About
This book is the modern neural networks foundation, and it's taught the way it should be.
The way neural networks function has changed, of course, and this second edition has to change too. It's all rebuilt around the latest versions of Python 3.14, NumPy 2.0 and PyTorch 2.0, so it'll be the only framework you'll need to get up to speed quickly. The likes of TensorFlow, Keras, RNNs, GANs and capsule networks are now a thing of the past. Now, the big players in the AI world are convolutional networks, attention and transformers, vision transformers, Kolmogorov-Arnold networks, state space models, diffusion transformers and multimodal language models. This book is all about building a single application using the same data, training it using a single pipeline. That way, you can compare it directly with other applications and see how it really compares. It's all done by hand in NumPy, then rebuilt in PyTorch, so nothing stays a black box.
This book is written for data scientists and AI engineers who want depth without the dependency bloat, keeping its toolkit to five libraries and its focus on understanding. The book makes you capable to read any new architecture paper and recognise the parts, because you'll have built them yourself.
Key Learnings
Build modern architecture by hand in NumPy, and then rebuild it in PyTorch.
Split work cleanly and write one training loop that drives architecture unchanged.
Diagnose overfitting with learning curves, then apply the full regularization toolkit.
Design convolutional networks that treat images as spatial objects, not flat vectors.
Implement attention and transformers from scaled dot-product to full encoder blocks.
Cut images into patches and train vision transformers from scratch.
Build diffusion models that generate crisp images through iterative denoising.
Compare architectures honestly using one dataset, one seed, one pipeline.
Recognize the reusable parts inside any new architecture paper you read.
Table of Content
Setting up Neural Network Stack
Data Pipelines with NumPy and Pandas
Feedforward Networks in Depth
Convolutional Networks for Visual Tasks
Autoencoders and Variational Autoencoders
Attention and Transformers
Vision Transformers
Kolmogorov-Arnold Networks
State Space Models
Diffusion Models and Diffusion Transformers
Multimodal LLMs
The way neural networks function has changed, of course, and this second edition has to change too. It's all rebuilt around the latest versions of Python 3.14, NumPy 2.0 and PyTorch 2.0, so it'll be the only framework you'll need to get up to speed quickly. The likes of TensorFlow, Keras, RNNs, GANs and capsule networks are now a thing of the past. Now, the big players in the AI world are convolutional networks, attention and transformers, vision transformers, Kolmogorov-Arnold networks, state space models, diffusion transformers and multimodal language models. This book is all about building a single application using the same data, training it using a single pipeline. That way, you can compare it directly with other applications and see how it really compares. It's all done by hand in NumPy, then rebuilt in PyTorch, so nothing stays a black box.
This book is written for data scientists and AI engineers who want depth without the dependency bloat, keeping its toolkit to five libraries and its focus on understanding. The book makes you capable to read any new architecture paper and recognise the parts, because you'll have built them yourself.
Key Learnings
Build modern architecture by hand in NumPy, and then rebuild it in PyTorch.
Split work cleanly and write one training loop that drives architecture unchanged.
Diagnose overfitting with learning curves, then apply the full regularization toolkit.
Design convolutional networks that treat images as spatial objects, not flat vectors.
Implement attention and transformers from scaled dot-product to full encoder blocks.
Cut images into patches and train vision transformers from scratch.
Build diffusion models that generate crisp images through iterative denoising.
Compare architectures honestly using one dataset, one seed, one pipeline.
Recognize the reusable parts inside any new architecture paper you read.
Table of Content
Setting up Neural Network Stack
Data Pipelines with NumPy and Pandas
Feedforward Networks in Depth
Convolutional Networks for Visual Tasks
Autoencoders and Variational Autoencoders
Attention and Transformers
Vision Transformers
Kolmogorov-Arnold Networks
State Space Models
Diffusion Models and Diffusion Transformers
Multimodal LLMs
Related Subjects
Extended Details
- EditionSecond Edition