Thursday, November 26, 2015

They are gone

I thought there would be music,
I was wrong,
Colours would become vivid,
Noise to birdsong.

My world should have changed,
I was wrong,
Like every other friend,
Like every other bond.

I should have noticed,
I was wrong,
You won't see your love,
Until they are gone.

Monday, November 23, 2015

My all

You are my light
my solace, my balm,
my peace, my calm
in this raging storm.

You are my shelter,
my cradle, my hope,
my thread of existence,
my only rope.

You are everything
to me, my love.

Saturday, November 14, 2015

What Educators Need to Learn From Machine Learning.

Machine learning. For someone who dreams of being an educator, this was one field which I did not expect to teach me anything with respect to Education. Never have I been so wrong. The very name should have told me something.

Machine Learning is centered around the idea that you can teach machines how to learn from the data it receives. The difference between teaching humans and machines is that for the machine you have to spell it out to completely. A human child already knows how to learn. A teacher's goal is to enhance that ability. What I present here are some things I learnt from studying Machine Learning.

There are two types of training in Machine Learning(ML). The first is supervised and the second is unsupervised. During supervised training, the computer is given input and shown expected output. Essentially what we are saying is "When you see this do this." The computer then learns on it's own to relate the input to the output. This method is fast, efficient and generally the road people take.

The other type of training is unsupervised. You give the computer some data and ask it "What do you see in this data?" generally speaking. This technique might not teach the AI what you intended but it allows it to find patterns that you yourself might have missed.

What educators must take away from this field is the fact that there are a lot of errors when teaching an AI anything. There is the problem of over training, causing the AI to perfectly know the training data but perform badly on new data which is previously unknown. This is akin to the habit of rote learning among students. While it allows you to perform incredibly well on the data you have been trained on, the performance drops drastically once new data is introduced.

Another facet is that AI are generally geared to do different tasks. If a Feed Forward Neural Network is asked to model a  time series it will fail. Not every student is the same. That does not mean that the current education methods count for nothing. What it means is that the educator needs to be especially perceptive of what the student is best suited for. Especially developing that need and letting the student develop that very skill is what an educator must do.

Coming back to rot learning, let me tell you about neural networks. Neural networks were inspired by the nervous system of humans. The brain is essentially a mass of neural networks. This lets us come to a conclusion that can help us teach better. The only way neural networks can learn anything new is by recalling that information again and again. However they are also incredibly good at building shortcuts. Recalling something the same way again and again slowly loses it's effect.

The challenge for the educator is to allow the student to recall the same information in different ways. One might take the example of the number PI. PI is the ratio of a circle's circumference to it's diameter. It is approximated as 22/7. A better approximation of PI is 355/113. Since PI is irrational it contains every possible number within it's never ending and never repeating fractional part.

With that we have created four different ways for the student to access the number PI. This increases the student's actual recall of the number PI itself. The human memory is unique in the fact that the more things you remember the more you can further remember. Let me illustrate. Your task is to remember this next drawing.

With the following bits of information you will know more about this character, thus giving you more ways to recollect it. Ultimately your memory of this character will be impeccable.
  • It is a mandarin character
  • It signifies the sun
  • Originally it did look like a sun.
  • It also signifies a single day 
Now that we know these bits of information we will remember this character better. Another bit of information is the evolution of this character as shown below.
Thus we can see that the more a student learns about a subject, the more they will be able to learn in the future.

With that I conclude this half-rant, half-wish of mine in the hope that I can one day be a good educator.

Wednesday, November 11, 2015

Statistcis of the 2014 General Election

The 2014 General Elections were very popular and saw the introduction of many new faces onto the political platform. Some interesting statistics from those elections and their results are shown here which might lead us to some interesting conclusions. To begin with, let us define our data set. The links used to provide data were:
With these links we had with us the election results of all candidates in the 2014 GE. Along with that we also had with us the number of criminal cases against approximately 1400 candidates.

Using DATA1 and DATA2 we create a list of candidates present in both the data sets. Thus we have ~1000 candidates. From this list we only use 3 pieces of information.
  1. Candidate name
  2. No of criminal cases against them
  3. Number of votes they got
With that we obtain some interesting figures:
  • The minimum vote anyone received was 1
  • The maximum anyone received was 758482
  • The maximum number of criminal cases was 382 against Uday Kumar SP
  • The second highest criminal count was against Sridip Bhattacharya at 57
  • Minimum number of criminal cases were 1
  • Correlation factor between criminal cases and votes was 0.032
Some things of note are:
  • The correlation is weak. Thus it is ignored. In order for it to be relevant, it must be at least greater than 0.1
  • Real criminals are still contesting for public office
For the time being the 2014 GE show nothing interesting statistically for me to discover at my level. I will consider it interesting the day the covariance crosses 0.1

Comments and similar efforts are welcome.

Thursday, November 5, 2015

Secure Distribution of Examination Question papers

During various examinations the secure distribution of Question Papers is an important aspect of the entire examination. In case the paper is compromised the entire examination must be scraped and redone. This introduces a lot of cost to the examination body.

Here we attempt to develop a mechanism to securely deliver question papers for an examination on the day of the examination. Our goals encompass:
  • Secure delivery of the document to exam centers
  • Robustness of system
  • Cost effectiveness
  • Scalability
In order to facilitate this system we will use already existing technology and infrastructure, namely Internet and mobile penetration of India. We will adopt the use of technologies such as websites, public key encryption and hashing.

Common terms used here are:
  • Host: The body/institution responsible for the conduction of the examination.
  • Exam: The examination under discussion.
  • Center: A place where candidates may take the examination. May be an institution/independent body
  • Paper: The examination question paper

The resources needed to complete this objective are:
  • A website controlled by the Host.
  • Printers at the Center
  • Proper backups for printing are necessary in case primary printers fail.
The method employed is:
  • The Host encrypts the paper using PrivateKey1.
  • The resulting document is again encrypted using PrivateKey2
  • This document is made available for download on the website that the Host controls N days prior to the examination, N being selected based on the scale of the examination.
  • Along with this PublicKey2 is made available on the website to authenticate that the downloaded document was indeed indeed generated by the Host.
  • On the day of the examination, the PublicKey1 is declared using one of the mass communication methods discussed later.
  • With PublicKey2 and PublicKey1 the Center can now read the paper.
  • The center now prints copies of the paper as per requirement.
  • It stamps each paper with it's seal/stamp/hologram to verify that this physical paper was generated at the Center.
This method has several weaknesses and strengths. Some are:
  • The points of attack on the paper are reduced to only the Website and the Center.
  • Attacks on the website require a high skill set and thus limits the number of attackers to the system.
  • With proper precautions the chances of take down are limited.
  • Proper security of the website ensures proper security of paper distribution.
  • Authentication of the paper is done through public-key cryptography
Some weaknesses are:
  • There is no way to ensure that the Center prints what is received. They may print a different paper.
  • This can be remedied with a punishment system. After the examination, the Host publishes the paper as plain text.
  • If this is found to be different from what was given to the candidate at the Center, the candidate can report it.
  • In case of a candidate reporting an institute, investigation comprising of the stamped paper with the candidate and the paper published will be undertaken by the Host.
  • If Center is found guilty, it can be blacklisted and possibly a lawsuit filed against it for breach of contract.

For mass communication, SMS, email, content on website can be used to distribute PublicKey1. This has a two fold advantage.
  • People who receive can also forward the messages to their peers.
  • Distribution is organic. False keys will be eliminated by peer review.

With such a system in place it becomes possible to distribute exam papers securely and without much risk.