Could you tell us about your fascinating journey into R&D in AI? How did you start at harmon.ie?
I started developing AI algorithms for handwriting recognition at my part-time student job while doing my Undergraduate degree in Computer Science. Since then, over the last 20 years or so, I have strived to combine my work in the industry with academic research. I did my Graduate degree in Computer Vision and completed my Ph.D. in Machine Learning while having quite an intensive career in the industry in parallel with my studies. In the industry, I’ve worked on all kinds of data and applications, including medical imaging, educational multimedia, mobile advertising, financial time series, video, text and speech processing for public safety, and other projects. When I began working with the product and business aspects of R&D, I felt that I needed to strengthen the relevant skills, so I went back to school and got an additional Master’s degree in Technology Management.
I still teach occasionally and advise graduate students in their research. I believe that staying connected to both the academy and the industry is necessary, especially in a field as fast-changing as AI. The industry experience teaches pragmatism and execution skills, while the academic research helps to acquire skills that are necessary for systematic analysis of problems. I believe that AI leaders need both for doing industrial R&D and for conducting applied research as well.
In my previous position, I ran an Applied Research team at the innovation center of a large corporation. There, I had to cope daily with large amounts of e-mail and information related to my team, internal and external partners, and government agencies. When I heard about harmon.ie’s mission of relieving the information overload of knowledge workers, I immediately felt connected to it. I wanted to take part in the journey toward this ambitious goal. It was clear to me that the viable solution for the problem should process very diverse data with a single AI model while considering multiple products and business constraints. That’s exactly the sort of challenge that I thrive on professionally.
As a leader in the AI tech industry, what unique challenges and opportunities did you find in your journey?
Every day there is new neural network architecture or a new algorithm being published. AI development platforms are getting more mature and easier to use, and specialized hardware is becoming cheaper and more powerful. These developments generate exciting opportunities for applying state-of-the-art AI technology to solving real business needs.
However, matching between a business use case and appropriate technology is a rather challenging task. Different business applications have distinct requirements for accuracy, computing costs, or training data availability. Sometimes, because of these practical constraints, the best available AI algorithm is not always the best choice for the particular business problem at hand.
Often, achieving the adequate performance of AI algorithms in real-time settings requires multiple iterations in production. It’s always a challenge to identify a combination of valuable business use case and an underlying AI model that results in, what I call an “elastic model utility.” Ideally, an initial solution should have value, even allowing for the limited accuracy of the first-generation AI model. However, when the model accuracy improves, the value to the customers should increase proportionally.
Could you tell us about the current benchmarks in AI progress and how you train your team in managing these rather high expectations?
In the academy, people often design a common set of benchmarks for well-known problems, so different solutions to the same problem can be evaluated and compared against each other. While this is still useful in industrial settings, the competitive advantage of a company depends on a deep and unique understanding of the problem of its customers. Thus, designating accurate product-specific benchmarks is frequently the right thing to strive for while utilizing AI in industrial settings. I’m encouraging my team to focus on defining the right KPIs for a problem we are working on. I believe that sometimes, task-specific KPIs are more important than comparing ourselves to state-of-the-art benchmarks on standard data.
What message do you have for young professionals in the AI technology industry?
I would recommend investing time in planning your career strategically. Decide where you see yourself in 10 or 20 years and try to get the necessary skills. For instance, if you see yourself involved in strategic product or R&D decisions, you will need to acquire the required knowledge and start to get some experience at making these kinds of decisions. Alternatively, one can choose to specialize in applying Machine Learning for Computer Vision, for example, and enjoy this career path very much. The problem is that in a complex field like AI, it takes a lot of time to get the necessary experience that combines Research, Engineering, Product, and Business aspects. However, that is what is expected today from an AI R&D leader. Therefore, deciding early on in your career what kind of skills and experience you want to acquire is very important
How do you handle the heterogeneity of data inputs?
As AI solutions try to aggregate and analyze diverse data, the issue of heterogeneous data becomes more and more important. Luckily, there is a family of representation learning algorithms that transform/embed heterogeneous data into a uniform latent space where each data point is represented by a vector in multidimensional space. This approach has been proven essential for successfully applying AI algorithms to images, text, knowledge, and social graphs. We have chosen representation learning as a core technology of our platform at harmon.ie
What are your future predictions for Digital Transformation based on AI, Data Science, and IoT?
