Keeping Track of Machine Learning Experiments
Muru Selvakumar (~muru_selvakumar) |
Getting started with machine learning is easy. You just open a text editor and pump out all the code and train. But the learned model weights and auxiliary data produced from training are harder to keep track of, even with handful of hyper-parameters. When studying different networks and models over different datasets, it is an unmanageable mess.
This talk aims to inform and encourage the machine learning newbies to develop practices to keep track of their experiment data, which includes model weights, metrics history and auxiliary data like preprocessed dataset, vocabulary in case of NLP tasks.
The entire machine learning exercise is iterative. We try a method and see how it performs and tweak something and try again. This can go on for a long time. We make mistakes in between which deteriorates performance of the model. So it helps to keep track of the data, just like we keep track of the changes to the code. In this talk, we would go through one of such methods, that grew organically out of the last two years of experimentation.
- Model life-cycle 10mins
- Data consumed and produced 5mins
- Our workflow 10mins
- QA 5mins
- Machine learning beginners
Selva Kumar is interested in doing a bit of painting and 3d modeling. Passionate about linguistics and languages & culture in general. He is interested in AGI. He is also a proud, free software evangelist.
He has two publications,
- An Attentive Sequence Model for Adverse Drug Event Extraction from Biomedical Text
- Compositional Attention Networks for Interpretability in Natural Language Question Answering.