Immortal Brains Blog
Explore the edges of technology, science and the mind.
Deep Learning Course
Seventeen lessons, from Python to diffusion models, taught top-down: every lesson starts with a working result in a few lines of code, then opens the box, and the mathematics arrives as answers to the questions the result raises.
Course · 4 lessonsDesign Patterns Course
All twenty-three Gang of Four design patterns, plus Simple Factory and Repository, in modern C#, in four lessons: what a pattern is, then the creational, structural and behavioural families. Every pattern has its intent in one line, the smallest C# that expresses it, when to use it and what it costs, and where it appears in low-level design problems such as the parking lot and the elevator.
Declarative Design Patterns in C#
What a design pattern is, what the creational, structural and behavioural families are, and how delegates, events, generics, records and pattern matching in modern C# replace most of the class hierarchies. The overview of a course covering all twenty-three Gang of Four patterns.
Software EngineeringCreational Design Patterns in C#
Simple Factory, Factory Method, Abstract Factory, Builder, Singleton and Prototype in modern C#, written with generics, default interface methods, Lazy<T> and records, each with its uses, trade-offs and examples from low-level design problems such as the parking lot.
Software EngineeringStructural Design Patterns in C#
Decorator, Adapter, Facade, Bridge, Composite, Flyweight, Proxy and Repository in modern C#, written with small wrappers, default interface methods, ConcurrentDictionary and DispatchProxy, each with its uses, trade-offs and examples from low-level design problems such as the parking lot.
Software EngineeringBehavioural Design Patterns in C#
Strategy, Command, Template Method, Chain of Responsibility, Observer, State, Iterator, Mediator, Memento, Visitor and Interpreter in modern C#, where delegates, events, records and pattern matching replace the class hierarchies, each with its uses, trade-offs and examples from low-level design problems such as the elevator.
AIEssential Python for Machine Learning
The Python you need for machine learning: pythonic idioms, lists, sets and dictionaries, mutability, everyday statements, and the NumPy and Pandas foundations every deep learning project is built on.
AIYour First Deep Learning Model
Train a 37-breed pet classifier in minutes with transfer learning, then see what goes in and what comes out: tensors, logits, softmax, and what makes a list of numbers a probability distribution.
AIPretrained Text Models and the Geometry of Meaning
Sentiment, text generation and semantic search with Hugging Face, then how meaning becomes geometry: vectors, lengths, the dot product derived from the law of cosines, projections and cosine similarity.
AIExploratory Data Analysis and the Statistics Behind It
Find and fix the problems in a dataset with Pandas, and the statistics that come with it: why variance squares and divides by n − 1, probability densities, the normal curve, correlation as an angle, and why scaling speeds up training.
AIIs My Model Any Good?
Why 96% accuracy can mean almost nothing: confusion matrices, precision, recall, F1, ROC and AUC, where cross-entropy comes from (likelihood and the Bernoulli distribution), and the binomial error bars on a test score.
AIPyTorch Fundamentals, and What a Matrix Really Is
Tensors, Datasets, DataLoaders, nn.Module, autograd and the GPU, written out by hand, and the matrix ideas underneath: matrices as functions, why multiplication goes row by column, and why mini-batches work.
AIInside a Neural Network
From the perceptron to multi-layer networks: the geometry of a neuron, a proof that one neuron can't learn XOR, why depth needs non-linearity, universal approximation, and choosing initial weights from variances.
AIHow Networks Learn: Gradient Descent and Backpropagation
Train a network in four lines, then open the box with a ladder of questions: tangent lines, limits and ε-δ, continuity, Taylor's theorem and the learning-rate limit, gradients, convexity and the chain rule.
AITraining Well: Regularisation, Learning Rates and Tuning
Why networks memorise and how to stop it: the bias-variance decomposition, weight decay as a Gaussian prior, why L1 gives zeros, dropout, early stopping, learning-rate schedules and random search.
AIConvolutional Neural Networks for Computer Vision
How CNNs see: convolution as a matrix with shared weights, translation equivariance, the output-size formula, receptive fields measured with autograd, and why skip connections keep deep networks trainable.
AIRecurrent Neural Networks for Sequences
Recurrent networks for sequences in PyTorch: the recurrence, backpropagation through time, and eigenvalues, which explain exactly why gradients vanish or explode, and how the LSTM's additive memory fixes it.
AINLP with Deep Learning: Word and Contextual Embeddings
How words become vectors whose geometry captures meaning: co-occurrence, pointwise mutual information, the SVD, word2vec from scratch, why analogies work, and contextual embeddings from BERT.
AIAttention and Transformers
The transformer built piece by piece in PyTorch: attention as a soft lookup, why we divide by √d_k, attention as an expected value, why order must be added, and sinusoidal positions as rotations.
AIWorking with Large Language Models
How LLMs predict, generate and become assistants, and how to adapt them: the chain rule of probability, perplexity, entropy and temperature, LoRA fine-tuning through matrix rank, and retrieval-augmented generation.
AIVAEs and GANs for Image Generation
Generating images with VAEs and GANs: multivariate Gaussians, the reparameterisation trick, the KL divergence derived, Jensen's inequality and the ELBO, and the calculus of the GAN's optimal discriminator.
AIDiffusion Models and Generative AI
Build a diffusion model that turns noise into data, with the maths of why it can jump to any noise level, then step back to generative AI as a whole: Bayes' rule, model families, synthetic data and responsibilities.
AICapstone Project
Build and defend a deep learning project of your own: project options, a step-by-step plan, a reusable PyTorch template, an evaluation checklist, and the bootstrap for telling real improvements from luck.
ImmortalityDigital Immortality: What Is Possible Today and What Is Not
Digital immortality means three different things: a chatbot that imitates you, a preserved brain, and a mind running as software. What has been done at each level, how far brain mapping has come, and why a perfect copy might still not be you.
ImmortalityBrain-Computer Interfaces: How They Work and What They Can Do
How a brain-computer interface turns neural activity into speech and movement: the four kinds of sensor, how the decoding model is trained, what has been achieved in people since 2004, and what is still hard.