neuralnetworksfromscratch.com

What It Is

Neural Networks from Scratch is a learning book about building neural networks from the ground up in Python. Its goal is not to teach readers how to quickly call TensorFlow/Keras or PyTorch, but to help them understand how data enters neurons, how activation functions work in hidden and output layers, how loss is calculated, and how optimizers and backpropagation work together to train a model. It is accompanied by sentdex’s free YouTube tutorial videos and example code, and the book also includes QR-code animations to help explain some concepts.

Core Content and Teaching Format

The material focuses on deep learning, neural networks, and numerical computing in Python. It starts with writing a single neuron, then gradually connects neurons into layers, implements activation functions such as ReLU, Softmax, Sigmoid, and Linear, calculates cross-entropy loss, and uses backpropagation for gradient computation. It also covers optimizers including SGD, AdaGrad, RMSprop, and Adam. In terms of format, it is closer to “self-study with a book + recorded videos/code supplements” rather than a live course or 1-on-1 instruction. The book first demonstrates concepts in pure Python, then implements similar operations with NumPy, which is helpful both for understanding the underlying mechanisms and for improving NumPy skills.

Pricing, Delivery, and Support

The main text does not disclose specific pricing, but it states that ebook, paperback, and hardcover versions are available, and that purchasing any version includes access to the ebook. The PDF is usually sent by email within a few minutes after purchase, while access to the Google Docs version may take up to 48 hours. Google Docs allows readers to highlight specific passages and ask questions in comments, with the authors and other readers able to help—this is one of the more distinctive learning-support features of the product. Physical books are printed on demand globally and are typically expected to arrive in 3–6 weeks.

Pros and Cons

Its strengths are that the content is focused and does not add a lot of filler “programming basics,” making it suitable for people who truly want to understand the internal mechanics of neural networks. The path from pure Python to NumPy is also better for building fundamentals than jumping straight into frameworks. The downsides are that the refund policy is quite strict, especially for ebooks and print-on-demand physical books, which should be purchased with the expectation that they are basically non-refundable. Shipping times for physical books can also be unpredictable. The teaching language is English, which may be a barrier for some Chinese learners.

Who It’s For and Access from China

It is suitable for learners who already have a foundation in Python and object-oriented programming, have math at roughly a high-school level, and are willing to supplement their knowledge of linear algebra. It is especially well suited to people who are not satisfied with simply copying ready-made models and want to build novel deep learning applications. The source text does not specify access conditions from China; the accompanying YouTube and Google Docs resources may have uncertain accessibility in mainland China, and payment methods are not disclosed. If access is restricted, alternatives include local textbooks, domestic machine-learning courses, or using Khan Academy first to strengthen the required math foundation.