cambrain-realtime-eeg-inference

name	cambrain-realtime-eeg-inference
description	CaMBRAIN methodology for real-time continuous EEG inference using causal Mamba state space models. First model enabling long-range streaming inference of variable-length EEG signals with >10x higher throughput.
version	1.0.0
author	arXiv paper extraction
arxiv_id	2605.28792
activation_keywords	["EEG","real-time inference","state space model","Mamba","causal model","continuous EEG","streaming inference","brain activity monitoring","neural signal processing"]
tags	["neuroscience","EEG","state-space-models","causal-inference","real-time-processing","deep-learning","computational-neuroscience"]
related_skills	["jet-eeg-flow-matching","eeg-foundation-model-adapters","mamba-spike-forecaster-bci","eeg-ieeg-bridge-bci"]

CaMBRAIN: Real-time, Continuous EEG Inference with Causal State Space Models

arXiv ID: 2605.28792 Authors: Abhilash Durgam, Nyle Siddiqui, Jeffrey A. Chan-Santiago, Qiushi Fu, Elakkat D. Gireesh, Mubarak Shah Submission Date: 2026-05-27 Categories: cs.AI, cs.HC, cs.LG

Overview

CaMBRAIN is the first Causal, Mamba-based state space model (SSM) capable of real-time inference of EEG signals. It addresses critical limitations in existing EEG deep learning approaches:

1. Quadratic scaling problem: Traditional attention-based models scale quadratically with sequence length
2. Fixed-length input constraint: Raw EEG must be processed in sliding windows, preventing global signal understanding

Core Innovation

Key Arguments

• Bidirectional approaches are needlessly expensive for EEG processing
• EEG is inherently causal and unidirectional - past signals influence future, not vice versa
• Causal SSM architecture is more appropriate than bidirectional attention

Technical Challenge

EEG exhibits extreme temporal dynamics:

• Crucial events can be extremely brief (fractions of a second)
• Events separated by long intervals (minutes)
• Current self-supervised objectives optimize for signal reconstruction
• These objectives fail to train hidden state to retain salient long-range context

Methodology

Multi-Stage Self-Supervised Training Pipeline

Designed specifically for streaming SSMs to:

1. Encourage long-range memory retention
2. Preserve linear-time complexity of state space models
3. Explicitly train hidden state for streaming inference context

Architecture Components

• Causal Mamba backbone: Unidirectional processing aligned with EEG temporal structure
• Streaming-friendly design: Continuous inference without sliding-window limitations
• Linear-time complexity: O(n) scaling vs O(n²) attention

Performance Results

State-of-the-art across 3 different EEG datasets:

• >10x higher throughput than existing models