Author: Ammar A. Raja
1. SHORT BIO OF THE AUTHOR
Dr. Ammar A. Raja is an assistant professor at COMSATS Institute of Information Technology, Pakistan. He received his PhD degree in Finance from The London School of Economics and Political Science (LSE) in 2012. Apart from conducting research in data analytics, he also publishes blog posts regularly on Data Science Central and considers himself as a “disruptive” data scientist.
2. SUMMARY OF THE BLOG POST
Science fiction indulges us in the fascination of creating humanoids, like the grotesque creature from Mary Shelley’s novel, Frankenstein. And for a long period of time, building a robot seemed to be a faraway dream.
The advancement in modern technology has opened up the door for robotics: scientists have managed to build robots that look and move like us. But those robots are not humans after all. Thus, how do we make a robot that goes beyond mere resemblance, and can think like a human and make its own decisions?
In Dr. Ammar A. Raja‘s blog post “How Data Science Apply to Robotics” on Data Science Central, he pointed out that there are two major problems scientists have encountered while applying data science to robotics.
Problem 1: How to program a robot without defining all the moves it is going to make?
Raja stated that it might sound like an intuitive idea to just program every move a robot is going to make. But it is also challenging to predict every action that a robot is bound to take. Even if we could, when there is a new functionality, do we reprogram the robot? This makes the brutal programming-every-move strategy an inherently flawed one.
Problem 2: How to reduce the computational complexity in real-time vision tasks?
Vision is a crucial sense for humans, and it is also an important input for robots. Scientists are endeavoring to equip robots with real-time vision to cope with various tasks. But robots do not see the world as humans do: as far as a robot is concerned, the world is merely millions of zeros and ones. However, real-time vision tasks usually involves a constant change in views for the robot. For example, in a task where a robot has to capture the movement of a human hand, every position change of the hand introduces a different set of zeros and ones, which leads to high computational complexity.
How can machine learning help?
Raja referred to machine learning as “the poster boy of data science” that might help solve problems in robotics. Using machine learning, only a few labelled data is needed to teach the robot a behavior.
For instance, one prominent example of machine learning is handwriting recognition, where labeled data examples are fed into the computers, enabling them to learn both the positive examples (i.e. the letter “a”), along with the negative ones (this is the “training” phase). The computers will then be given new data, and we would evaluate the accuracy of the recognition (“the test phase”). A satisfactory accuracy would be an indication of a qualified classifier for the handwriting recognition task. The accuracy of autonomous handwriting recognition by computers has dramatically improved in the past decade, thanks to the availability of data for the purpose of machine learning, as well as the increasing computational power that equips us with the ability of coping with computation-intensive vision tasks, which comprises a good portion of the challenges in robotics.
Moreover, Raja considered reinforcement learning, a branch of machine learning, as “the closest that machine learning can get to the way how humans learn”. He explained the concept of reinforcement learning in laymen’s terms: reinforcement learning is concerned with how computers (or agents) ought to take actions in an environment to maximize the amount of awards. And the actions that computers take will either bring rewards or penalties. Therefore, computers could learn from penalties, and remember what actions should be taken in order to receive awards in certain scenarios.
There are already a considerable number of real-world applications of reinforcement learning, such as recommendation systems where items could be recommended to customers based on their purchase behaviors, etc. Raja’s personal favorite, as he wrote, would be teaching the computer to play Super Mario Bros using reinforcement learning.
Machine learning is not only helping machines learn about the world that we live in, but also allowing us to unravel our fascinations that we used to think would only exist in science fiction.
This a blog post is written by an enthusiastic data scientist, Dr. Ammar A. Raja, who has a background in finance. It gives a short overview of how data science, machine learning in particular, can be applied to robotics. He adopts the example of handwriting recognition to demonstrate how the successful application of machine learning can help tackle vision challenges in robotics. And he is optimistic about the future of machine learning, which will continue to help unravel humans’ fascination that used to only exist in science fiction.
Review by: Olli Huang | Editor: Joni Chung