Deep Reinforcement Learning
Supervised learning has a fixed dataset. Reinforcement learning creates its own, which changes everything.
-
-
Model Compression and Deployment
The model is the easy part. Everything after training is where deployments fail.
-
Adversarial Robustness in Deep Learning
A change too small to see can flip a classifier’s answer. That is a property of the models, not a bug in one of them.
-
Uncertainty Estimation and Model Calibration
A softmax output of 0.99 is not a 99% chance of being right. Here is how far apart those two things are.
-
Deep Learning Interpretability
A model that is right for the wrong reason will fail the moment the reason changes.
-
Graph Neural Networks
Images have a grid, sentences have an order, and a molecule has neither. Message passing is the generalization.
-
Self-Supervised Learning
Labels are the expensive part of supervised learning. These methods construct the supervision from the data itself.
-
Autoencoders and Representation Learning
Reconstruction is a pretext. What you actually want is the code in the middle.
-
Vision Transformers and Modern Vision Models
A transformer sees an image as a sequence of patches and assumes nothing about geometry. That is both its strength and its data requirement.
-
Building a Transformer Language Model in PyTorch
Every component explained separately, now assembled into something you can train tonight.
-
BPE, WordPiece, and SentencePiece Tokenization
Every model’s view of language is filtered through its tokenizer, and most surprising model behavior starts there.
-
Transformer Architecture and Self-Attention
Remove recurrence, keep attention, and add back the position information you just threw away.
-
Sequence-to-Sequence Models and Attention
Attention was invented to fix one specific failure: compressing a whole sentence into one vector.
-
RNN, GRU, and LSTM Recurrent Networks
Gates exist for one reason: to give the gradient an additive path through time.
-
Word Embeddings
A word’s meaning is approximated by the company it keeps. That single idea produces vectors you can do arithmetic on.
-
Neural Style Transfer
The one task where you freeze the network and train the image instead.
-
Face Recognition and Metric Learning
When every person is a class and new people arrive daily, classification stops being the right formulation.
-
Semantic Segmentation and U-Net Explained
Classification per pixel sounds simple until you have to recover the resolution you threw away.
-
Object Detection with YOLO, IoU, and NMS
Detection is classification plus regression plus a deduplication problem nobody mentions until it breaks.
-
PyTorch Image Classification Fine-Tuning
Almost nobody trains an image classifier from scratch. Here is what the alternative actually looks like.
-
CNN Architecture Comparison
Each architecture solved one specific problem left over from the previous one.
-
Convolutional Neural Networks
A convolution is a fully connected layer with two constraints. Those constraints are the whole idea.
-
Data Mismatch, Transfer, and Multitask Learning
When training data and production data differ, the usual bias-variance diagnosis gives the wrong answer.
-
Machine Learning Model Improvement Strategy
The hard part is not training a model. It is knowing which of ten possible improvements to try next.
-
Distributed Training with PyTorch DDP
Multiple GPUs are not free speed. Here is what actually has to be synchronized.
-
PyTorch Mixed Precision and GPU Training
Most training loops are slower than they need to be, and the bottleneck is rarely where people assume.
-
Reproducible Deep Learning Training
An experiment you cannot repeat is an anecdote, not a result.
-
PyTorch Training Loop
Six lines in the right order, and the reasons each one has to be there.
-
PyTorch Tensors and Autograd
Autograd is the computation graph you already built, constructed for you at runtime.
-
Hyperparameter Tuning for Deep Learning
Tuning is a search problem with a budget. Spend it on the parameters that matter.
-
Preventing Overfitting in Deep Learning
A practical framework for recognizing overfitting and choosing the right regularization method.
-
Normalization Layers in Deep Learning
Every normalization layer answers the same question differently: which axes do we average over?
-
Learning Rate Schedules and Warmup
The learning rate is the single most important hyperparameter, and a constant value is almost never the right choice.
-
Deep Learning Optimization Algorithms
Understand what popular deep-learning optimizers remember, how they update parameters, and when to tune the learning rate.
-
Gradient Checking for Neural Networks
A wrong gradient produces a model that trains slowly instead of crashing. Here is how to catch it.
-
Weight Initialization and Gradient Flow
Initialization is not a detail. It decides whether a deep network trains at all.
-
Deep Learning Loss Function Design
The loss function is the only place where you tell a network what “better” means. Choose it deliberately.
-
Softmax and Multiclass Classification Explained
Extend binary logistic regression to many classes with softmax, cross-entropy, and a numerically stable NumPy implementation.
-
Deep Neural Network Backpropagation
Follow data and gradients through an L-layer neural network with dimension rules and reusable NumPy formulas.
-
Building a Shallow Neural Network with NumPy
Implement a one-hidden-layer neural network from scratch and understand every forward and backward equation.
-
Computation Graphs and Backpropagation
A visual and numerical explanation of how local derivatives combine into backpropagation through a computation graph.
-
Neural Network Forward Propagation
A beginner-friendly explanation of neural-network structure, notation, hidden layers, and forward propagation with a verified NumPy example.
-
Logistic Regression for Binary Classification
A practical guide to logistic regression, from probability and cross-entropy to vectorized gradients and a complete NumPy implementation.
-
NumPy Vectorization and Broadcasting
Replace slow loops with clear matrix operations and learn how to prevent silent broadcasting and shape errors in NumPy.
-
Probability and Information Theory
Every loss function in this series is a probability statement in disguise. Here is the vocabulary.
-
Linear Algebra for Deep Learning
A neural network is a stack of matrix multiplications. This is the algebra that makes that sentence mean something.