This repository contains research code developed as part of an ongoing project targeting submission to ICML 2026.
The work investigates how computer vision–inspired techniques—such as representation learning and predictive modeling—can be adapted to enhance anomaly prediction in time series data.
Our goal is to build robust and generalizable frameworks that bridge methodological advances in machine learning with real-world data science applications.
Anomaly prediction in time-series data is a fundamental challenge across diverse domains, including finance, healthcare, industrial IoT, energy systems, and cybersecurity. Anticipating irregular patterns enables fraud prevention, early diagnosis, predictive maintenance, grid stability, and intrusion forecasting. Leveraging advances in representation learning and Computer Vision, this project introduces vision-inspired predictive frameworks to deliver robust, generalizable solutions beyond domain-specific heuristics.
|--- VQVAE/ # Vision-inspired quantization modules
|--- jepa/ # Predictive embedding architecture
|--- preprocess/ # Data preprocessing and normalization
|--- train/ # Training pipelines
|--- results/ # Logs, evaluation outputs, and visualizations
|--- data/ # Dataset loaders and utilities
This project is ongoing research.
Code here is primarily for experimentation and reproducibility.
- Research domain: Time-Series Anomaly Prediction, Deep Learning, Machine Learning
- Techniques adapted from Computer Vision into temporal modeling
Details may evolve later.