Kalhan's Blog

Deep learning has revolutionised technology, giving us everything from uncannily smart chatbots to medical imaging that can spot diseases better than the human eye. Yet, for a long time, a central mystery has haunted the field: why do these enormous models work so well? According to classical statistics, a model with billions of parameters—far more than its training data—should fail spectacularly. It ought to memorise the data, noise and all, and be unable to generalise to new, unseen examples. But deep neural networks defy this wisdom. They generalise brilliantly. How do we explain this apparent magic? There isn't one single answer. Instead, researchers view the problem through several different theoretical "lenses." Here are six of the most important ones. 1. The Linearisation Lens: The Neural Tangent Kernel (NTK) ⚙️ The NTK offers a startling insight: what if, under the right conditions, a massively complex neural network is just a simple, linear model in disguise? The...

Open Source Computer Science Degree is not a new concept. It's been in the online learning community for a while. The most prominent resources for this concept degree are,  OSSU  and  ForrestKnight . From these two, OSSU provides a very comprehensive roadmap. They also have created a community that helps the progression through the roadmap. Inspired by them I have decided to create my own version of this using only Coursera as the content provider. This is also something I'm trying myself at the moment (to some extent). What I have done here is quite simple. I have studied the OSSU and ForrestKnight degree roadmaps and I'm substituting the courses in them with relevant courses (preferably I will focus on specializations) in Coursera. There are a few reasons for focusing on Coursera, There is an option for university students to take courses their university has sponsored for free. (My reason) -  Link There is a  Coursera+  option which just has a monthly c...