sankho roy·Oct 1Wait! You Changed ReLU… Did You Change the Initialisation Too?”The hidden relationship between activation functions, variance propagation, and stable neural networks.
sankho roy·Jul 12The Science Behind Million-Token Context Windows: RoPE, ALiBi, and YaRNHow RoPE, ALiBi, and YaRN quietly solved one of the biggest challenges in modern Transformers.A response icon1A response icon1
sankho roy·Jul 11The ₹300 Crore Question Every Boardroom Asks — And Why Traditional Marketing Mix Models Can No…“Half the money I spend on advertising is wasted; the trouble is I don’t know which half.” — John WanamakerA response icon1A response icon1
sankho roy·Jun 25We Are Not Astrologers, We Are Machine Learning ScientistsWhy Predicting a Number Is Easy, but Predicting Uncertainty Is Science
sankho roy·Jun 20If Autoencoders Can Compress Data, Why Did We Invent Variational Autoencoders?No worries if you don’t understand the math. Just understand the story.
sankho roy·Jun 20Your Model Scored 96% Accuracy. Think That’s Success? Think Again.How a highly accurate model can still lose money — and why calibration and threshold tuning are the real keys to business success
sankho roy·Jun 18We never use MSE for Classification, But “WHY”?If MSE works perfectly for regression, why does almost every classifier use Cross-Entropy instead? After all MSE values at 0 or 1 are same…
sankho roy·May 20PyTorch Hooks and Integrated Gradients: Understanding What Your Neural Network SeesLearn how forward hooks, backward hooks, and Integrated Gradients help interpret deep learning models in PyTorch.
sankho roy·May 16NTK-Aware Interpolation in YaRN — The Missing Intuition Behind Long Context LLMsLarge Language Models like GPT, Llama, Qwen, and DeepSeek can now handle context windows of 128K, 1M, or even more tokens.A response icon1A response icon1
sankho roy·May 16The Evolution of Deep Learning OptimizersFrom Gradient Descent to Muon, SOAP, and Distributed Shampoo