Scalable Data Science and Deep Learning with H2O.ai

Named one of Fortune's 2014 Big Data All-Stars, Arno Candel is the Chief Architect at H2O.ai, a distributed and scalable open-source machine learning platform. H2O is a leading open-source in-memory machine learning platform designed for distributed compute clusters, with a vision to democratize scalable machine learning with its open-source Java code base that integrates with everyday tools such as R, Python, Hadoop and Spark.    At the Deep Learning Summit, in San Francisco on 28-29 January, Arno will present 'Scalable Data Science and Deep Learning with H2O'. This talk will spotlight H2O Deep Learning and its ease of use and scalability on large real-world datasets and showcase its versatility across multiple applications. I spoke with Arno ahead of the event to hear more about his role at H2O.ai and what we can expect next in the deep learning field.  How did you start your work in deep learning? I started working on distributed deep learning in December 2013 when I joined H2O.ai. My first task was to improve our existing codebase to achieve competitive performance on the famous MNIST dataset. After a few weeks, I was able to obtain record low test set errors for fully-connected feed-forward neural networks by adding input dropout. H2O Deep Learning has since evolved to become more fully featured, user friendly, automated and production ready.  Tell us about your role at H2O.ai, and what you're currently working on. I’m H2O’s Chief Architect and one of the main committers to our open-source codebase. My background is in high-performance computing and physics and I focus on improving H2O’s machine learning algorithms in general. Recently, I’ve been working on topics such as cross-validation, observation weights, distribution functions, stochastic gradient boosting and novel parallelization schemes for deep learning. We are fortunate to get great product feedback from our customers and from the open-source user community.  What industries do you see deep learning having the biggest impact on? Most human endeavors are already data-informed, and many will be data-driven soon. The advent of IoT alone will lead to the running of billions of devices with some form of artificial intelligence. Whether it’s crop control systems in agriculture, personalized shopping in retail, home automation in utilities, peer-to-peer lending in finance or the emergence of self-driving cars in the automobile industry, we are about to witness significant change from the application of deep learning techniques. The impact will be enormous and it will revolutionize the way we live.  What do you feel are the leading factors enabling recent advancements in deep learning? It’s a result of the democratization of hardware, software and data. Neural network research has bloomed in recent years with easier access to computers and large labeled datasets, both of which are critical for deep learning. The availability of open-source software has given millions of people access to world-class machine learning tools in recent years.

What do you feel is essential to future progress in the field? There’s a huge gap between reading about a new method in a publication to having a robust and scalable system in production. Most organizations are overwhelmed by the pace of advance in deep learning specifically, and machine learning in general, and many are still hanging on to outdated rule-based or simple statistical models that perform far worse than the state of the art.  A barrier to the speed of adoption is the lack of production-ready tools that incorporate the latest research in a scalable and robust way. Data scientists shouldn’t be implementing and debugging algorithms, their time would be better spent gaining insights and optimizing workflows. Another barrier is the lack of interpretability of some of these incredibly accurate but complex models. Decision makers are hesitant to deploy models that are difficult or impossible to understand. More work on improving model interpretability and on reducing model complexity without sacrificing accuracy is needed.  What’s next for H2O.ai? At H2O.ai, our engineers are improving existing machine learning algorithms and adding new ones, but the main focus of our work is on improving customer experience in operationalizing scalable machine learning workflows on large datasets. So we spend a lot of effort on quality control and customer support. H2O.ai is uniquely positioned in that we’re a customer-centric company, but our product is open-source and anyone can download it.  What developments can we expect to see in deep learning in the next 5 years? I expect that deep learning techniques will become a lot more computationally efficient. Current deep learning techniques are still too power-hungry compared with the gold standard of mother Nature. It still takes too much brute force to train even simple concepts. I hope that we can reduce the training complexity and end up with a simpler set of elementary building blocks that can learn equally well or even better. This is similar to finding a unifying theory in physics such as Special Relativity or Quantum Electrodynamics that now elegantly explains previously almost impossible to express concepts. In deep learning, there are multiple pushes in this direction. For example, new recurrent neural network models can cleverly shift focus on certain areas of attention instead of looking everywhere at once. Other ideas include Hinton’s “dark knowledge” concepts on model compression and specialist networks.  Which technologies you’re excited about becoming commonplace in our daily lives? When do you think these will be available? I can’t wait for a robot that prepares a 12-course Michelin-rated dinner and serves it to me and my family while we relax at the dinner table with fully personalized massage chairs (ok, that part is easy). Perfect timing and doing the dishes afterwards would be welcome additions. Charlie Chaplin brought a glimpse of that vision to the masses in Modern Times. It’s remarkable how little we’ve achieved towards this goal in the 80 years since then, but I’m hopeful that deep learning can speed things up a bit...Arno Candel will be speaking at the RE•WORK Deep Learning Summit in San Francisco, on 28-29 January 2016. Other speakers include Andrew Ng, Baidu; Clement Farabet, Twitter; Naveen Rao, Nervana Systems; Pieter Abbeel, UC Berkeley; and Andrej Karpathy, Stanford University.The Deep Learning Summit San Francisco 2016 is now sold out! Please see our upcoming Deep Learning Summits in Boston, London, Singapore, San Francisco and New York on our events page here.