BCNE: Brain-Dynamic Convolutional Embedding

• BCNE is an unsupervised deep manifold learning framework that maps multichannel brain data into low-dimensional trajectories while preserving temporal and spatial correlations.
• It constructs correlation-aware 2D images from reweighted time-series data using temporospatial precedent and Gromov–Wasserstein alignment before applying a 2D convolutional network.
• Experimental results across fMRI, hippocampal spikes, and sensorimotor recordings show BCNE effectively reveals brain state transitions, learning stages, and active–passive dissociations.

Searching arXiv for BCNE and related context papers. Search query: "Brain-dynamic Convolutional-Network-based Embedding BCNE arXiv" Brain-dynamic Convolutional-Network-based Embedding (BCNE) is an unsupervised deep manifold learning framework for dynamic brain data that maps multichannel recordings $X \in \mathbb{R}^{N \times T}$ to low-dimensional trajectories $\{y_t\}_{t=1}^T$, with $y_t \in \mathbb{R}^d$, intended to reflect evolving cognitive or behavioral states over time. Its defining strategy is to avoid applying dimensionality reduction directly to raw time-point vectors. Instead, BCNE first constructs a temporospatial correlative representation that encodes temporal autocorrelation and inter-channel correlation, converts each time point into a correlation-aware 2D image, and then learns a low-dimensional embedding through a 2D CNN and a recursive Kullback–Leibler manifold objective. In the reported experiments, this framework was used to reveal scene transitions in naturalistic fMRI, learning-stage organization in hippocampal population activity, and active–passive dissociations in sensorimotor electrophysiology (Zhou et al., 7 Aug 2025).

1. Definition and analytical scope

BCNE is designed for temporally extended, non-stationary neural processes such as watching a movie, learning a spatial task, or performing active and passive movements. The fundamental objective is to represent each time point $t$ by a low-dimensional vector $y_t$ such that the ordered sequence $\{y_t\}$ forms a visually interpretable and quantitatively useful “brain-state trajectory.” The target is not merely low reconstruction error or classification performance, but an embedding geometry that preserves meaningful local neighborhood structure while also exposing larger-scale temporal organization (Zhou et al., 7 Aug 2025).

The framework is motivated by specific limitations of conventional dimensionality-reduction pipelines. Static, snapshot-based methods such as PCA, t-SNE, UMAP, and PHATE treat each time point independently and therefore ignore temporal continuity; in the reported characterization, this often yields fragmented or noisy embeddings in which adjacent time points can scatter arbitrarily. Static connectivity analyses summarize over long windows or full runs and therefore lose fine-grained temporal structure. Direct manifold learning on raw time-series vectors $x_t \in \mathbb{R}^N$ also ignores two structures that BCNE treats as central: temporal correlations at different lags within each channel and correlative structure across channels (Zhou et al., 7 Aug 2025).

A common misconception is to equate BCNE with a generic dimensionality-reduction method applied to neural recordings. In the implementation described, BCNE is more specific: it is a temporospatially aware pipeline in which temporal dependence is handled before the convolutional network and spatial organization is induced through a correlation-based channel layout. The learned embedding is therefore downstream of a deliberately structured representation rather than a direct projection of raw $N$-dimensional observations (Zhou et al., 7 Aug 2025).

2. Temporospatial correlative representation

The first stage of BCNE constructs a representation that combines temporal autocorrelation and spatial correlation. Given $X \in \mathbb{R}^{N \times T}$, BCNE computes, for each channel $i$, an autocorrelation function

$\{y_t\}_{t=1}^T$0

then smooths each autocorrelation curve using a centered moving average,

$\{y_t\}_{t=1}^T$1

A dropoff point $\{y_t\}_{t=1}^T$2 is defined as the first lag where the smoothed autocorrelation becomes negative. This determines a temporal correlation matrix $\{y_t\}_{t=1}^T$3: $\{y_t\}_{t=1}^T$4 The original data are then reweighted as

$\{y_t\}_{t=1}^T$5

so that each $\{y_t\}_{t=1}^T$6 becomes a lag-weighted average of temporally correlated neighboring time points rather than a simple local average (Zhou et al., 7 Aug 2025).

BCNE then imposes a spatial organization on channels. From the temporally reweighted data $\{y_t\}_{t=1}^T$7, it computes a sample covariance matrix $\{y_t\}_{t=1}^T$8, its inverse $\{y_t\}_{t=1}^T$9, and a channel–channel interaction matrix

$y_t \in \mathbb{R}^d$0

This $y_t \in \mathbb{R}^d$1 is paired with a $y_t \in \mathbb{R}^d$2 grid, where $y_t \in \mathbb{R}^d$3, whose geometric distance matrix $y_t \in \mathbb{R}^d$4 is

$y_t \in \mathbb{R}^d$5

BCNE aligns the geometry of $y_t \in \mathbb{R}^d$6 and $y_t \in \mathbb{R}^d$7 by a Gromov–Wasserstein coupling

$y_t \in \mathbb{R}^d$8

with uniform marginals and $y_t \in \mathbb{R}^d$9, solved by an entropically regularized Sinkhorn algorithm. A hard assignment derived from $t$0 maps channels to grid positions, and each time point is converted into an image

$t$1

The result is a sequence of 2D images in which temporally smoothed channel values are arranged so that correlated channels are spatially proximate, allowing 2D convolutions to act on a correlation-aware layout rather than an arbitrary channel ordering (Zhou et al., 7 Aug 2025).

