ICML 2019 Expo Demo
July 12, 2020
Addressing practical barriers to optimizing applied learning models
There are a wide variety of learning-model optimization challenges that are well-researched in academia. By deeply exploring each particular challenge, this academic research creates a robust foundation for applied research situations. But bridging the gap from the academic to the applied setting is not as simple as transferring the code or capability from a paper to your notebook.
Data preprocessing, metric selection, architecture selection, distributed training, and hyperparameter search each introduce complications in the modeling process, and these complications are often exacerbated in applied circumstances. Some of these - such as iterating on the metric - are tough to solve without deep domain and machine learning expertise. But others, such as distributed training, can largely be addressed with software solutions, modeling techniques, or a combination of the two.
In this demo, SigOpt’s research shares modeling techniques and software products that are the result of addressing some of these real-world barriers to building productive models. Included in this demonstration are the following research areas:
Multimetric Optimization: Solving for competing objectives by efficiently providing a variety of optimal modeling configurations that can be compared Multitask Optimization: Using algorithmic techniques to optimize particularly expensive models by learning from partial cost versions of the model to reduce wall-clock time Reproducibility: Running, tracking and analyzing experiments in a way that lends itself to efficiently cataloging and reproducing models or experiments for alternative use cases Transferability: Transfer insights from any given model configuration to new models designed to solve similar problems but with novel datasets Distributed Training: Orchestrating access to compute to efficiently distribute training in parallel to reduce the wall-clock time to optimize any model