Solving the challenges of massively parallel hyperparameter tuning at scale

As anyone who has trained deep learning models knows, a single run can take days or even weeks to finish, draining considerable time and budget from any organization. But what if we could shrink that timeline and reach stronger results along the way?

Earlier this quarter, our chief scientist and a visiting professor of machine learning walked members of a regional deep learning (DL) community group through the problem of massively parallel hyperparameter optimization and introduced a fresh way to think about model training. Attendees received a grounded overview of Automated Machine Learning (AutoML); identified two core challenges within this space, namely hyperparameter optimization and neural architecture search (NAS); and came away understanding how NAS is really a specialized subproblem of hyperparameter optimization.

The talk laid out how modern hyperparameter optimization actually behaves in a deep learning setting, zeroing in on the practical obstacles engineers hit when they juggle a large number of tunable hyperparameters alongside training runs that can stretch from days into weeks. To search these sprawling spaces properly, DL practitioners have to evaluate an enormous number of configurations. And while they would love to lean on parallel and distributed computing to accelerate the hunt, they almost always face orders of magnitude more configurations than they have parallel workers, which quietly erodes the payoff from many conventional parallelization schemes.

The team set out to close this gap by proposing ASHA, a lean and dependable hyperparameter tuning algorithm with strong theoretical footing that leans on asynchronous parallelism and aggressive early-stopping. Their thorough empirical study shows that ASHA edges out leading hyperparameter tuning approaches such as Population Based Training (PBT) and Bayesian Optimization and Hyperband (BOHB) ; scales linearly with the number of workers across distributed settings; reaches a high-quality configuration in half the time required by a well-known internal tuning service in a run with 500 workers; and holds its own against purpose-built NAS methods on standard benchmarks.

In fact, Gradyn's proprietary hyperparameter optimization method builds directly on the research-grade ASHA approach and folds it cleanly into our end-to-end deep learning platform.

View the full presentation to dig in deeper:

If you want to go further on the subject of hyperparameter optimization, we suggest starting with these resources:

Massively parallel hyperparameter tuning at scale

Architecture search explained

The next era of machine learning tooling

Which hyperparameter optimization challenges have you run into, and how did you work around them? Begin a conversation with us.