arXiv weekly scan

AI-generated summary

• Adapts chemical foundation-model representations to reduce collisions between differently labeled molecules after conversion to short bit strings.
• Evaluates blood–brain barrier permeability prediction using ChemBERTa and simulated quantum circuits.
• Reports higher accuracy than a logistic-regression baseline on identical inputs at 10 qubits; this is a restricted model comparison.

Significance

Examines how information lost during discretization affects subsequent classification.

Relation to lab research

Offers a reference for assessing information loss when compressing brain representations. Adaptation to fMRI or EEG remains untested.

Summary

• EEG-PRIME is a novel two-stage foundation model for EEG decoding designed to overcome cross-dataset and cross-subject generalization issues.
• It employs masked pretraining with spectral augmentation followed by prototype-aligned instruction tuning, incorporating task-semantic, dataset-specific, and subject-invariant conditioning for robust decoding.
• EEG-PRIME demonstrates superior performance on diverse EEG tasks and exhibits strong zero-shot transfer capabilities on unseen datasets, outperforming existing baselines and foundation models.

Significance

This work represents a significant advancement in EEG decoding by developing a foundation model capable of robust and generalizable performance across diverse datasets and subjects without extensive calibration, paving the way for more practical brain-computer interfaces.

Relation to lab research

The development of EEG-PRIME aligns with the SNU Connectome Lab's Neuro-X initiative, as it focuses on creating generalizable AI models for neural signal decoding, a critical component for advancing neurotechnology.

Summary

• EEG-PRISM is a novel method that translates time-channel attribution scores from EEG foundation models into more physiologically relevant domains, specifically the frequency and source domains, without altering the original model.
• The method utilizes linear transformations, including an invertible DFT for frequency mapping and an approximately invertible EEG generative model for source localization, to provide clinically intuitive interpretations.
• Evaluations in both simulated and real EEG data (epilepsy and autism) demonstrated EEG-PRISM's effectiveness in recovering ground-truth spectral activity and localizing seizure onset regions and predictive biomarkers with reasonable accuracy.

Significance

This work introduces a crucial tool for making advanced EEG foundation models interpretable in a clinically meaningful way, thereby enabling better understanding of neural mechanisms and identification of disease biomarkers.

Relation to lab research

EEG-PRISM's focus on interpretable AI for EEG foundation models directly aligns with SNU Connectome Lab's Neuro-X initiative, which aims to develop advanced neurotechnology with a strong emphasis on understanding and explaining neural data.

Summary

• The paper proposes Brain Latent Predictive Model (BLPM), an EEG-language foundation model that redefines EEG decoding as a continuous semantic embedding prediction problem.
• BLPM introduces a Continuous EEG Latent Predictive (CELP) encoder for learning transferable representations and a Multi-Query Semantic Decomposition (MQSD) module for aligning EEG with textual semantics in a shared latent space.
• The model aims to overcome limitations of masked autoencoding and autoregressive modeling in current EEG foundation models by prioritizing task-relevant semantics and addressing the mismatch between continuous neural dynamics and discrete token spaces.

Significance

This work establishes a novel paradigm for EEG-language foundation models, demonstrating improved generalization and alignment of continuous EEG signals with natural language semantics.

Relation to lab research

This research is highly relevant to the 'Large Brain Models' and 'Neuro-X' initiatives at the SNU Connectome Lab, as it advances the development of foundation models for integrating brain activity with artificial intelligence and language processing.

Summary

• Merges redundant EEG tokens while preserving feature magnitude to reduce the sequence length processed by Transformer layers.
• Compares compression methods across EEG models and datasets; performance relative to the uncompressed model varies by setting.
• Reports reduced inference time in a local-device deployment case study with LaBraM.

Significance

Examines the trade-off between token compression, classification performance and execution time; a device benchmark does not establish clinical utility.

Relation to lab research

Suggests a compression method to test in Neuro-X EEG models, with evaluation of which predictive information is retained or lost.

Summary

• This paper introduces a novel security risk in EEG foundation models: adversarial transfer attacks from public encoders to private downstream models.
• They propose SW-ProxyCE, a query-free attack framework that leverages shrinkage-whitened class prototypes from a small labeled reference set to generate transferable adversarial examples.
• SW-ProxyCE effectively attacks inaccessible downstream models across various EEG tasks and foundation encoders, outperforming task-agnostic attacks and demonstrating that foundation model transferability does not imply adversarial robustness.

Significance

This research highlights a critical security vulnerability in the deployment of EEG foundation models, urging the development of robust defenses to protect private downstream applications.

Relation to lab research

This work is highly relevant to Neuro-X, as it addresses critical security and robustness challenges pertinent to the deployment and trustworthiness of large-scale neural data models.

