AiThority Interview Series With Raj Minhas, VP, Director of Interaction and Analytics Laboratory at PARC, a XEROX Company

Know My Company

Tell us about your journey into technology and how you arrived at PARC?

Even as a young kid, I was always interested in technology and cool new toys enabled by them. For my eighth birthday, I had received a radio-controlled model airplane that I was completely fascinated with – my father told me that if I wanted to build things like that, then I would need to become an engineer. And that started my journey. After a Ph.D. from the University of Toronto in control system design, I joined Xerox Research Center Webster in upstate New York and have been with Xerox Research ever since. After Webster, I joined Xerox Research Center India and led its growth from 2011-2013. In 2013, I came to PARC and currently manage the AI Research lab.

What is PARC and how does it fit into the world of Data Science and Human-Machine collaboration?

PARC (i.e. Palo Alto Research Center) is a fully owned subsidiary of Xerox Corporation and operates in an open-innovation ecosystem. We are the innovation engine for Xerox and develop new technology options, enable new service offerings, and incubate new business opportunities for Xerox. In addition, we also work closely with numerous commercial clients internationally including many Fortune 500 companies and startups.

We help create design and technology innovations that drive many of their new products and services. Finally, we do a lot of research on behalf of several US federal and state agencies (e.g. Defense Advanced Research Projects Agency). We do research in a wide range of areas and the diversity of our researcher is astonishing given the relatively small size of our organization.

We focus on six overlapping areas:

1. Microsystems and Smart Devices
2. Digital Design and Manufacturing
3. IoT and Machine Intelligence
4. Novel Printing
5. Digital Workplace
6. AI and Human-Machine Collaboration

Our people have backgrounds in energy systems, sensor design, material deposition, polymers & composite materials, printed electronics, thin-film electronics, optoelectronics, user experience design, cyber-physical system design, condition-based maintenance, system security, computational geometry, conversation agents, computer vision, model-based reasoning, and interactive machine learning. Most of our work is at the intersection of multiple such competences.

We have been involved with AI research from the very beginning – in fact, one of our Research Fellows, Danny Bobrow, wrote the very first Ph.D. thesis in AI under the supervision of Prof. Marvin Minsky at MIT. We continue to push the frontiers of AI with cutting-edge research in Explainable AI, Computational Design, Hybrid Model-Based, and Data-Driven modeling.

We are trying to develop the foundations of Human-Machine collaboration by bringing together key concepts such as explainability, physical groundedness, and cognitive modeling. We want to build AI systems that can be easily trained by the people who will use them. We want these systems to be partners in design and exploration.

As a tech leader, how do you prepare for an AI-driven world?

Alan Kay, a distinguished computer scientist, and PARC alumnus, once said that the best way to predict the future is to invent it. That ethos is still very much alive at PARC.

We are preparing for the AI-driven world by helping create it. From very early stage research to novel applications in numerous domains, we are involved in all stages of the innovation lifecycle. We are not trying to do this alone but, as mentioned earlier, we are denizens of the open innovation ecosystem. We partner with top academics and commercial partners around the world to develop innovative technologies. We work with Fortune 500 companies and startups to transition those technologies to the world.

What message/advice do you have for young professionals looking to build a career in AI/ML?

Jump in – there are lots to do and it’s easier than ever to get started. There is a lot of easily accessible learning material and code available on the web – there many online courses, blogs, and podcasts that can help you understand the state of the art as well as open problems. Read widely and don’t focus just on the computer science literature. Machine learning currently focuses on statistical models that are learned empirically from data alone. I strongly believe that the path forward for AI/ML is by incorporating more direct knowledge and constraints about the world. So, I would recommend learning about dynamical system design and modeling.

What is the state of ‘AI for Businesses’ in 2018-2020? How can business development teams’ better benefit from leveraging PARC as a platform?

As William Gibson said, the future is here but it is unevenly distributed. There are many companies that are natively digital and data-driven – they are able to adopt and utilize AI/ML in a natural way. But the vast majority of large and small businesses are just starting that journey. They realize the vast potential that can accrue to them by the use of AI and data-driven processes. We have a lot of clients who fall in this category and who are in various stages of that journey.

We help them in various ways. First and foremost, we strive to be a trusted advisor to them. Sometimes they simply need our help to understand the landscape and make an informed choice between options they are already considering. Most of the times, we help them identify business processes that can benefit from AI, build the financial model, envision the new product or service, and then create a proof of concept. PARC has a very multi-disciplinary workforce and we are able to bring together the right team for any given task.

Tell us more about your AI research programs and the most outstanding digital campaign at PARC.

At a high level, the AI research at PARC has two themes:

1. Hybrid AI combining model-based and data-driven capabilities, and
2. AI for the design of systems

Physical model-based techniques are based on first principles and, therefore, can learn from very little data and are more robust to extrapolations to previously unseen data regimes. But building accurate models based on first principles models is very time-consuming. On the other hand, accurate data-driven techniques are easier to build as long as a lot of relevant data is available. But for a number of safety-critical cyber-physical systems, very little data is available under failure conditions because they are (thankfully) designed to be robust. Hybrid AI combines the best of both worlds – it allows normal behavior data to be used to build and validate physical models which can then be automatically augmented to include failure modes (based on an understanding of first principles).