I think people will start to take smart behavior of things and systems around them for granted. Inability to “connect the dots” and adapt to user input would be considered a “bug” in the system, rather than a lack of smart features. That will trigger massive collection of data about every aspect of product usage (including IoT sensors) and, perhaps, the rise of Q&A Data Science teams, which will work on isolating and fixing these bugs. It would be interesting to see large Q&A teams trying to catch dumb behaviors of a product so they can be fixed before the product is released to customers.
How do you foresee issues in data management and security, especially of applications that run-in tandem with Microsoft enterprise software?
Organizations see their data as one of their main business assets. Therefore, they are very careful in their decisions about exposing their data, even to their vendors or partners. The ideal solution is to bring AI software to a secure data environment, and not try to move data out of the customer cloud. For Microsoft Office 365, this is done by deploying third-party services into a customer’s Azure tenant. This way, the customer data never leaves the security boundary of the tenant. There are some emerging technological solutions to data management and security problems, including privacy-preserving data mining algorithms and homomorphic encryption. However, there is still a long way to go before these technologies can be used in practical settings.
What do you see as the future roadmap for companies that are yet to build their own Data Science/AI teams?
There are many more opportunities around AI than are currently being explored by businesses. The first goal of in-house AI teams would be to explore and evaluate these opportunities in the context of a specific organization. Frequently, this means establishing data collection processes and mapping the essential business flows of data-driven decisions. The participation of an in-house AI team is essential at this stage, as deep knowledge of what is possible algorithmically is crucial to ask the right questions and for defining the most valuable data collection points. Once data is collected, AI teams should analyze the collected data for discovering inefficiencies or opportunities for automation. Once these opportunities are mapped, the solution can be built internally or procured from a vendor.
How do you prepare for an AI-centric world? How do you inspire your people to work with technology?
I think that every technology or business professional should be acquainted with the basic notions of Machine Learning. Luckily, there are so many online resources for learning AI-related tech in so many ways and on so many levels. I found that making the right resource available to the team with some on-the-job self-guided training provides a good starting point. I believe that with the right help, any professional can get into the basics of AI and start leveraging it in his or her work.
What start-ups are you keenly following?
I’m trying to follow startups that deal with breakthrough technologies like quantum computing, homomorphic encryption, or neuromorphic hardware.
What’s one technology that will be outdated by 2025?
Paper and pencil, and whiteboard and markers are still widely used. I believe mechanical content capture and presentation methods will be replaced by digital technologies that will be based on voice, brain-computer interface, and digital ink.
What’s something you do better than others – the secret of your success?
I’m a bit obsessed with getting to the most basic and most simple description of everything: business problems and processes as well as algorithms and technologies. I believe that uncovering the right name and definition for phenomena is a precursor for dealing with it later. It takes some practice, but it helps me a lot in my work.
Who’s the one person in the industry whose answers to these questions you would love to read?
I’ve always been interested in the intersection between technology and economics. I would love to learn from great visionaries like Ray Kurzweil and Yuval Harari.
Thank you, Sasha! That was fun and hope to see you back on AiThority soon.
Sasha Apartsin, VP of Artificial Intelligence at harmon.ie. has extensive experience working with enterprise systems and machine learning, including as Head of AI Research at Mortorola’s Tel Aviv Innovation Lab, and as AVP of Research (ML/Data Science) for Citi Innovation Lab, and working on AI projects for public safety use cases.
He joined harmon.ie because he was drawn to the idea of organizing information for the end user in an office environment to improve every minute of employees who operate in a knowledge-rich environment.
harmon.ie makes user experience tools for the digital workspace, built to deliver information in an intelligent way and unlock the full power of Microsoft Teams and Office 365. Its flagship solution breaks down data silos from Office 365 apps by grouping information using Descriptive Labels. Supported by cognitive science and powered by machine learning, harmon.ie’s Smart Assistant helps organizations bring together all their information and improve productivity. The company is a Microsoft Partner and App of the Year Finalist.