AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

A Hierarchical Tokenization Framework for Voxel-Level fMRI Representation Learning

Karan Singh, Yue Zhao, Mohammadhassan Abbasi, Ehsan Adeli

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Contrastive representation theory · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HBAR, a hierarchical tokenization framework that organizes voxel-level resting-state fMRI into sub-parcel voxel groups, parcels, and functional networks for compact multi-scale representation learning.

Problem

Voxel-level fMRI is high-dimensional, and conventional parcel-based representations discard fine-grained spatial information; a compact yet detailed representation is needed for neuroimaging learning.

Approach

A hierarchical tokenization over resting-state fMRI spanning sub-parcel voxel groups, parcels, and large-scale functional networks, giving about 415x compression while retaining fine-grained spatial detail beyond parcel-level methods.

Key results

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use atlas-based hierarchy instead of flat tokenizer at 415xHBAR-415 · HCP-YA reconstruction (415x)parcel correlation: 0.691 -> 0.746 (vs flat RVQ)Improved
use hierarchical HBAR instead of MS-VQVAE at 415x for voxel reconstructionHBAR-415 · HCP-YA reconstruction (415x)voxel correlation: 0.427 -> 0.405 (MS-VQVAE vs HBAR)Hurt
use hierarchy instead of flat BSQ at 52xHBAR-52 · HCP-YA reconstruction (52x)voxel and network correlation: voxel 0.529 vs 0.526; network 0.968 vs 0.965No clear effect
use hierarchy instead of flat BSQ at 52xHBAR-52 · HCP-YA reconstruction (52x)parcel correlation: 0.890 -> 0.920Improved
keep anatomical chunk-to-parcel grouping instead of shuffling (ablation)HBAR-52 · HCP-YA reconstruction (52x)parcel correlation: 0.886 -> 0.920Improved
keep anatomical grouping instead of shuffling (ablation)HBAR-52 · HCP-YA reconstruction (52x)voxel and network correlation: voxel 0.523 vs 0.526; network 0.966 vs 0.965No clear effect
use fixed atlas pooling instead of learned cross-attention pooling (ablation)HBAR-52 · HCP-YA reconstruction (52x)parcel correlation: 0.876 -> 0.920Improved
use fixed atlas pooling instead of soft alignment pooling (ablation)HBAR-52 · HCP-YA reconstruction (52x)parcel correlation: 0.796 -> 0.920Improved
HBAR instead of flat tokenizers for linear-probe demographics at 415xHBAR-415 · HCP-YA age/sex linear probe (415x)AUROC: sex 0.701 vs 0.722 (RVQ); age 0.482 vs 0.607 (BSQ)No clear effect
HBAR instead of flat tokenizers for AR readout at 415xHBAR-415 · HCP-YA AR readout (415x)AUROC: age 0.506 -> 0.584; sex 0.556 -> 0.638 (vs flat BSQ)Improved
HBAR instead of flat BSQ for sex AR readout at 52xHBAR-52 · HCP-YA sex AR readout (52x)AUROC: 0.640 -> 0.792Improved
add temporal-difference loss terms (ablation)HBAR-52 · HCP-YA 52x dev ablationvoxel recon MSE: 0.810 -> 0.673Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy