Neural networks for developers who skipped grad school

What it does

This repository houses the code and codelabs for a six-episode video series aimed at software developers who want to become machine learning practitioners. It walks through building and training neural networks with TensorFlow, covering dense and convolutional networks, recurrent networks, attention mechanisms, and reinforcement learning. Each module pairs theory with a concrete project—handwritten digit recognition, Shakespeare text generation, airplane detection in aerial imagery, and a Pong-playing agent.

The interesting bit

Rather than staying in toy territory, the curriculum mixes low-level TensorFlow mechanics with production-grade patterns like the Estimator API, batch normalization diagnostics, and deployment on Google Cloud ML Engine and Cloud TPUs. It also tackles non-differentiable problems such as training a game agent via policy gradients, which sits outside standard supervised learning.

Key highlights

• Six self-contained episodes with videos, slides, codelabs, and Jupyter notebooks
• Code samples progress from MNIST (99%–99.5% accuracy) to RNN-generated Shakespeare and YOLO-based airplane detection
• Explicit coverage of training diagnostics: learning rate decay, dropout, overfitting, and verifying batch normalization
• Includes reinforcement learning with policy gradients applied to a Pong agent
• Production deployment patterns on Google Cloud ML Engine and Cloud TPUs

Caveats

• Explicitly disclaimed as “not an official Google product”
• Some episodes link to code in external repositories (attention tutorial, Shakespeare RNN)
• The repository is primarily a curated index and collection of notebooks, not a unified library

Verdict

Ideal if you learn best by watching a talk and immediately hacking on the accompanying notebook. Less useful if you need a drop-in library or a quick reference guide; this is a syllabus, not a framework.