How to Stabilize the Control System of an Autonomous Vehicle
What is your background? Can you please tell us a little bit about your career path.
Well, my undergraduate degree is in electrical engineering from the Indian Institute of Technology, Roorkee. During those years, I was mostly working on various projects ranging from control, signals, power systems, etc. But I was mostly fascinated by control theory given the fact that it could be applied to robotics or any other dynamic system. I decided to pursue that field as a postgraduate at Purdue University. However, the major changing factor came when my advisor suggested to take some machine learning courses. So, here I am now, working on transfer learning, a key sub-field of machine learning that tries to make it label-efficient. The interesting thing is that control theory and machine learning has so much in common that it was not difficult for me to migrate. Both these fields have feedback, optimization, and input/output/ states as part of the framework. In fact, the famous Bellman equation is used both in optimal control and reinforcement learning.
What is your current specialization?
My current specialization is transfer learning, which studies methods on how to transfer and adapt existing models from previous machine-learning tasks to new tasks. I am working specifically on two problems related to transfer learning, i.e. domain adaptation and small sample learning. Domain adaptation tries to minimize data distribution discrepancies that may exist between training and testing conditions. Small sample learning allows for the learning of new tasks from only a few labelled data.
My work has an impact on any industry that uses AI or machine learning to produce decision models. A learning system has data collection and data annotation as part of its pipeline. This stage of data processing is very cumbersome and takes up a major chunk of the resources and time. Transfer Learning can reduce this data processing time by requiring to label only a few data samples or just use unlabeled data.
What are the most important problems your customers are met with?
In the machine learning community, the most important problem that is preventing models from large-scale deployment in the real world is the issue of interpretability. Current deep neural network-based models are high performing and produce close to human levels of performance, but they are not interpretable. More specifically, we are not clear as to why the neural network model is producing an output given a certain input. It is mainly because of the black-box nature of neural network models where each neural connection does not have a semantic meaning. This problem is certainly important for AI decision models in healthcare where wrong diagnoses can lead to fatality. Similarly, black-box perception systems in autonomous vehicles, when uninterpretable, can lead to accidents mainly because it cannot provide correct decisions, especially for edge-case data samples.
One obvious solution to the interpretability problem is the use of simpler models, such as decision trees, linear models, etc. However, the recognition performance of these models are poor compared to deep-learning models. Deep-learning models are less interpretable but better performing compared to these simpler models. The goal is to have something in between that is both interpretable and well performing enough. As a matter of fact, new design methodologies for deep networks could be explored that constrain the neural connections to explain a part of the task of interest.
What is the most remarkable experience you’ve had during your professional career in this field?
Most difficult problems have the simplest solution. That’s what I learned from this remarkable experience. This was a team-based project for an FSAE Motorsports competition that I worked on during my undergraduate studies. We were supposed to build a hybrid motorsports vehicle for the competition. Me and a group of teammates were in charge of the control system for the vehicle. The control system had current input from the batteries and torque output to the wheels. Somehow we could not cause the torque outputs to stabilize. The torque would overshoot and would never settle to the baseline.
Weeks passed and we could not find a solution. We tried to change the magnitude of the current input, but it would not stabilize the torque. After some brainstorming, we tried to attack the problem from the first principles. Our control system was a PID controller, the parameters of which decide the stability. These parameters are related to the resistances, capacitances, and inductances used in the circuit and we had initially calculated the parameters and found out the system to be stable. Upon further inspection we found out that the wire used for the resistance had actually been coiled and it was producing an additional inductance. This caused a change in the parameters of the PID controller and therefore caused the instability. After that, we put the wire into a rigid pipe so that it would not get coiled, and consequently there was no stability in the torque output. This is a perfect example of a simple solution to a frustrating problem.
What are the sources of knowledge you use to improve your skills?
Since I work in an academic environment, research papers are my ultimate source of knowledge. However, when some concepts or ideas become incomprehensible, I refer to video lectures and blogs. Having said that, there were some important courses that have shaped my career direction. They are “Machine Learning” by Andrew Ng, “Convex Optimization” by Stephen Boyd, and a mathematical statistics course that I took at Purdue. I generally read Internet resources and not as many books, but I would definitely recommend “Learning from Data” by Yaser Mostafa and “Deep Learning” by Goodfellow et al. These books are a treasure for anyone trying to break into machine learning and AI. As for the future, I would like to study financial models, since I have zero knowledge of finance. I am especially interested in how I can make my transfer learning skills useful in the finance domain.
Questions asked by Alex Savinkin
Former number cruncher in investment funds & strategy consulting. One of Geekforge Founding Fathers. Blockchain and technical singularity true believer.
