By Shivam Panda
In collaboration with Arnav Sampigethaya
Machine Learning can be incredibly daunting to think about. How can something without a brain learn? How can something with no ‘creativity’ create? How can we replicate human thought in the form of some lines of code? These are questions that everyone had and are intimidated by.
What separates traditional programming from Machine Learning is the fundamental of what’s happening. In a traditional program, you input data and the method and you get an output. For example, if I wanted a program to tell me the sum of two numbers, I would say a=5, b=10, the method of finding the sum is a+b, and the program would output 15. In Machine Learning, you give the program the data and the correct answers and the machine tries to discern the formula or method. You input a hundred unique pairs of numbers and their sums and the machine will figure out what the relationship is. While on an easy task like adding two numbers, traditional programming might be more intuitive, there are many cases where traditional paradigms fail us.
The most popular form of Machine Learning is supervised learning, in this form of learning, we tell the program the correct answers. Just like in our brain, in our programs, we have neurons. The more neurons we have, the higher the complexity of our model. A collection of neurons is called a layer. There has to be an input layer and an output layer (for obvious reasons), but what makes Neural Networks so revolutionary is the hidden layers.
Source- Towards Data Science
Essentially what happens is the input data is processed through a layer which tries to form an equation aX1 + bX2 + cX3 …… = output layer. The machine learning algorithm will try to perfect the values of a,b,c to best match the correct answers it’s been given. Fitting of a model to training data is one of the most important aspects of Machine Learning. In statistics, a fit refers to how well you approximate a target function. Overfitting refers to a model that reflects the training data too well. Overfitting happens when a model learns everything in the training data including the details and the noise making it so precise that it cannot be used to predict values for new data. The ways overfitting can happen is if you go over the data too many times (epochs in technical terms), if you have too many neurons or too little training data with many epochs. Underfitting means that the model can neither predict accurate values for the training data nor for new values. This is a much rarer occurrence as accuracy can be tracked and quickly remedied in earlier stages. The tough task for any Machine Learning engineer is finding the sweet spot. The trouble is, we don’t know the values of a,b,c etc and there’s no way to make the machine tell us. AI is a black-box which means we can’t look inside and see what it’s thinking.
The biggest trouble in Machine Learning is finding enough training data. This is why companies collect your data and why data is so valuable. The captchas and tests you do to prove you are human are training robots to distinguish between commonplace objects and better understand text. Machine Learning is a beautiful thing, it’s like creating life, but that life is extremely specialized to do one task and that task only.
Machine Learning and AI Algorithms dominates the majority of today’s software, be it open to public or private. Private corporations use AI programs in data analysis to fetch raw data that each employee puts out and creates helpful statistics that can help optimize workforce efficiency. Data Analytics is considered a very niche position these days and is something that is in high demand. Using Machine Learning predictions of the future are made and preparations are done accordingly.
Particle physicists all over the world are colliding super small particles at super-high speeds and seeing how it breaks and checking what's new. That seems so common like, “Ah okay, regular scientist stuff!” but it's actually not that simple. At speeds close to the speed of light, sensor reaction times are not accurate enough to pick up the changes in the environment caused by these collisions so there is no direct way of learning what happened inside the collision chamber. ML uses data collected after the collision to simulate the path of the collisions particles as a function in time and can calculate their positions at any time before that. Deterministic equations of such complexity aren't fathomable by the human brain, so we have artificial intelligence to thank for some of the most recent breakthroughs.
Celestial phenomena take millions of years to perceive due to the vast distances that come into the picture. The first supernova explosion was perceived in the 1800s but through ML we know that it happened when Alexader was out to conquer the world.
Do you ever open YouTube and see something you find interesting and proceed to watch the video? It's not a mere coincidence you ‘stumbled’ across that video. It was put there by a machine learning algorithm that identifies genres and channels you watch frequently, catalogues it to your account and ‘intuitively’ gives you suggestions. This works the same for targeted adds on shoddy websites, product suggestions of e-com websites and even you google search auto-complete.
Modern medicine has identified over 200 pathogenic diseases and there exists uncountable medicinal cures made privy to the public courtesy of the pharmaceutical industry. AI is used to run simulations on cures, prospective medicinal drugs, and new methods of treatments. These programs execute intricate algorithms to determine new and valid combination of atoms to form complex molecules to base a new drug on, therefore getting rid of a huge hurdle for those in the research and development sector.
Forensic scientists use machine learning to cross-check the probability of the method of death and check the validity of the evidence used to come to certain conclusions. This advance has made the work of real-life detectives much easier.
Machine Learning is a tool of the future, with it so much is accomplishable, and because of it so much of the impossible is now possible, so much of the unfathomable is now fathomable, and to those who are scared by the genius of it I ask, do you fear it for what it is, or what it can be? Because we control what it can be, and we control what to do with the knowledge of it.
Comments