Experiment Exchange

Maintaining oil and gas machinery is expensive—but predictive maintenance models can help engineers minimize repairs and downtime.Shayan Mortazavi and Alex Lowden, Data Scientists at Accenture in the Industrial Analytics Group, work on the development of predictive maintenance models to minimize downtime of systems. In this episode, they discuss the complications when building these models, such as limited access to failure data and the massive number of features available, as well as the need for explainability and interpretability in their models. They also share how SigOpt’s parallelism feature allowed them to accelerate model development.1:27 - Intros3:05 - Machinery maintenance, then vs now4:06 - Goals of maintenance6:49 - Challenges of predictive maintenance for oil and gas8:31 - Human in the loop element10:07 - Interpretability11:42 - Using SigOpt to optimize hyperparameters13:50 - Managing multiple LSTMs16:38 - Using SigOpt's multimetric optimization18:36 - Predicting ultimate machine failure20:39 - Getting teams on board with AI-based tools23:21 - Overconfidence of AI Learn more about SigOpt at sigopt.com and follow us on Twitter at twitter.com/sigoptLearn more about Accenture: https://www.accenture.comSubscribe to our YouTube channel to watch Experiment Exchange interviews: https://www.youtube.com/channel/sigopt

How do you design a better glass?Paul Leu, Associate Professor of Industrial Engineering at the University of Pittsburgh, shares the insights behind his interdisciplinary work using machine learning to design and test novel glass structures. He and Michael McCourt discuss the collaboration between Pitt and SigOpt, the challenges of glass design and testing, and what's ahead for the Year of Glass.0:23 - Intros5:14 - Paul's interdisciplinary philosophy7:04 - Nanomaterials research11:48 - Designing for the properties of light14:35 - How Paul and his team use SigOpt17:50 - Designing better solar panels21:05 - How Paul uses SigOpt's advanced features22:51 - The Year of Glass24:16 - Paul's work with MDS-Rely Learn more about SigOpt at sigopt.com and follow us on Twitter at twitter.com/sigoptLearn more about LAMP: https://lamp.pitt.edu/ Learn more about MDS-Rely: https://mds-rely.org/ Learn more about the Year of Glass: https://www.iyog2022.org/ Subscribe to our YouTube channel to watch Experiment Exchange interviews: https://www.youtube.com/channel/sigopt

Given a property, what’s the material or the molecule that achieves it?This is the question behind some of Rafael Gomez-Bombarelli's latest work. Tune in as SigOpt's Michael McCourt interviews the Assistant Professor of Materials Processing at MIT about his development of machine learning strategies to design new materials—including fluids, cloths, metals, and nanomaterials. 1:28 - Intros2:18 – Using ML to predict physical properties of molecules3:48 – Rafael's active learning process5:23 - ML costs6:34 – How Rafael uses SigOpt in his work7:59 - Modeling the color change properties of molecules9:26 - Multi-fidelity methods12:12 – Continuum simulations14:57 – Inverse design problems18:50 - Workshops at ML conferences20:47 – What's next for RafaelLearn more about SigOpt at sigopt.com and follow us on Twitter at twitter.com/sigopt Subscribe to our YouTube channel to watch Experiment Exchange interviews: https://www.youtube.com/channel/sigopt

Aluminum design is an incredibly complicated business. Not only do you have to get the model design right—it also has to work in the real world. In fact, the aluminum soda can is one of the most engineered products in your house right now. In this episode, SigOpt’s Head of Engineering Michael McCourt talks with Vishwanath Hegadekatte, R&D Manager at Novelis, about how he's using tools like SigOpt to optimize aluminum production and design—as well as considering environmental impact to build better products and conserve resources. 2:16 - Intros 4:03 - Metrics for aluminum design 4:50 - How Novelis is pioneering physics-informed machine learning6:40 - How Novelis uses SigOpt8:35 – Industry is on par with academia for physics-informed machine learning 12:18 - "The optimum is not always the best choice" 14:45 - What's next from NovelisLearn more about Novelis: https://www.novelis.com Learn more about SigOpt at sigopt.com and follow us on Twitter at twitter.com/sigoptSubscribe to our YouTube channel to watch Experiment Exchange interviews https://www.youtube.com/channel/sigopt

