Skip to main content

Machine Unlearning #5 (Reinforcement Learning)

Machine Unlearning is a series broken up into tiny, one-minute readable pieces to humor our ever-shortening attention span. Sharing the links to every single piece right below:


Reinforcement Learning is a slightly different learning model than the other techniques that we have discussed previously. Therefore, I wouldn’t be able to explain the same using the fruit basket example that we have been using all this while. Let’s replace apples and oranges with self-driving cars!


Suppose that you are at Google or Tesla and are trying to train a car to drive by itself! How would you go about that? Driving requires the knowledge of much more than turning the ignition on and steering the wheel. You should know to keep the side of the road, to stop at the red signal, or to keep off the footpath, for instance.


You decide not to overwhelm the poor machine by laying out all the rules at once. Instead, you take an “on the fly” approach, by letting the car take decisions in a real-time manner, and by giving feedback on the merit of the decision taken.


If the car stops at a red light, you reward it by giving positive feedback. If the car sways off the road and runs over people sleeping on the footpath, you punish it by giving it negative feedback. The car learns by repeating the decisions that gave it positive feedback and avoiding the decisions that gave it negative feedback.


Now, for the real world story. My mother - a brown-skinned Indian - used to be teased by some of her brown-skinned relatives, for being dark-skinned. As a result, she grew up with the notion that dark-skinned was bad (negative feedback). Whenever she used to try out a bright colored outfit, they would chide her by saying that the bright colors did not compliment her complexion. The repeated feedback reinforced this belief in her mind, and she resorted to wearing pale-colored clothes during the colorful years of her youth. It took her years of unlearning to realize that dark colors did in fact suited her well.


Not just her relatives, but the teachers at her school, who are supposed to be implanting progressive thoughts in young minds, were guilty of reinforcing regressive ideas. She used to recall accounts of the nuns who ran the Convent School she was a student of slut-shaming students for coming to school with a bright bindi on their forehead. Sweet Jesus!


Comments

Popular posts from this blog

Machine Unlearning #0 (Intro)

You might be familiar with the term Machine Learning. Worry not if you have not, cause I have tried to give a gist of the concept here. The term has been in the limelight of late and has been tossed around rather liberally to denote anything related to artificial intelligence, robotics, and data mining. Machine Learning, as the name suggests, could simply mean the field of study of enabling the “machines” (computers) to “learn” from past experiences and make informed decisions in the future.   Wait a minute! Learning from past experiences is something humans do, right? Exactly! The computer folks want computers to behave more and more like us. As if there aren't enough of us already. As the machines are becoming more like us, we are becoming more like them. Introspection time! Most of us wake up every morning like clockwork! Then we rush through the morning routines - get dressed, wade through the traffic, and reach our offices or schools or wherever people expect us to be. We spe

The High State

 Before The Judgement I believe I must begin by addressing the pressing question - Was planning a vacation in the midst of a pandemic a recommended move?  No. Yet we went ahead with it. Here is why.  We (Nithya & I) were newly married, and our vividly planned vacation at the island of Langkawi was stolen away from us by the virus. Our stay in Delhi was coming to an end due to job-related moves, and we felt it would be a waste not to utilize this opportunity in exploring at least one of the tourist hot spots easily accessible from the national capital region. Let us end this section by answering another question - Are the reasons listed above good enough to risk a vacation during a pandemic? No. We had taken a calculated risk. Arrival at Manali There are two phases to this - planning and execution. We had not started planning with Manali in mind. There were numerous choices - starting from Jaipur and Amritsar to Nainital, Shimla, and Manali. After a bit of reading and deliberations,

Machine Unlearning #1 (Classification)

You can’t conclude a discussion on Machine Learning without mentioning classification. Classification is a machine learning technique where the machine is trained to predict the label of the given input data. Alright, let’s cut the jargon and get some real-world examples. Oranges and Bananas. Let’s assume that we have a box of fruits that contain some oranges and some bananas. You are asked to pick one fruit at random and tell if it is an orange or a banana. Pretty basic, right? For us, it is straightforward. We would know the answer at first sight. But, how would a computer be able to tell the difference? In classification, the machine would first be trained on some pre-labeled data. It would be shown an orange and we would tell it that the fruit is an orange. The machine would study the orange and remember its features - orange color and round shape. Then it would be shown a banana and the process is repeated. What are these features? A feature is anything that helps us uniquely labe