Behind the scenes with tech trailblazers: Meet Florin Manaila

How did you first get interested in AI and supercomputing?

I’ve always been fascinated by the learning curve of how we discover and adapt to new things. Even since I attended university, I wanted to come up with better answers for automating repetitive tasks (like measuring how much water there is on a crop parcel based on a satellite image). The magic of supercomputing for AI isn’t that you need to carefully design complex elements such as compute, network and storage. It’s more about how such systems are architected to support hundreds of data scientists and hundreds of experiments (natural language processing, sound, computer vision and more) running in parallel. Those AI supercomputing systems become live entities with their own rules, constraints and demands augmenting human research. This is like space exploration for me.

What aspects of cognitive systems do you specialize in?

I specialize in architecting and designing AI clusters for distributed deep learning training at scale. Among all the areas of AI, I like computer vision most. Images are universal; they’re not bound to language or culture. Therefore, the processes of learning from images can be applied everywhere. For example, you can train a deep neural network in Germany and use it in Iceland by retraining with additional images category or objects. This is called transfer learning, where a model developed for a task is reused as the starting point for a model on a second task. Even a small data set of 650 GB with 2 million files can challenge your design, but think about a 1 petabyte data set. What will happen?

What are some of the most interesting AI projects you’ve seen in your work?

Each project is very interesting, because each organization using AI today is innovating on completely new ideas. Some clients are looking for small-scale AI deployments but using existing enterprise AI technology for fast adoption, such as to identify plant diseases by training convolutional neural networks to classify leaves, to track the activity of various beetles in a lab, or to accelerate healthcare diagnostics with deep neural network models based on X-rays or CT scans.

I got to work on a large AI implementation for research with IBM Fellow John Cohn, and nearly every aspect of the original design was challenged by the users and the client team, until it evolved into something unexpected, driven by innovation and the needs of the client. When you start a new project, it can be hard to anticipate where you might end up, and that’s part of the excitement. An AI supercomputing system might need to accommodate the needs of hundreds of researchers.

I’m also interested in the energy efficiency and carbon offset of large AI clusters. If you train a generative adversarial network (GAN) model on eight IBM Power System AC922 servers (32 GPUs in total) for a month, you’ll need to offset the equivalent of 20 trees. How do you design a system at a scale that will have hundreds of users and help lower the CO2e?

How does IT infrastructure factor into the success of an AI project?

Essentially, the AI infrastructure (hardware and software) you’re using influences your research. Therefore, it’s essential to meet the following criteria:

◉ User-centric (freedom of choice)
◉ Simple and open
◉ Scalable and efficient
◉ Reliable
◉ Federated

At scale, simplicity will present an advantage, and the open source technology will drive a smoother and faster adoption among data scientists.

Can you give some examples of ways you seen AI transform business for IBM clients?

Take a client that’s using IBM Visual Insights to count bad seeds and predict the quality of its products for better pricing and positioning in the market. This project is changing perspectives on how a technology can be used to shape the future of company strategy. Another example I can think of is a client using Watson Machine Learning Community Edition on the Power AC922 for an entirely new business creating audiobooks from e-books to help users with vision impairment. Isn’t this transforming the way we were creating products and strategies in the past?

What’s your favorite part of your job?

I like working on client challenges and coming up with inventive ways to deliver AI solutions that meet their needs. Sometimes I need to go back and run experiments and benchmarks to understand the limits and come up with new ideas that will fit the original project requirements. Our ability to be successful, in my view, isn’t just about understanding the problem but experimenting with it to fully understand the constraints. I need to know what problems I am solving with my design. Only in this way can you come up with a good solution.

When you think about the future of AI for business, what do you imagine?

Today, innovation in AI is driven by hardware, followed by the software ecosystem. Nothing is possible without accelerated computing (GPUs, TPUs, FPGAs, ASICs). Even if those massive accelerated systems can prove that they can perform a specific prediction in a few seconds compared with conventional workflows, there’s still resistance from many companies in adopting AI. Time will prove that machine learning solutions really can be better than conventional tools. I imagine a future for AI where it will be a part of every business process in every company, which will open different challenges than those we face today.

Source: ibm.com