Artificial Intelligence Lab — Research Contributions

This repository hosts all benchmarking, evaluation, and fine-tuning experiments on large pre-trained time-series models (LPTMs) under Dr. Prakash’s supervision.

Benchmark Datasets

I evaluated Foundational Time-Series Models on eight standard datasets:

• ETT1
• ETT2
• Flu-US
• PEM-Bays
• NY-Bike Demand
• NY-Taxi Demand
• Nasdaq
• M4

Model Suite & Comparisons

I compared and fine-tuned the following Foundational Time-Series Models:

1. LPTM
2. MOMENT
3. TimesFM
4. Chronos
5. MOIRAI
6. TinytTimeMixers

All experiments included systematic fine-tuning on each model to assess adaptability across datasets.

Backend API

In backend/, you’ll find a RESTful Flask API deployed on a private NVIDIA DGX server using reverse-proxy and continuous testing with Postman & Ngrok. It supports:

• Model loading & versioning
• Dataset uploads
• On-the-fly fine-tuning
• Inference endpoints

Data & Visualizations

• Data: Raw benchmark datasets live in data/.
• Visualizations: All performance charts and plots are saved under benchmark_visualizations/.

Experiments

Detailed notebooks and logs of each experimental run are in experiments/. These include:

• Training/fine-tuning configurations
• Metrics tracking (MAE, RMSE, etc.)
• Ablation studies and hyperparameter sweeps