Similarly, we can combine physical information (e.g. the structure of road networks, reflection spectrum properties of biological cells and tissues) when performing computer vision tasks. This results in AI systems that are more robust and require less training data compared to the state of the art.

Design of a physical system typically requires exploration of a large design space. For example, there are many ways in which a block of metal can be turned into a desired machine part. AI utilizing a knowledge of computational geometry can help find with the efficient exploration of this design space.

The size and complexity of the design space increase exponentially when we move from the traditional subtractive manufacturing to the new wave of additive (i.e. 3D) manufacturing. The current CAD/CAM/CAE tools are incapable of handling these complexities so we are building the next generation design tools for additive and hybrid additive-subtractive manufacturing.

This is also an area where we are putting a lot of effort into human-machine collaboration. In the current paradigm, the human designer comes up with a design and the computer tools test properties such as feasibility, manufacturability, durability etc. We are trying to build systems where the AI can do a lot more and be a partner in the exploration of the design space. For example, instead of simply testing properties of a given design, the system should be able to generate candidate designs based on high-level requirements (ala Jarvis from Iron Man).

Philosophy in AI/Machine Learning 

Can AI systems learn from people? Can people learn from AI?

Absolutely. AI systems learn from people already albeit in simplistic ways (e.g. labeled data, reinforcing feedback etc). Researchers around the world are working on ways to make this easier and more ongoing. The converse is also true – people learn from AIs especially when it comes to areas that rely on the exploration of large state spaces. The evolution of chess is a good example – the style of play has changed dramatically since the advent of chess AI programs.

What are the major challenges for AI technology companies in making it more accessible to local communities? How do you overcome these challenges?

I think the AI needs to be transparent, predictable, and intuitive. Nobody thinks much about the AI algorithms running inside a washing machine, a Roomba vacuum cleaner, a GPS Pathfinder, a Nest thermostat, or a camera’s exposure metering system. In all these cases, the AI technology is invisible to the user – it performs the task in a predictable and intuitive way. These examples may seem trivial now but they represented the state of the art not too long ago. In fact, this is a well-observed phenomenon called the AI Effect, whereby the role of AI is discounted when it becomes commonplace. As the current state of the art in AI becomes more predictable, intuitive, and mundane, it will also just become a part of the fabric of life.

On what foundation/principles should businesses and society chase the future with AI and machine intelligence?

Any powerful technology has the potential for being misused and AI is no different. In my opinion, the upsides far outweigh the potential downsides. So the principles for the pursuit of AI/ML should be no different than the principles for the development for other impactful technologies: international participatory decision making that determines what applications of AI are acceptable to the global human community. The humankind has done this successfully in a myriad of other domains and there is no reason why we can’t do it for AI applications.

Tell us “The Good, The Bad, and The Ugly” of AI/ML technologies.

To me, the good and the ugly are the two sides of the same coin. AI has the potential to, and is already starting to, have a huge impact on all aspects our lives: society, democracy, health, wellness, agriculture, mobility, finances, environment, exploration, etc. Just like digital computing and global networking, the impact of AI will be global and ubiquitous.

And just like those technologies, the good will far outweigh the bad and the ugly. But it won’t happen on its own. As AI researchers and global citizens, we need to develop and demand capabilities that make AI more explainable, transparent, unbiased, intuitive, and manageable.

Would you agree that ‘Weaponization of AI” is a real danger? How steps do you take as a tech provider in preventing the misuse?

Weaponization of any potent technology is a danger – AI is no different. We are doing research to make AI models be more transparent and explainable so that the inherent assumptions and biases, as well as the outcomes, can be readily examined. Currently, the complexity of the AI/ML models renders them as black boxes – data goes in and decisions come out. Either because of legislation (e.g. GDPR’s right to an explanation) or due to consumer demand, it will become necessary to open but this black box and lay the contents bare. We are playing our part in making that happen. 

The Crystal Gaze 

What AI start-ups and labs are you keenly following?

DeepMind Health. If they succeed in their mission to reduce the time for patients between test and treatment, it can have a huge societal impact and be a model for other countries.

What technologies within AI and computing are you interested in?

In addition to the areas already being researched at PARC, I am very interested in machine learning algorithms for quantum computing.

As a tech leader, what industries you think would be fastest in adopting AI/ML with smooth efficiency? What are the new emerging markets for AI technology markets?

AI is transforming many industries (e.g. healthcare, transportation, quantified self) and will yield big impacts. But there are a number of things we still need to solve (e.g. robustness, transparency, unbiasedness) before there can be “smooth efficiency” in areas like autonomy, diagnosis, personalized medicine. So the smoothest adoption will likely be in areas where the business impact is high and the risk to life and limb is minimal: Industry 4.0, mining, high-speed trading etc. I think every industry will be impacted eventually.

What’s your smartest work-related shortcut or productivity hack?

Exercise before starting the workday.

Tag the one person in the industry whose answers to these questions you would love to read:

Prof. Michael Jordan, UC Berkeley

Thank you, Raj! That was fun and hope to see you back on AiThority soon.