How to make model training and re-training faster and cheaper? We envisioned a multi-use AutoML, aiming to create a solution superior to other HPO services in the market.
Created a novel solution for AI/ML for Techila's computing environment.
Used Bayesian optimization, gradient-free optimization, random&grid search as well as reinforcement learning.
Extended the product to a multi-use AutoML with black box optimization capabilities.
Techila Distributed Computing Engine is a next generation compute grid. It is a user-friendly solution that can improve the productivity of users from the research and development to applications that are deployed within production environments.
In this project, Techila wanted to create a novel solution for AI/ML for their computing environment. Together we identified hyperparameter optimization as a potential candidate. Our data scientists started working as part of Techila’s team, aiming to create a solution superior to other HPO services in the market.
Starting with this idea, we also extended the product to a multi-use AutoML with black box optimization capabilities. In practice, it takes in parameters and returns a score, which is used to fine-tune the input parameters. The solution is instantly usable by data science teams, reducing the need for manual processing of parameters. This makes model training and re-training a lot faster and, most importantly, cheaper.
For tech enthusiasts, we used for example Bayesian optimization, gradient-free optimization, random&grid search as well as reinforcement learning to come up with the state-of-the-art solution, with tools like Scikit-Learn, CatBoost, XGBoost, LGBM, Keras/Tensorflow and Pytorch.
These glimpses are just the tip of the iceberg. Together, let's turn your "what-ifs" into "how-tos," shaping a future we dare to dream.
Contact us