3. Convolutional mapping and recursive manifold objective

The direct input to BCNE is the image sequence $t$2, with $t$3. The default network is a 2D CNN followed by a dense stack. The convolutional module contains four convolutional layers with 16, 32, 64, and 64 feature maps. Their outputs are flattened and passed through fully connected layers of dimensions

$t$4

where $t$5 or $t$6. The final dense layer output is the embedding $t$7. The implementation described uses no temporal convolution; each time point’s image is processed independently, and temporal structure has already been injected through the reweighted input representation (Zhou et al., 7 Aug 2025).

BCNE learns the embedding by matching pairwise similarity distributions in high- and low-dimensional spaces. For inputs $t$8 and $t$9, the high-dimensional similarities are defined by a Gaussian kernel,

$y_t$0

In the embedding space, similarities are defined by a Student-$y_t$1 kernel,

$y_t$2

with $y_t$3. The stagewise loss is

$y_t$4

This is a parametric manifold-learning objective rather than a reconstruction loss, a predictive loss, or a contrastive objective (Zhou et al., 7 Aug 2025).

A central feature is recursive manifold refinement. Let $y_t$5 denote the output of dense layer $y_t$6. At recursion stage $y_t$7, BCNE computes a new similarity distribution $y_t$8 on $y_t$9 and minimizes

$\{y_t\}$0

Layers up to stage $\{y_t\}$1 remain trainable. Reported recursion depths are typically $\{y_t\}$2 or $\{y_t\}$3: earlier stages preserve local neighborhoods, whereas later stages integrate increasingly global structure. Training is fully unsupervised and batch-wise. The paper states that labels are used only after training for evaluation and interpretation, and it identifies a “balance” hyperparameter, defined as dataset size divided by batch size, with values around $\{y_t\}$4–$\{y_t\}$5 performing best (Zhou et al., 7 Aug 2025).

4. Experimental domains and reported empirical patterns

The reported evaluations span three modalities and three distinct neurobehavioral settings, each chosen to test whether a low-dimensional trajectory can expose structure that is difficult to recover from static or time-agnostic dimensionality reduction (Zhou et al., 7 Aug 2025).

Dataset	Modality and sample	Reported BCNE structure
Sherlock	fMRI, 16 participants	Scene transitions, event boundaries, ROI-specific trajectories
Rat hippocampus	CA1 spike trains, 4 rats	Learning stages, spatial position, movement direction
Macaque S1 area 2	Electrophysiology, 1 rhesus macaque	Active–passive separation, reach-angle trajectories, limb-position geometry

In the Sherlock dataset, participants viewed a $\{y_t\}$6-minute clip of Sherlock, and BCNE was applied to four ROIs: early visual (307 voxels), high-level visual (571 voxels), early auditory (1018 voxels), and posterior medial cortex (481 voxels). The reported trajectories for high-level visual and early auditory cortex formed cohesive, continuous shapes with clear temporal progression and scene segmentation, whereas early visual and posterior medial cortex were more dispersed. Recursion sharpened scene boundaries, and by Recur 3 the embeddings exhibited more refined event structure than at Recur 0. In 2D KNN classification of 39 scenes, BCNE achieved the highest accuracy across all four ROIs, with an average improvement over T-PHATE of approximately $\{y_t\}$7. For behaviorally rated event boundaries, BCNE yielded larger $\{y_t\}$8 than PCA, t-SNE, UMAP, PHATE, T-PHATE, and CEBRA in early visual, high-level visual, and early auditory cortex, and was only slightly lower than CEBRA in posterior medial cortex while showing less variance (Zhou et al., 7 Aug 2025).

In the rat hippocampus experiments, four male Long-Evans rats traversed a 1.6 m linear track, with 48–120 putative pyramidal neurons recorded per rat and spike trains binned into 25 ms bins. BCNE produced trajectories that evolved smoothly along the track and separated early and late learning stages; when colored by movement direction, leftward and rightward traversals formed distinct but related manifolds. The paper reports that, from Recur 0 to Recur 3, KNN learning-stage accuracy approximately doubled from $\{y_t\}$9 to $x_t \in \mathbb{R}^N$0. In 2D, BCNE achieved the highest learning-stage accuracy and the highest direction accuracy; in 3D, T-PHATE was slightly better for direction. BCNE also achieved the highest trustworthiness across methods, while continuity remained high, and representational similarity analysis (RSA) with spatial position was consistently strong, only slightly below CEBRA in some 3D conditions. Roll-shift tests showed that embedding-position correspondence degraded sharply with large temporal misalignment, with $x_t \in \mathbb{R}^N$1-values remaining near zero (Zhou et al., 7 Aug 2025).

