Practical Deep Learning for Coders (based on fastbook)

Topics covered

• •Understand what machine learning and deep learning are, and how a neural network differs from a regular program.
• •Train your first working model (an image classifier) in a few lines of code.
• •Master key terms: model, parameters, loss, metric, training and validation sets, overfitting, transfer learning.

Topics covered

• •Go through the full project lifecycle: data collection → training → interpretation → application deployment.
• •Understand production risks: data mismatch, out-of-domain data, feedback loops.
• •Learn to recognize the ethical problems of ML systems: bias, feedback loops, disinformation — and know the tools for mitigating them.

Topics covered

• •Understand the mechanics of training 'under the hood': tensors, gradients, SGD, loss functions — and write a training loop by hand.
• •Master cross-entropy, softmax, learning rate selection, discriminative learning rates, and fine-tuning strategies.
• •Learn to work with multi-class, multi-label, and regression tasks, and apply advanced techniques: normalization, progressive resizing, TTA, Mixup, label smoothing.

Topics covered

• •Understand the idea of latent factors and embeddings, and build a recommender system on MovieLens (dot-product and neural-network variants).
• •Master working with tabular data: categorical features via entity embeddings, random forests, gradient boosting, and neural networks — and when to choose which.
• •Learn to interpret tabular models: feature importance, partial dependence, OOB estimation — and spot the pitfalls (extrapolation, leakage, domain shift).

Topics covered

• •Master the ULMFiT pipeline: pretrained language model → fine-tuning on your own corpus → text classifier.
• •Understand text processing: tokenization, numericalization, and batching for language models — and be able to build custom pipelines using fastai's mid-level API.
• •Understand how recurrent networks are built: from the simplest RNN to an LSTM with regularization (dropout, weight tying, AR/TAR).

Topics covered

• •Understand convolution as an operation: kernels, stride, padding, channels — and build a convolutional network from scratch.
• •Master techniques for stable training of deep networks: batch normalization, the 1cycle policy, activation diagnostics.
• •Understand ResNet (skip connections, bottleneck blocks) and learn to interpret CNNs via CAM/Grad-CAM and PyTorch hooks.

Topics covered

• •Understand the design of fastai's applied architectures: heads and bodies for vision (including U-Net and siamese networks), NLP, and tabular data.
• •Understand optimizers (momentum, RMSProp, Adam) and the callback mechanism that drives the training loop.
• •Build a neural network truly from scratch: matrix multiplication, forward/backward pass, backpropagation — and assemble your own Learner in plain Python/PyTorch.

Topics covered

• •Systematize what you've covered: a map of the course's topics and the connections between them.
• •Draw up a personal growth plan: projects, community, writing about your work, teaching others.
• •Carry your own capstone project through to completion and make it public.