Live Neural Processing

Decoding Brain Signals
In Real-Time

Revolutionary AI framework that simultaneously denoises and classifies motor imagery from EEG signals without user calibration

62.1%
LOSO Accuracy
109
Subjects
64
EEG Channels
View on GitHub Learn More
PyTorch 2.0+ Python 3.8+ Streamlit MNE-Python
Problem & Solution

Addressing Critical BCI Challenges

Revolutionizing motor imagery classification through joint learning

Challenge 1: Noisy EEG Signals

EEG signals are contaminated with artifacts from EMG, eye movements, and environmental noise, making accurate classification difficult.

Challenge 2: Subject Variability

Traditional BCIs require extensive calibration for each user, limiting practical deployment and scalability.

Our Solution: Joint Learning

Simultaneously denoise and classify through shared representations, enabling zero-shot cross-subject generalization.

Innovation

Key Contributions

01

Joint Learning Framework

Simultaneous optimization of denoising and classification objectives through a shared encoder architecture.

  • End-to-end trainable
  • Task-aware denoising
  • No separate preprocessing
02

Subject-Independent Inference

Leave-One-Subject-Out (LOSO) training enables zero-shot inference on completely new users.

  • No calibration required
  • 109 subjects evaluated
  • Ensemble strategies
03

Production-Ready System

Interactive Streamlit interface with real-time inference and comprehensive visualizations.

  • Web-based interface
  • Signal visualization
  • Export capabilities
Technical Details

Model Architecture

Hybrid EEGNet-inspired design with dual-head outputs. Click on any block to learn more!

Input EEG (noisy)
shape: (B, 64, 641)
Unsqueeze dim=1
→ (B, 1, 64, 641)
Shared encoder (EEGNet-inspired)
Conv2d — temporal filter
1 → 16 filters, kernel (1, 64), padding=same
BatchNorm2d(16)
Conv2d — depthwise spatial filter
16 → 32 filters, kernel (64, 1), groups=16
BatchNorm2d(32) + ELU
Separable Conv2d
Depthwise: (32→32, kernel(1,16), groups=32)
Pointwise: (32→32, kernel(1,1))
BatchNorm2d(32) + ELU + Dropout(0.25)
→ (B, 32, 1, 641)
branches into two heads
Denoising head
Squeeze dim=2
→ (B, 32, 641)
Conv1d — refine features
32→32, kernel=9, pad=4
BatchNorm1d(32) + ReLU
Conv1d — project to channels
32 → 64, kernel=1 (residual)
Residual add
x_hat = x_noisy + residual
x_hat — denoised EEG
Classifier head
AdaptiveAvgPool2d (1,1)
(B,32,1,641) → (B,32,1,1)
Flatten — (B, 32)
Dropout(0.50)
Linear(32 → 2)
raw class logits
Softmax (inference only)
left / right MI probability
Joint loss
L_total = CrossEntropy(logits, y) + 0.2 × MSE(x_hat, x_clean)
F1=16 D=2 temporal_kernel=64 enc_dropout=0.25 cls_dropout=0.50 λ=0.20
Adam lr=1e-3 · batch=32 · epochs=50 (best checkpoint)
Data

PhysioNet EEGMMI Dataset

Large-scale motor imagery database with 109 subjects

109
Healthy Subjects
64
EEG Channels
160 Hz
Sampling Rate
4 sec
Trial Duration

Motor Imagery Tasks

Task 2

Left vs Right Hand

Runs: R04, R08

Classes: T1 (Left) | T2 (Right)

Status: ✓ Completed (109 subjects)

Task 4

Fists vs Feet

Runs: R12, R16

Classes: T1 (Fists) | T2 (Feet)

Status: ⏳ Evaluation Pending

Performance

Experimental Results

Comprehensive LOSO evaluation on 109 subjects

LOSO Task 2 Performance

62.1% Mean Accuracy
±12.6% Std • F1: 0.595 • SNR: +0.007 dB

Accuracy Distribution

62.1%

Mean accuracy across all 109 subjects

F1-Score (Macro)

0.595

Balanced performance across classes

SNR Improvement

+0.007 dB

Conservative denoising preserves signal integrity

Subject Variability

40%
96.7%

Wide performance range across subjects

Key Insights

Cross-Subject Generalization: Successfully achieves 62.1% accuracy without subject-specific calibration

High Variability: ±12.6% std indicates significant inter-subject differences remain challenging

Top Performers: Best subjects (S072, S085) achieve 96.7% accuracy, showing model capacity

Signal Preservation: Modest SNR improvement (+0.007 dB) indicates conservative, task-aware denoising

Interactive

Streamlit Web Application

Production-ready interface for real-time EEG inference

localhost:8501

Streamlit Interface

Application Features

Multi-Strategy Inference

Ensemble (3-10 models), Single Model, or Best Model selection

Auto-Detection

Handles both PhysioNet multi-trial and single-trial EDF formats

Interactive Visualizations

Probability distributions, SNR improvement, confusion matrix, signal comparison

Export Options

Download results as CSV, NPY, or NPZ with timestamped filenames

Launch Application

Run locally with: streamlit run app.py

Technology

Tech Stack & Dependencies

Deep Learning

PyTorch 2.0+ torchvision CUDA

EEG Processing

MNE-Python 1.5+ scipy numpy

Visualization

Streamlit Plotly matplotlib

Analytics

scikit-learn pandas tqdm
Get In Touch

Contact & Resources

GitHub Repository

View source code, documentation, and contribute

github.com/AKASH7294 →

Email

Questions or collaboration inquiries

Contact via GitHub

Documentation

Comprehensive README and usage guides

View Docs →

Citation

@misc{mi_joint_denoise2026,
  title={AI-Driven Joint EEG Denoising and Motor Imagery Classification},
  author={AKASH7294},
  year={2026},
  publisher={GitHub},
  url={https://github.com/AKASH7294/mi_joint_denoise}
}