hanbuck30 / pytorch_esn

Pytorch version of ESN. Learning algorithms include FORCE, inverse matrix, and Gradient Descent. Unlike traditional ESNs, large storage nodes can be used using memory much more efficiently.

Hi there, I'm Hanbeot Park👋

I'm a dedicated Researcher in the fields of Bio-Plausible Neural Networks, Brain-Computer Interfaces (BCI), and Financial Data Analysis. Currently pursuing my Master's degree in Data Engineering at Pukyong National University, my research focuses on the prediction of nonlinear time series data in BCI and financial applications. I’m also deeply engaged in developing next-generation neural networks to enhance the efficiency and performance of existing models.

🚀 Technical Skills

Languages

Specializations

Tools & Frameworks

Databases

DevOps

💼 Research Experience

🔬 Ongoing Research

• BXAI Lab, Pukyong National University - Graduate Researcher
  • Conducting research on interpolation methods from low-channel to high-channel EEG data, aiming to enhance the precision of BCI systems.
  • Developing encoding techniques for Spiking Neural Networks (SNNs) to explore their potential in high-performance computing.
  • Enhancing Word Recognition in Imagined Speech Using Non-Invasive EEG for Improved BCI Applications
  • Engaged in predictive modeling of nonlinear time series data for financial and BCI applications, utilizing state-of-the-art neural networks.

🧠 Past Research & Studies

• Multi-modal Learning (CV, NLP) - Dacon Project

  • Conducted research on improving image classification by analyzing and enhancing how text within images can be better utilized for classification purposes. This involved developing techniques to effectively integrate the visual and textual components of images to improve overall model performance.

• Deep Reinforcement Learning (DeepRL) - NeuroMatch Project

  • Conducted research aimed at advancing bio-plausible neural networks by applying deep reinforcement learning techniques. The project involved training an Ant agent within the OpenAI Gym environment to perform desired actions using bio-plausible learning rules. The goal was to enhance the performance of bio-plausible neural networks by incorporating biologically inspired learning mechanisms.

• Computer Vision (CV) Applications - Capstone Project

  • Conducted research on ICT technology that detects umbrellas using computer vision, enabling automatic door closure during rainy weather. This project aimed to integrate real-time weather responses into smart home systems, enhancing convenience and safety.

• Stress Prediction - Previous Study

• Neuromorphic Computing Research - Previous Study

• Inner Speech Decoding - Previous Study

• ESN Research - Previous Study

• Stock Prediction - Previous Study

• Explainable-AI-Design-ESN-Based-Model-for-Depression-Diagnosis -Previous Study

🔧 Notable Projects

NeuroTalk Decoding

Traditional invasive methods have often achieved high performance in decoding. However, decoding using non-invasive methods like EEG is very challenging, especially for imagined speech, where understanding a user's intent through mere thoughts is particularly difficult. To address this, we developed a method that achieved high performance.

• Tech Stack: Python, PyTorch, NumPy, MATLAB, Pandas
• Features:
  • Achieved a high F1 score
  • Plan to further develop this method not only for decoding but also for language processing
  • It takes about 7 Hours for training.

Nonlinear Time Series Prediction for Financial Data

This project investigates the use of advanced neural network models for predicting nonlinear time series data, specifically focusing on applications in BCI and financial markets.

• Tech Stack: Python, Scikit-learn, Pandas, Pytorch
• Features:
  • Optimized the hyperparameters of the ESN and the TA using a heuristic optimization technique, GA (Genetic Algorithm)
  • Aimed for higher profits by utilizing optimized algorithms and TA tailored to each stock
  • Achieved stable and high returns by considering not only the profitability but also the Maximum Drawdown (MDD)

ESN-Pytorch

Pytorch version of ESN. Learning algorithms include FORCE, inverse matrix, and Gradient Descent. Unlike traditional ESNs, large storage nodes can be used using memory much more efficiently.

• Tech Stack: Python, NumPy, Pytorch
• Features:
  • When using many nodes, you can effectively manage and utilize memory
  • By using batch processing, tasks can be performed at high speed through parallelization
  • It can be utilized as needed through various learning algorithms

📈 GitHub Stats

📊 Most Used Languages

🌱 What I'm currently learning

• Meta-Reinforcement Learning (Meta-RL)
• Latest state-of-the-art neural networks architectures for time series data
• Neuroscience and its intersection with AI
• Quantum Computing (Future Interest)

📫 How to reach me

• Email: [email protected]
• LinkedIn: linkedin.com/in/hanbuck30

Feel free to reach out if you want to collaborate on research projects, discuss neural networks and their applications, or explore the exciting fields of BCI and financial data analysis!

🌟 Fun Fact

I'm fascinated by the idea of replicating the human brain within a computer, and I'm closely following the developments in Spiking Neural Networks (SNNs) to make this a reality. I'm also deeply interested in EEG encoding and decoding, and how these can be applied to cutting-edge AI systems.

---

"Research is formalized curiosity. It is poking and prying with a purpose." – Zora Neale Hurston