Walking on the thin intersectional lines of disciplines.

Elishai Ezra Tsur, a multidisciplinary researcher, talks about the challenges that conventional academic mindset brought to his professional life. He, DeWolf, and Supic introduce us with their viewpoint about "data science" and its role in their research. In their recent work published in this issue of Patterns, they tackle the inverse kinematics problem using brain-inspired neuronal architectures.

Main text

Tell us about your paper in this issue of Patterns

Elishai Ezra Tsur: Our work uniquely addresses one of the most fundamental challenges in robotics: inverse kinematics in convoluted environments. Inverse kinematics is a computational process for deriving a robot’s configuration (orientation of each of its joints), given its desired target position in space. Inverse kinematics is often underdetermined, involving redundancy resolution, where one solution is chosen from an infinite set. Underdetermined inverse problems are significant concerns in many other fields as well, ranging from medical imaging to hydrology and pharmacokinetics. This work represents a cross-disciplinary effort, integrating insights from neuroscience, computer science, robotics, and control theory. With this work we celebrate abstraction, transdisciplinarity, and collaboration among the academic, startup, and industrial worlds.

Tell us about your background (personal and/or professional)

EET: I founded the Neuro-Biomorphic Engineering Lab (NBEL) following 10 years of traversing numerous disciplines: biology, philosophy, history, neuropsychology, bioengineering, computer science, and neuroscience. My research endeavors often involve different schools of thoughts. I’ve used algorithms from computer graphics to elucidate flow patterns embedded within fluid dynamics, used artificial intelligence to design mechanical systems, and used cloud computing to create federated databases on vascular diseases. In NBEL, we focus on brain-inspired (neuromorphic) engineering. We use neuronal architectures to elucidate mechanisms of biological visual systems, to control autonomous vehicles, to manipulate robotic arms, and even to track black bears’ hibernation physiology during the winter.

Travis DeWolf: I received my PhD in systems design engineering, focusing on computational neuroscience from the University of Waterloo, where I was a member of the Computational Neuroscience Research Group. I study the motor control system, working to uncover the planning, execution, and adaptation mechanisms that drive movement. I am the co-founder of Applied Brain Research (ABR), where I lead the autonomous systems group, which focuses on developing neurorobotic systems using spiking neural networks and neuromorphic hardware.

Lazar Supic: I received my PhD at the University of California Berkeley College of Engineering. At UC Berkeley, I was also a postdoctoral researcher at the Berkeley Deep Drive (BDD) institute and the Berkeley Wireless Research Center (BWRC), doing research on deep neural networks (DNNs) optimization for energy-efficient, low latency edge computing. I am currently a research scientist at Accenture Labs in San Francisco, where I lead several projects related to neuromorphic computing and robotics, including control algorithms for neurorobotics and event-based visual odometry.

What motivated you to become a data researcher? Is there anyone/anything in particular that helped guide you on your path?

EET: Over the years, I found that data liberates my models from full mechanical and mathematical descriptions. For example, engineers usually use numerical or analytical approaches to solve inverse kinematics. These approaches are generally subjected to complete mechanical descriptions and known environments. However, a data-driven neural network may provide a unified framework for inverse kinematics, supporting robotic systems with arbitrary complexity, operating in arbitrarily convoluted environments, subject to any constraints. Data-driven approaches alleviate the requirement for detailed descriptions, emphasizing the uniqueness of brain-inspired computing: adaptively learning to navigate the changing world around us. A key enabler for me was the highly crafted abstraction layers provided by open-source frameworks like TensorFlow and Nengo. The second enabler is the high level of training and the independent thinking and resourcefulness I strive to cultivate in my graduate students. With the availability of computing resources, the only limit for data-driven science is creativity and imagination.

TD: Better understanding the mechanisms that underlie the brain’s functionality has been a primary motivation throughout my career. Both in hopes of being able to more effectively address injuries and sickness in the brain and improve our ability to develop efficient, adaptable engineering solutions. In current technology, there are artificial systems that can outperform humans on some specific tasks. But these systems are highly engineered for their application, and the flexibility and robustness of the brain as a general solutions machine remains far beyond anything we can build. There is something in the balance of the brain’s highly stereotyped larger-scale architecture and highly plastic smaller-scale structures that has proven incredibly effective, and working toward better understanding it is incredibly exciting and a constant source of inspiration

LS: I remember designing a speech recognition algorithm for mobile robot control during my master’s program in electrical engineering at the University of Belgrade, Serbia. I was amazed by the fact that I could infer the identity of a person from audio data. Each person has a unique vocal tract, a set of vocal frequencies that are singularly unique to them, which is reflected in the data. I learned firsthand that data could give us unique insights about us or the world around us and that, often, uncovering deeply hidden patterns is only possible through data. For example, during my PhD program at UC Berkeley, I used data processing techniques, such as singular value decomposition, that enabled me to effectively look inside a radiation detector and determine where a gamma ray interacted with the sensor to create a gamma ray image of the environment. That was also the first time I really faced the challenge of processing enormous amounts of data and experienced firsthand the power of new programming languages, such as Python, which effectively revolutionized experimental research on extensive datasets. During my postdoctoral research on DNN optimization and compression, I developed data processing and computation tools for training state-of-the-art DNNs on enormous datasets (e.g., ImageNet) in a reasonable amount of time. So, I would say that the emergence of these new powerful data tools was tremendously helpful to me on my path.

