What is arhams-deeplearning-roadmap?

This repository is a curated two week self study plan for learning deep learning from the ground up. It is not a piece of software you install and run. Instead, it is a structured guide made of markdown lesson files that link out to existing videos, books, and courses from well known teachers such as Andrej Karpathy, 3Blue1Brown, and Sebastian Raschka. The path starts small: learners begin with a short, plain NumPy neural network, included in the repository, that recognizes handwritten digits, and are asked to understand and fix small intentional flaws in it. From there, each of about twelve lessons builds on the last, moving through backpropagation, basic language models, attention, building a small version of a GPT style model, and tokenization, before reaching pretraining a base model and then post training it with techniques like supervised fine tuning, direct preference optimization, and reinforcement learning for reasoning. The final lessons cover current fine tuning methods like LoRA and suggest a capstone project. Every lesson offers two parallel options: a primary video to watch, or a full written alternative covering the same material for someone who prefers reading. Each lesson follows the same pattern of stating a goal, explaining why it matters, linking resources, asking the learner to build something hands on, and ending with a checkpoint question to confirm understanding before moving on. The author states that every linked resource was checked to make sure it still works, rather than assumed from memory. A small included folder holds the reference handwritten digit recognizer mentioned above, written in plain NumPy without any deep learning framework. The rest of the repository is the lesson files themselves, plus a single combined file for anyone who prefers one long document. Most of the roadmap does not require a graphics card, though the sections on training larger models point learners toward renting cloud computing time. The project is released under the MIT license, which allows free use, sharing, and modification.