Adopting Hyperparameter Tuning to Get the Best Possible Deep Learning Models

We’re back again with the articles and speaking to AI trends that were touched upon during Munich AI Summit 2019 hosted by Develandoo AI Innovation Lab. We’ve written about such topics as data protection and privacy in AI, machine learning in the banking industry, women in AI, the importance of adopting new data-driven solutions, and others.

Today we’re going to dive deep into another thought-provoking talk ”Pragmatic hyperparameter tuning” by Timon Ruban from Luminovo.

Timon graduated top of his class from ETH Zurich. During his MSc at Stanford, he dug his heels into deep learning. He helped teach Andrew Ng’s machine learning class, applied computer vision algorithms at Google’s smart imagery team, and now leads Luminovo’s machine learning efforts.

He did a quick review of different hyperparameter tuning strategies, keeping in mind the constraints faced in real-world ML projects and finished the talk with a deep dive on Population-Based Training (PBT) and a combination of Bayesian Optimization and Hyperband (BOHB) two simple but efficient algorithms for automated hyperparameter tuning.

He starts his topic by giving an overview of the deep learning lifecycle for deploying models into production and raises two important questions; how do you split your data and how do you evaluate? 

”Making efficient use of your available resources (scarce) to find the best possible model for a given dataset and evaluation metric.”

This is what hyperparameter tuning is all about, and he highlights two points in this definition ”given dataset” and ”evaluation metric” which assumably answer the above-mentioned questions.

He also gives insights into 3 paths towards better machine learning models;

  • waiting for better hardware
  • making more efficient use of existing hardware
  • finding better apriori insights

Timon further explains how hyperparameter tuning is done today, shares his own experience of how his machine learning journey has started and dives deep into the topic based on some algorithms. 

This was just a snippet from our speakers’ talk. If you want to learn more about the topic, you can visit our youtube channel where you’ll find the videos of all the Munich AI Summit talks.

We are thankful to all our sponsors (Fujitsu, Women in AI, Liquid Newsroom, Scylla and Urgestain, Wayra Germany), participants and the speakers for attending our Munich-AI Summit 2019, the one and only free event in Munich area related to AI.

As we have already announced our Meetup group ‘’Power Humanity With AI’’ is going Global & Virtual because of the unprecedented global situation caused by the outbreak of COVID-19. This means in the upcoming period, Develandoo will be organizing and running a series of webinars and virtual online sessions which will include leading professionals in the field of AI and Machine Learning. 


Develandoo is the leading AI company in Armenia. As an innovation lab, Develandoo has successfully created over 30 products for global enterprises such as KPIT Germany and Welocalize USA.

Develandoo has also incubated three in-house startups which are leading the markets in their areas of expertise:

  1. Scylla – World’s leading gun detection system
  2. Cibola – World’s first in-store analytics system
  3. Protogen – Generic Tabular Data platform

For more information, please visit –





  • Topics:
  • Meetup

Top Stories

High Five! You just read 2 awesome articles, in row. You may want to subscribe to our blog newsletter for new blog posts.