Monday, March 28, 2016

Awesome words

 A constantly updated list of words I like.

Tuesday, March 22, 2016

Teacher Relative Marking In Schools

Machine learning has brought to my attention the horribly wrong interpretations possible due to data not being on the same scale. When information is not measured by the same stick, the result is that the bigger stick dominates the inferences you draw from the data.

In schooling, standardized testing is a method with good intentions in mind. It enforces the same standards on all the people taking the test. In the context of national education however, I think there might be something that the authorities have missed.

Not every teacher is a great one. I have been fortunate enough to have been taught by good teachers. I have also had the fate of being instructed by absolute teach-shop employees. since on a national scale we would like to measure the learning ability of a child, it would make sense to check if there is anything to be learnt at all where the child is studying.

I propose that along with the students, the teachers also take board exams and so on. Then when the students get their results, their performance is weighed in relation to how good their teacher is.

A student scoring 50/100 when his/her teacher scored 30/100 is a good student. They have learnt a lot. A student scoring 50/100 when the teacher scored 90/100 is not such a good learner.

This would be a more fair scale than every student getting the same paper irrespective of the environment they studied in.

I do realize that this could have potential disasters since the student's performance no longer depends on them only. It would put enormous pressure on the teachers as well since they would effectively keep on taking exams throughout their career.

One potential fallacy is that over time teachers might get exceedingly good at taking tests and thus fail the system altogether. They will learn the common questions and so on. Since subjects are finite in nature at the school level, this is bound to happen. It might result in the teachers getting better at their subjects over time though, I cannot predict this well enough.

To conclude, I say that on the day of the exam, let the teacher sit with the students ad face the music. Then let the students be judged relative to the teacher.

Sunday, March 20, 2016

India's Ignorance Towards AI

South Korea broadcasted the AlphaGo vs Lee Sedol nationwide. What happened after the match is now being called the AlphaGo Shock. The machine won 4-1 in a breathtaking battle, often making moves previously unseen in the professional world of competitive Go.

After DeepBlue beat Kasparov in 1997, Go was the game everyone looked up to as the one thing humans were better at. This arose from the astounding complexity of the game. A chess board is an 8x8 grid, a Go board is a 19x19 grid. This, along with the rules of the games makes Go a much more complex game.

As of March 17 2016, South Korea has declared that it will introduce 1 Trillion Won into AI research over the next five years. Go has been a powerful game in the South Korean community since a long time.

The 4-1 game was followed by a spate of newspaper headlines worrying that South Korea was falling behind in a crucial growth industry.

In all this time, the Indian government is yet to publicly embrace AI. We are so blissfully ignorant that not even one of our headlines showed this news. As a Machine Learning student, I understand the implications this might have for India.

While we crown ourselves with titles like "Tech Giant" and so on, the world gleefully moves on to bigger things. We have become the telephone operators of the world while they have moved on to better things.

While the world measures Internet speed in MBps we are still on Mbps (the difference in B and b is significant in case you did not know). While the world is using computers to find significant information and insights from data, we are literally still accounting with pencil paper systems.

While this is not entirely the government's fault a large share is theirs to take. The lack of computer science jobs in the central government is a major drawback. Computer science is no longer a "helper" science used to aid other "real" sciences. It is a titan in it's own right.

The government's refusal to look at this field with an equal eye will harm us in the long run. Take Modi's God-like election victory. That is what Machine Learning does. Modi has seen the power of Machine Learning, yet refuses to acknowledge the work it could do for India.

I am almost inclined to say that like the Lords of old looking at Babbage's difference engine, all Modi can think of is how to use it. To truly understand it's potential we need someone who understands the machine. Someone like Ada of Lovelace.

If India does not put centralized effort into the field of AI, it will result in a handicap which will be impossible to overcome. All of the major data collecting companies like Google, Facebook are not India based, hence crippling the nation with respect to data collection. One might even argue that this has already happened.

The lack of government interest in research as shown by their archaic recruitment procedures for research scholars will only harm this nation.

All of this will happen not because of someone's action, rather because of inaction. Technically ignorant people have been making decisions for technology based on advisers. We require people who understand the technology.

After all is done and dusted, I expect nothing to happen. I expect that India will take pride in it's inertia as it always has and at some point in the future people like us who understand what has happened will be forced to leave this nation we love so much.

All because of the government's ignorance towards AI.

Saturday, March 5, 2016

The Machine Learning behind Voting.

Deciding who comes to power in a Democracy is a big decision. On it rests the future of the nation. The people had better not get this decision wrong. So how do we go about ensuring that our decisions are good? Voting is the answer almost everyone came up with around the world. Is there a reason it works?

Sure there is. It's called ensemble methods. What this area of Machine Learning talks about is the question "Can we have a lot of weak classifiers make a strong one?". Can we have a lot of not-so-good decision makers and somehow derive a good decision from them?

The answer to that was voting. If you have a lot of people who must vote on a certain yes/no decision, it does not matter how good they are at making that decision (as long as they are slightly better than a random decision), a good decision will be made.

Since every voter sees only a few qualities of the candidate, they make decisions based on what they see in him/her. That creates a RandomForest like situation. As we know Random Forests apply well to almost any given data set.