Explore Brain-Inspired Machine Intelligence for Connecting Dots on Graphs Through Holographic Blueprint of Oscillatory Synchronization

TL;DR

HoloBrain and HoloGraph are introduced, a brain-inspired framework that models oscillatory synchronization to overcome limitations of graph neural networks and enable more efficient, robust learning.

Abstract

Neural coupling in both neuroscience and artificial intelligence emerges as dynamic oscillatory patterns that encode abstract concepts. To this end, we hypothesize that a deeper understanding of the neural mechanisms governing brain rhythms can inspire next-generation design principles for machine learning algorithms, leading to improved efficiency and robustness. Building on this idea, we first model evolving brain rhythms through the interference of spontaneously synchronized neural oscillations, termed HoloBrain. The success of modeling brain rhythms using an artificial dynamical system of coupled oscillations motivates a "first principle" for brain-inspired machine intelligence based on a shared synchronization mechanism, termed HoloGraph. This principle enables graph neural networks to move beyond conventional heat diffusion paradigms toward modeling oscillatory synchronization. Our HoloGraph framework not only effectively mitigates the over-smoothing problem in graph neural networks but also demonstrates strong potential for reasoning and solving challenging problems on graphs.

Figures (14)

• Figure 1: From HoloBrain to HoloGraph.a Neuroscience $\rightarrow$HoloBrain. Task stimuli drive neural activity measured by fMRI. HoloBrain models oscillatory coordination on the structural network—via Kuramoto‐style dynamics with attending‐memory (control)—to recover task-evoked synchronization patterns. b Machine learning $\rightarrow$HoloGraph. Inspired by this mechanism, HoloGraph replaces heat-diffusion message passing with neural-oscillation–based synchronization on graphs, using outcome-specific feedback to align distant nodes into meaningful clusters.
• Figure 2: From wave interference to CFC patterns: physics-inspired fingerprints for neurodegenerative diseases. The physics insight of cross-frequency coupling (building block of HoloBrain) is analogous to the wave interference principle a, which both yields constructive (red) and destructive (blue) interference patterns on the screen and cross-frequency couplings, as shown in b. We present the node-wise averages of CFC patterns along with quantitative measures (sign consistency degree) that highlight disease-specific interference patterns in c AD using ADNI dataset (https://adni.loni.usc.edu/data-samples/adni-data/), d PD using PPMI dataset https://www.ppmi-info.org/, and e FTD using NIFD dataset (https://memory.ucsf.edu/research-trials/research/allftd). '$\ast$' denote statistically significant group differences at the levels of $p<0.05$ (two-sided Mann–Whitney U), respectively. The brain images are generated using Surf Ice rorden2025surficehttps://www.nitrc.org/projects/surfice/ (Surf Ice 6), a software for reading surface- and volume-based neuroimaging data.
• Figure 3: Model interpretability: task-specific neural synchronization in phase space. Each brain region is mapped to a unit-norm feature and visualized by its phase on the unit circle. Color coding: in panel (a) the colors denote canonical subnetworks; in panel (b) the colors denote congnitive tasks. a HCP-A examples. Columns show the evolution through HoloBrain: $\mathbf{X}^{(0)} \!\rightarrow\! \mathbf{Y}^{(0)} \!\rightarrow\! \mathbf{X}^{(1)}\ldots \mathbf{X}^{(L)}$. Tighter dots indicate stronger synchrony; e.g., VISMOTOR concentrates in visual/sensorimotor systems, whereas REST shows stronger default-mode synchrony. b Group-level dynamics. As depth increases, regional phases within the same task collapse toward a task-specific attractor (higher sign-consistency), while different tasks remain separable, revealing distinct synchronization fingerprints for each cognitive state.
• Figure 4: Whole-brain and regional interference patterns learned by HoloBrain.a Whole brain. For each cohort/task comparison, we show the group-average CFC matrices (red/blue colors indicate constructive/destructive interference). The box plots quantify the sign-consistency degree along the dominant off-diagonal. An asterisk denotes a significant difference between the two tasks (two-sided Mann–Whitney U, $p<0.05$). b Regional exemplars. Cortical regions with the strongest learned synchrony are highlighted on the brain surface (visualized using Surf Ice rorden2025surficehttps://www.nitrc.org/projects/surfice/), and their regional CFC matrices are shown below each parcel. These examples illustrate that HoloBrain captures task-specific interference fingerprints both at the whole-brain level and within characteristic regions.
• Figure 5: Overall performance of HoloBrain.a Comparison of nine methods on HCPA, HCPYA, HCPYA-WM, ADNI, PPMI, and NIFD datasets. Bars show accuracy (Acc), precision (Pre), and F1-score (F1), with red asterisks marking cases where HoloBrain significantly outperforms competing methods (paired $t$-test, $p<0.05$ and $p<0.001$). Red fonts denote the best performance. b Neural synchronization levels, both at the whole-brain and regional scales, exhibit significant group differences between CN and AD, CN and PD, and CN and FTD. The group comparison results at the whole-brain level are shown in the shaded bounding box. At a significance level of $p<10^{-5}$, we display the brain regions with significant health vs. disease differences as well as the distributions of KOP degree for CN (green) and disease (red) subjects. In addition, brain regions with reduced synchronization levels associated with disease pathology are highlighted in green, while those with increased synchronization are marked in red. The brain images are visualized using Surf Ice rorden2025surfice (https://www.nitrc.org/projects/surfice/). c Clustering results of HoloBrain versus spectral clustering, compared against the ground truth. Red circles highlight discrepancies from the ground truth. Different clusters (colors) are mapped in brain surface using ParaView (v5.10.1) ahrens2005paraviewhttps://www.paraview.org/.