How do you detect fraud when less than one percent of your network’s users are bad actors? In this episode, SigOpt’s Head of Engineering Michael McCourt speaks with Venkatesh Ramanathan, a Director of Data Science at PayPal, about his work using Graph Neural Networks to detect fraud across large financial networks.0:23 - Intro3:08 - AI/ML at AOL4:24 - The scale of data today 6:11 – The tradeoffs of accuracy and interpretability 7:54 - What are Graph Neural Networks? 9:18 - Robustness of GNNs; how they work with blockchain networks 10:57 - The need for robust hardware for GNNs 12:44 - How PayPal uses SigOpt for hyperparameter search 15:12 - The importance of sample efficiency 16:51 - What's next for Data Science at PayPal 20:52 - Opportunities for academia to power industry insightsLearn more about SigOpt at sigopt.com and follow us on Twitter at twitter.com/sigopt Subscribe to our YouTube channel to watch Experiment Exchange interviews: https://www.youtube.com/channel/sigopt

Machine learning holds significant promise for fields like proteomics, therapeutics, and more—but blockers like access to datasets and issues of health privacy make progress complicated. Alexander Johansen is a Ph.D. student in computer science studying the intersection between computer science, bioinformatics and digital health.Within his lab at Stanford University, Alexander has applied Natural Language Processing to proteins, explored the history of wearables data privacy, and more. In this episode, SigOpt’s Head of Engineering Michael McCourt speaks with Alexander about his pioneering work and how SigOpt has played a role in advancing progress.1:29 - Intro 2:09 - What Alexander's lab works on5:40 - Who makes up Alexander's team8:39 - Should only the FDA be involved?12:31 - NLP for proteomics 17:44 - How Alexander uses SigOpt to power his research21:24 - How Alexander collaborates with other areas of study24:28 - Picking the right health data to prioritize Follow Alexander on Twitter: https://twitter.com/AlexRoseJo Learn more about the Stanford Center for Personalized Health: https://stanford-health.github.io/ Learn more about SigOpt at sigopt.com and follow us on Twitter at twitter.com/sigopt Subscribe to our YouTube channel to watch Experiment Exchange interviews: https://www.youtube.com/channel/sigopt

Small to medium enterprises make up the majority of the companies in the world, yet they're often underserved when it comes to AI. Pablo Zegers, Co-Founder and VP of Product, is seeking to change this through Anastasia AI, which democratizes access to time series analysis for a variety of businesses. In this episode, they discuss this work, the environmental impacts of today's models, and what's next for Anastasia AI. 2:19 - The state of AI in Latin America9:14 - Why Anastasia is focusing on time series forecasting first in a modular system 11:15 - How SigOpt helps Anastasia get better results for customers 15:00 - Working with X-RNN models 18:43 - What's next for Anastasia AI this year 19:48 - Environmental impact of training AI Learn more about Anastasia AI at https://anastasia.ai Read the IEEE Spectrum Magazine article about ImageNet energy usage: https://spectrum.ieee.org/deep-learning-computational-cost Learn more about SigOpt at sigopt.com and follow us on Twitter at twitter.com/sigopt Subscribe to our YouTube channel to watch Experiment Exchange interviews: https://www.youtube.com/channel/sigopt

Our brains only use about 30-40 watts of power, yet are more powerful than neural networks which take extensive amounts of energy to run. So what can we learn from the brain to help us build better neural networks? Join Michael McCourt as he interviews Subutai Ahmad, VP of Research at Numenta, about his latest work.In this episode, they discuss sparsity, bioinspiration, and how Numenta is using SigOpt to help them build better neural networks and save on training costs.1:31 - Background on Numenta2:31 - Bioinspiration3:47 - Numenta's three research areas4:06 - What is sparsity and how does it function in the brain?7:15 - Training costs, Moore's Law, and how deep learning systems are on a different curve9:58 - Mismatch between hardware and algorithms today in deep learning11:04 - Improving energy usage and speed with sparse networks14:10 - Sparse networks work with different hyperparameter regimes than dense networks14:18 - How Numenta uses SigOpt Multimetric optimization15:48 - How Numenta uses SigOpt Multitask to constrain costs18:06 - How Numenta chose their hyperparameters19:40 - What's next from NumentaLearn more about Numenta at numenta.com and follow them on YouTube at www.youtube.com/c/NumentaTheoryRead Jeff Hawkin's book, A Thousand Brains: A New Theory of IntelligenceLearn more about SigOpt at sigopt.com and follow us on Twitter at twitter.com/sigoptSubscribe to our YouTube channel to watch Experiment Exchange interviews at www.youtube.com/channel/sigopt

Join us as Michael McCourt, SigOpt's Head of Engineering, hosts frank discussions with industry and academic leaders – where we find out how the latest developments in AI are transforming our world.Learn more about SigOpt at sigopt.com and follow us on Twitter at twitter.com/sigoptSubscribe to our YouTube channel to watch Experiment Exchange interviews: www.youtube.com/channel/sigopt