In the macaque S1 area 2 dataset, one rhesus macaque performed an eight-direction center-out reaching task under active and passive conditions. The BCNE embeddings were reported as the only ones that clearly separated active and passive trajectories in both 2D and 3D while retaining angular structure. Active trajectories appeared more linear and direct, whereas passive trajectories appeared more curved or circular. In active mode, RSA between embedding-space geometry and actual limb-position geometry averaged approximately $x_t \in \mathbb{R}^N$2 in 2D and $x_t \in \mathbb{R}^N$3 in 3D; T-PHATE was second best at approximately $x_t \in \mathbb{R}^N$4 and $x_t \in \mathbb{R}^N$5, respectively. For eight-angle KNN classification, where chance was $x_t \in \mathbb{R}^N$6, t-SNE and BCNE achieved almost $x_t \in \mathbb{R}^N$7 accuracy, with UMAP and T-PHATE above $x_t \in \mathbb{R}^N$8 (Zhou et al., 7 Aug 2025).

5. Position within dynamic embedding research

Dynamic network embedding, in the formulation of the survey “A Survey on Dynamic Network Embedding” (Xie et al., 2020), studies a time-evolving graph

$x_t \in \mathbb{R}^N$9

and seeks snapshot-specific low-dimensional node representations that preserve structural properties and temporal dynamics. That survey proposes five categories of methods: matrix factorization–based, Skip-Gram–based, autoencoder-based, neural-network-based, and other methods. It also emphasizes recurring mechanisms such as reconstruction objectives, temporal smoothness, predictive losses, attention, and incremental updates (Xie et al., 2020).

BCNE is not part of that survey’s original corpus, but the taxonomy suggests a natural placement. It aligns most closely with the neural-network-based dynamic embedding family, because it combines deep convolutional processing with temporally informed representation learning. At the same time, its implementation differs from canonical dynamic graph models. It does not operate on explicit graph snapshots $N$0, does not use graph message passing, and does not model temporal dependence with an RNN or temporal self-attention. Instead, temporal structure is injected by autocorrelation-based reweighting, and spatial organization is created by a correlation-preserving image layout before standard 2D convolutions are applied. This suggests that BCNE should be understood as a brain-dynamic embedding method adjacent to spatio-temporal neural embedding, rather than as a direct instance of graph convolutional dynamic embedding in the sense of EvolveGCN, DySAT, or DyRep (Xie et al., 2020).

This contextualization also clarifies what the term “convolutional-network-based” denotes in BCNE. In the reported architecture, convolution acts on a 2D image induced by temporospatial correlation and Gromov–Wasserstein alignment, not on an explicit brain graph. A plausible implication is that BCNE occupies an intermediate position between dynamic network embedding and deep parametric manifold learning: it inherits the dynamic-embedding concern with temporality and inter-entity structure, but operationalizes those concerns through image construction and KL-based manifold preservation rather than through evolving graph filters or event-driven node updates.

6. Interpretation, limitations, and prospective extensions

BCNE is explicitly presented as an unsupervised exploratory tool. Scene labels, learning stages, movement direction, active–passive condition, and limb kinematics are not used during training; they are used only after training for interpretation, KNN evaluation, RSA, and related analyses. Another common misconception is that BCNE is an autoencoder or predictive model. In the reported formulation, there is no explicit reconstruction loss, no predictive loss, and no contrastive supervision. The core optimization target is recursive minimization of $N$1 between pairwise similarity distributions (Zhou et al., 7 Aug 2025).

Interpretability arises primarily from the geometry of the learned trajectories and from post hoc alignment with external variables. The paper emphasizes coloring trajectories by scenes, learning stages, directions, position, or active–passive condition, and it uses RSA to compare distances in embedding space with distances in physical or behavioral spaces. It also notes a conceptual form of back-projection: because the correlation-based image layout preserves channel identities, one can relate regions of the manifold back to ROIs or neural populations. However, the reported implementation does not include explicit saliency maps (Zhou et al., 7 Aug 2025).

The stated limitations are substantive. Temporal correlation is truncated at the first zero crossing of the smoothed autocorrelation, so long-range correlations and repeated motifs beyond that dropoff are neglected. Performance remains sensitive to the balance parameter, recursion depth, and convolutional/dense architecture, even though robustness is reported to be better than for some baselines. The study emphasizes $N$2D and $N$3D embeddings for interpretability rather than higher-dimensional embeddings for prediction. Generalizability beyond the three studied datasets remains open, as does extension to modalities such as EEG, MEG, or intracortical arrays. The paper further identifies supervised, multimodal, and predictive extensions as future directions, including long-range temporal models, trajectory-based biomarkers for clinical settings, and uses in neurofeedback or brain–computer interfaces (Zhou et al., 7 Aug 2025).

Taken together, these features define BCNE as a specialized framework for extracting low-dimensional brain-state trajectories from high-dimensional dynamic recordings through temporospatial preprocessing, correlation-aware image construction, convolutional mapping, and recursive manifold refinement. Its significance lies less in introducing a generic deep architecture than in specifying a particular computational pipeline for making dynamic neural structure visually and quantitatively legible across heterogeneous experimental paradigms.