock time, and subspace angle to the offline solution.
To expose hyperparameter fragility, we fix a tuning budget for each circuit (the same number of learning-rate configurations drawn from the same grid). We then report both the best result and the fraction of configurations that converge rather than oscillate or diverge. Fragility under a fixed budget is a practical failure mode; it is not a nuisance to be tuned away.
The conditioning axis is controlled in two ways. On the Bach & Jordan latent-factor generator used by Bio-CCA, we hold the canonical correlations fixed and sweep within-stream $\kappa$ over several orders of magnitude. On real data, we measure the empirical $\kappa$ of each stream and add a variant with temporal context stacking, which raises $\kappa$ without changing the shared signal.
A secondary axis tests a biological objection the Bio-CCA authors raise themselves: their derivation assumes a 1:1 ratio of interneurons to pyramidal cells, while cortex has roughly four pyramidal cells per interneuron. We shrink the interneuron population and measure the loss in stability and captured correlation.
Baselines
The naive reference point is a random projection of matching rank, together with the raw correlation of the population mean of each stream. The second figure echoes the comparison in Canonical correlations reveal co-variability between spike trains and local field potentials in area MT), where mean spike rate and mean LFP correlated at only −12% while CCA found components at 29%, 26%, and 21%. The ratio between the naive score and the ceiling tells us how much there is to capture.
The ceiling for any linear circuit is offline regularized linear CCA. For Split-MNIST, Deep Canonical Correlation Analysis reports a test total correlation of 28.0 for linear CCA, 33.0 for kernel CCA, and 39.7 for deep CCA. We reproduce the linear figure under their protocol, and a circuit "succeeds" on this dataset if it reaches at least 90% of the reproduced linear value within one pass of streaming data. For XRMB, the same paper reports 88.6 total correlation in 112 dimensions for an eight-layer deep CCA. We found no reported linear CCA number for XRMB on Lacuna, so we compute it ourselves. Deep CCA serves as an upper reference showing what nonlinearity adds; it is not a competitor, because these linear circuits cannot exceed linear CCA. For Mediamill and AV ChiME3, no reported linear-CCA correlation was found on Lacuna, and we compute the ceiling.
The strongest existing online competitors are the non-biological streaming solvers Gen-Oja, MSG-CCA, and Asym-NN. Bio-CCA reportedly beats them in sample and runtime efficiency on synthetic data, but no reported number for that comparison was found on Lacuna. All solvers get the same data order, the same minibatch size of one, the same tuning budget, and the same compute. Beating them means reaching the 90% ceiling in fewer samples at matched convergence rate across the tuning grid, on real data and not only on synthetic data.
The ablation that isolates our key ingredient feeds each circuit inputs that have been whitened offline using training-set covariances, which sets $\kappa = 1$. If conditioning drives the gap, this ablation should recover most of the ceiling and collapse the $\kappa$-dependence. If the gap persists, conditioning is not the explanation, and the claim fails.
Why not just…?
Why not just report offline linear CCA and skip the circuits? For engineering, offline CCA is the right answer, and we treat it as the ceiling. The scientific claim under test, however, is that neurons can reach this solution with local rules from a stream. Only measuring circuits against the ceiling can test that, and the measurement is exactly what the existing papers omit.
Why not just compare to deep CCA on XRMB, as the obvious gap suggests? A linear circuit losing to an eight-layer network says nothing about whether the circuit works. Deep CCA's 88.6 mixes the cost of locality with the cost of linearity. Our ceiling-relative metric separates the two, and deep CCA appears only as a reference for how much the linearity assumption costs.
Why not just tune the learning rates harder? This is a real threat. With unlimited grid search, the circuits may converge everywhere, and the two-layer network in Blind Nonnegative Source Separation Using Biological Neural Networks already relied on extensive grid search. We therefore report the full curve of best result against tuning budget. If the $\kappa$-dependence disappears at modest budgets, we will say so, and the conclusion becomes "fragile but fixable" rather than "conditioning-limited."
Why not just use the existing online ML solvers? They use non-local updates, and they need a fixed rank where Adaptive Bio-CCA adjusts its rank. They are, however, our strongest competitor on efficiency. If they also slow down linearly in $\kappa$, then the slowdown is a property of streaming CCA in general, not of biological circuits, and we would report it that way.
Decisive Checks
The first check is the slope of log samples-to-convergence against log $\kappa$ on the controlled synthetic sweep. A slope near one, flattened by the prewhitening ablation, supports the claim. A flat slope without prewhitening refutes it.
The second check is whether the measured $\kappa$ of real streams (Split-MNIST halves, XRMB acoustics and articulators, Mediamill, AV ChiME3 audio and lip features) predicts each circuit's captured fraction of the ceiling across datasets.
The third check runs the same prewhitening ablation on the single-layer and two-layer nonnegative ICA circuits from Biologically plausible single-layer networks for nonnegative independent component analysis. If prewhitening erases the single-layer deficit, the conditioning explanation extends beyond CCA.
All code, tuning grids, and full convergence-rate tables will be released, so that the empirical record these models lack becomes reusable.
References
- Neurons as Canonical Correlation Analyzers
- A Biologically Plausible Neural Network for Multichannel Canonical Correlation Analysis
- Normative Framework for Deriving Neural Networks with Multicompartmental Neurons and Non-Hebbian Plasticity
- Biologically plausible single-layer networks for nonnegative independent component analysis
- Linear Convergence with Condition Number Independent Access of Full Gradients
- Canonical correlations reveal co-variability between spike trains and local field potentials in area MT)
- Deep Canonical Correlation Analysis
- Blind Nonnegative Source Separation Using Biological Neural Networks