Machine learning in Python
I explore the vast landscape of machine learning in Python, covering everything from deep neural networks and natural language processing to reinforcement learning and production deployment.
A topic page: a summary of what my posts say, with each claim quoted from the post it came from. The posts are the authority. Last checked 2026-11-06.
Summary. I explore the vast landscape of machine learning in Python, covering everything from deep neural networks and natural language processing to reinforcement learning and production deployment. I demonstrate how specific architectures, optimization techniques, and data pipelines allow models to solve complex problems like image recognition and sequence forecasting.
What the posts say
The SELU activation function allowed a 100-layer network to reach 82.0% accuracy.
“In my run, a bend called SELU (scaled exponential linear unit), with matching starting numbers, took the 100-layer network from 69.9% with ReLU to 82.0%.” — Scikit-Learn (Part 11)
The Huber loss is less affected by a few very wrong answers than standard regression.
“The Huber loss squares small errors but only counts large ones in proportion, so a few odd houses do not pull the whole model towards them.” — Scikit-Learn (Part 12)
The prefetch function allows the model to work on one batch while the next is prepared.
“A: It prepares the next batch while the model works on the current one.” — Scikit-Learn (Part 13)
The Xception model achieved 91.8% accuracy after a two-step transfer learning process.
“Then it unfreezes the network from layer 56 on, and continues for 10 epochs with a step size ten times smaller (0.01): that gave 91.8% by the end.” — Scikit-Learn (Part 14)
A simple forecast provides a necessary baseline to determine if a complex model is worth the effort.
“The lesson: work out what a simple guess scores before building a model, because it is hard to beat by much, and it tells me whether a clever model is worth having.” — Scikit-Learn (Part 15)
The Transformer uses positional encodings to maintain word order since it lacks a recurrent memory.
“The notebook adds a pattern of sine and cosine waves to each word, different for each position, so that word order is in the numbers.” — Scikit-Learn (Part 16)
Diffusion models are more stable than GANs because they learn a simple, fixed task.
“A diffusion model avoids that.” — Scikit-Learn (Part 17)
The replay buffer prevents the network from forgetting earlier lessons by providing a variety of past steps.
“The DQN trains on random batches from it, instead of on the most recent steps, because steps in a row are too alike, and training on them makes the network forget what it learned earlier.” — Scikit-Learn (Part 18)
The TensorFlow Lite conversion reduced the model size by about 3.6 times.
“Converting for a phone with TensorFlow Lite made the same model about 3.6 times smaller, 171,748 bytes to 47,928.” — Scikit-Learn (Part 19)
A model must be judged on data it has not seen to avoid memorization.
“Judge a model on data it has not seen. A model can memorise what it learned from and look perfect.” — Scikit-Learn (Part 20)
The posts behind this page
- Scikit-Learn (Part 11) — Training deep neural networks: why deep networks are hard to train, and the starting numbers, activation functions, batch normalization, optimizers, and dropout that make it work
- Scikit-Learn (Part 12) — Custom models and training with TensorFlow: tensors, automatic slopes with GradientTape, and a training loop written by hand instead of calling fit()
- Scikit-Learn (Part 13) — Loading and preprocessing data with TensorFlow: a tf.data pipeline that reads many files, shuffles, and batches, the TFRecord format, and Keras layers that turn raw values into numbers
- Scikit-Learn (Part 14) — Deep computer vision with convolutional neural networks (CNNs): filters that slide across an image, pooling, a CNN for clothing, a pretrained ResNet, and transfer learning for flowers
- Scikit-Learn (Part 15) — Processing sequences with recurrent neural networks (RNNs) and CNNs: forecasting daily train ridership, comparing a naive guess, a statistical model, and RNNs, and forecasting several days ahead
- Scikit-Learn (Part 16) — Natural language processing with recurrent networks and attention: a Shakespeare character model, movie review sentiment, translation, the Transformer, and pretrained models from Hugging Face
- Scikit-Learn (Part 17) — Autoencoders, GANs, and diffusion models: networks that squeeze data and rebuild it, and networks that invent new clothing pictures
- Scikit-Learn (Part 18) — Reinforcement learning: an agent that learns to balance a pole by trial and error, with policy gradients and deep Q-networks
- Scikit-Learn (Part 19) — Training and deploying models at scale: serving a TensorFlow model with Docker over REST and gRPC, updating it live, shrinking it for a phone, and spreading training over several devices
- Scikit-Learn (Part 20) — Looking back over nineteen chapters of Hands-On Machine Learning: how models learn, which parts are still current, and what I would do differently