Unleashing the Magic of Deep Learning
Chronicles of Deep Learning.
DeepLearning
Introduction deep learning
why we need to know about Deep Learning.
Large Language Models: From Tokens to Text
LLMs are, mechanically, a stack of transformer layers predicting one token at a time. Here’s how that simple idea, scaled up, produces something that can hol...
LoRA: Fine-tuning Giant Models Without Retraining Them
Why updating billions of frozen weights is wasteful, and how a couple of small low-rank matrices can do the job instead.
Mixture-of-Experts (MoE): Huge Models, Cheap Inference
How routing each input to only a few ‘expert’ sub-networks lets models grow to trillions of parameters without a proportional compute bill.
Neural Networks: How They Actually Work
Before transformers, MoE, or Mamba make sense, you need the fundamentals: what a neuron computes, how a network learns, and why ‘deep’ just means ‘more layer...
Retrieval-Augmented Generation: Giving a Model a Memory It Doesn’t Have to Memorize
An LLM’s knowledge is frozen at training time and crammed into its weights. RAG sidesteps both problems by letting the model look things up instead of memori...
State Space Models and Mamba: Beyond Quadratic Attention
Why attention gets expensive as sequences grow, and how Mamba borrows ideas from control theory to process long sequences in linear time.
Tokenization: How Neural Networks Read Text
Neural networks only understand numbers. Tokenization is the step that turns ‘Hello world’ into something a model can actually compute on — and the choices m...
Vision Transformers (ViT): Attention Instead of Convolutions
How the transformer architecture that took over NLP learned to see images, one patch at a time.