What is the definition of data science in your opinion? What is a data scientist? Do you self-identify as one?

EET: Data scientists transform data into knowledge. I do not identify myself as a data scientist but rather an engineer striving to build brain-inspired machines. Brains use data to create a useful mental world through which their emerged consciousness can reside. Data science is essential; it does, however, constitute parts of the whole.

TD: I would characterize a data scientist as an interdisciplinary specialist who is able to apply the scientific method and a deep understanding of statistics to uncover insights hidden in data. The larger and less complete the dataset, the more difficult this task becomes. In developing computational models of the brain, being able to process the enormous amount of information from experimental neuroscience and build, compare, and contrast different models requires data scientist skills. My work now often has an engineering focus, taking insights we’ve gained and applying them to new problems, but still often requires putting on the data scientist hat.

LS: I would say that a data scientist is someone who uncovers hidden patterns and insights about us and the world around us from data. I am not a data scientist by training, but I would argue that everyone who does research is necessarily a bit of a data scientist. Designing an effective experiment and then collecting, cleaning, processing, and interpreting the data is critically important for experimental lab work. Each step in the data processing pipeline, comprising data collection, preprocessing, preparation, and data representation, requires meticulous effort to obtain meaningful results. Even in theoretical research, which often includes simulations, the scientist must prepare, process, and interpret data from simulations. In that sense, in this data-driven age, every scientist is partly a data scientist.

Why did you decide to publish in Patterns?

EET: I think that Patterns fills a significant gap in the data-intensive-focused publication space. It uniquely provides a high-quality venue, rigorous review, and the high visibility associated with Cell Press.

Tell us about any barriers you faced in pursuing data science as a career

EET: I always wanted to lead independent research in a university and embed a trail-blazing mentality and creative thinking within bright young minds. However, admitting into the highly disciplinary, conservative, and competitive world of academia was a significant challenge, as my CV did not reflect one area of expertise. In the past decade, I published my work in journals focusing on toxicology, mechanical engineering, fluid mechanics, electronics, biomedical devices, brain sciences, data science, computer-aided design, education, and computer vision. On top of my diverse academic training, this eclectic body of works turned out to be a high barrier to the conventional academic mindset. As a matter of fact, the only reason I ended up with an academic appointment was the courage of one woman that recognized the potential. She assembled a professional committee, uniquely comprising experts from different disciplines to make an assessment. Data-driven research thrives on multidisciplinary thinking, and it takes unorthodox academic structures to acknowledge it.

How do you keep up to date with both advances in data science techniques and the advances in the field/domain that you work in?

EET: Keeping up to date with the field’s advancements is crucial to running an innovative research group. To stay updated, I participate in at least four conferences per year, I do a lot of peer reviews of my colleagues’ works, I edit special issues in various venues, I write papers and grants and carefully address the reviewer’s concerns, I use Google Scholar notification system to know when someone cites my work or when a colleague published a new manuscript, and yes, I also allow Google Scholar to recommend recent interesting articles every weekend.

TD: Part of my work at ABR is to stay current with the most effective methods and advances in the field, to make sure that we’re able to continue achieving cutting-edge results. We have time set aside each week specifically for investigating the latest state of the field, and we help keep each other informed. I’m also fortunate to collaborate with leading researchers, such as Dr. Ezra Tsur, who provide unique and valuable perspectives and constantly push the field forward.

Which of the current trends in data science seem the most interesting to you? In your opinion, what are the most pressing questions for the data science community?

EET: I have always been inspired by the engineering of moving devices. Writing code that breaches the boundaries of the virtual-computer world has a unique creative value. One important thing that organisms do better than any other device is to move. We move efficiently and adaptively without holding accurate numerical models. Instead of using exact pre-trained models, we use data to learn and adapt in real-time to the physics of our environment, allowing accurate, modular, and versatile motion planning. I would argue that this type of learning—physics-aware learning in real-time—is the next big step forward in data-driven robotics.

LS: The most exciting trend to me is the advances and open questions in data fusion, which seeks to combine data from different sources. In my opinion, getting good quality data from new hardware modalities and creating algorithms that enable real-time processing are key challenges. I believe that neuromorphic hardware and spiking neural networks can play an essential role in addressing these challenges in the future.