Arun Waves

July 27, 2016

Neural Networks – scary good :-)

For the last year or so, there has been a steady increase in chatter about Artificial Intelligence/Machine Learning/Neural Network when it comes to anything online. After a brief stint with Fuzzy Logic during my undergrad, I never got an opportunity to dig into this topic further, other than keeping myself up to date with current events or milestones like the advent of IBM Watson (Jeopardy), Siri/language processing, self driving cars/various DARPA challenges, DeepBlue (chess), DeepMind (AlphaGo) etc. Almost by serendipity, I stumbled onto a Coursera course, Machine Learning by Andrew Ng (of Standford/DeepMind/Baidu fame) and found that my maths skills from my PhD in Quantum Mechanics fits right into this magical world of Neural Networks and Machine Learning. Cool, I dove right into the ocean of Deep Learning, layers,  gradients, features etc 🙂

Here is a ‘small’ result from the coding exercises of Andrew Ng’s Machine Learning course in Coursera. The task is to code a system that can recognize hand written single digit numbers and I am elated to report that Neural Networks is ‘scary good’ at it. What does ‘scary good’ mean? I did not have to explicitly code how numbers look, meaning I need not know Arabic numerals at all to develop a system that can recognize it from a non-standard medium (hand scribbled). Check out the results in GIF format below (Note: Yann LeCun’s LeNet5 has done this and more, way back in 1998);

Neural Network


Blog at