Summary

• STEAM is a novel spatio-temporal alignment Mixture-of-Experts (MoE) model designed for EEG decoding, addressing challenges in generalizability and adaptation costs of conventional BCI algorithms.
• It employs a dual-branch spatio-temporal encoder with a Shared Soft Mixture-of-Experts (SSMoE) module to align and exchange information between spatial and temporal representations.
• The model utilizes a hierarchical pre-training strategy that allows for general initialization and subsequent specialization to specific EEG paradigms, achieving superior performance across diverse datasets and evaluation settings.

Significance

This work introduces a foundational model for EEG decoding that significantly advances the generalizability and adaptation efficiency of BCI systems, crucial for widespread adoption in clinical and rehabilitative applications.

Relation to lab research

STEAM aligns with the Connectome Lab's focus on Large Brain Models and Neuro-X by developing a sophisticated deep learning architecture for robust and generalizable neural signal decoding, applicable to understanding and leveraging complex brain activity.

Summary

• EEG foundation models exhibit a pervasive low-frequency bias in their learned representations, irrespective of data scale, model capacity, or pretraining objective.
• This bias stems from the interplay between EEG's $1/f^α$ spectral characteristics and neural networks' natural inclination to prioritize low-frequency components, exacerbated by $\ell_2$ reconstruction losses in masked autoencoders.
• FAME (Frequency-balanced Masked Autoencoding) is proposed, which balances spectral supervision by reconstructing time-frequency activity in predefined EEG bands with independent standardization and equal loss weighting, leading to improved, spectrally balanced representations and state-of-the-art performance on 24 out of 41 downstream tasks.

Significance

This research identifies a critical and persistent spectral bias in EEG foundation models, offering a novel frequency-balanced solution that significantly improves representation learning and downstream task performance, pushing the boundaries of transferable EEG AI.

Relation to lab research

This work directly impacts the development of Large Brain Models and Neuro-X initiatives at SNU Connectome Lab by providing crucial insights and a solution for training more effective and balanced EEG foundation models, which are fundamental components for understanding brain activity across diverse tasks and populations.

Summary

• EEG-JEPA introduces a structured latent-prediction framework for EEG foundation modeling, moving beyond simple masked waveform reconstruction to infer latent states.
• The framework employs a masked context encoder and predictor to infer contextual latent states from a target encoder observing complete input, utilizing Neurotopology-Aware Multi-scale Electrode-Temporal Masking (N-MET) for structured electrode-time prediction.
• EEG-JEPA demonstrates significant improvements in frozen multitask transfer (from 40.49% to 50.42%) and full fine-tuning (from 68.98% to 70.65%) compared to masked waveform reconstruction, setting new benchmarks on EEG-FM-Bench.

Significance

This work establishes a more effective pretraining strategy for EEG foundation models, enabling them to learn more transferable and neurally relevant representations by predicting structured latent states rather than noisy raw signals.

Relation to lab research

This research is highly relevant to the 'Large Brain Models' focus of the SNU Connectome Lab, as it develops advanced foundation models for EEG, a key modality for understanding brain function, and optimizes their pretraining for better generalizability and transferability across various tasks.

Summary

• Models whole-brain fMRI as patch sequences, removing non-brain background tokens.
• Uses UK Biobank, ABCD and HCP for autoregressive pretraining, followed by HCP sex classification.
• Evaluates computational savings and classification performance; prior-model comparisons use different data splits.

Significance

Develops a computationally efficient approach to whole-brain learning that retains spatial detail.

Relation to lab research

Extends Neuro-X’s whole-brain representation learning through computational efficiency.

Summary

• Examines relationships among genes, brain, behavior and environment in childhood using ABCD data.
• Analyzes covariation among 33 polygenic scores, MRI modalities, cognition, mental health and environmental measures.
• Finds covariation involving cognitive-related scores and brain measures; separate PGS–non-imaging components include neighborhood income, poverty and parental acceptance.

Significance

Places genetic differences alongside brain measures and lived context to study individual differences.

Relation to lab research

Provides a concrete study of relationships across genetic, brain, behavioral and environmental levels.

Summary

• Combines spatiotemporal attention with channel permutation equivariance for EEG representation learning.
• Fine-tunes the pretrained model for emotion recognition with 32 channels and motor imagery with 64 channels.
• Remains stable under synthetic channel permutations; relative performance varies by task.

Significance

Treats channel ordering as a design and transfer-evaluation problem in EEG modeling.

Relation to lab research

Develops the lab’s EEG representation learning around differences between recording setups.

Summary

• Critiques the traditional linear regression assumption between LLM activations and fMRI responses.
• Shows that nonlinear multimodal mappings recover brain-language correspondence that linear encodings underestimate.
• Connects predictive-processing views of cortex with autoregressive Transformer internal states.

Significance

Re-evaluates the mathematical foundation of how artificial neural representations correspond to biological neural activity.

Relation to lab research

Connectome Lab preprint; theoretical grounding for the Neuro-X Large Brain Model initiative.