Course overview

Understand the data structures that neural networks operate on at every layer.

Learn how small changes in weights ripple through a network and drive learning.

Ground your understanding of loss functions, predictions, and model uncertainty in statistical thinking.

Get comfortable with the scientific Python stack you'll use alongside PyTorch throughout the course.

Configure a clean, reproducible PyTorch environment with GPU acceleration ready to go.

Master the core object in PyTorch and learn how to create, inspect, and manipulate tensors confidently.

Perform efficient mathematical operations on tensors, including PyTorch's powerful implicit broadcasting rules.

See how PyTorch silently builds a graph of operations to compute gradients automatically during backpropagation.

Learn the simple but critical skill of shuttling tensors and models between devices for accelerated training.

Discover PyTorch's base class for all models and learn how to structure your networks cleanly and extensibly.

Build the core computational units of a network and introduce non-linearity that enables complex learning.

Choose the right objective to minimize and the algorithm to minimize it, from MSE to Adam.

Internalize the four-step cycle that drives learning in every neural network you'll ever train.

Persist your trained weights to disk and restore them for inference or continued training.

Use PyTorch's native abstractions to feed batched, shuffled data into your models with minimal boilerplate.

Wrap any data source — images, CSVs, audio — into a PyTorch-compatible dataset class.

Artificially expand your training data and improve generalization through on-the-fly transformations.

Apply sampling strategies and loss weighting to stop your model from ignoring minority classes.

Eliminate data bottlenecks using multiprocessing, pinned memory, and prefetch techniques.

Learn how convolutional filters detect spatial patterns in images hierarchically and efficiently.

Control the spatial resolution and receptive field of your network with these essential layer parameters.

Leverage models trained on millions of images to dramatically accelerate your own training.

Decide how much of a pretrained model to update based on your dataset size and task similarity.

Go beyond classification to locate and identify multiple objects within a single image.

Convert discrete tokens into dense, meaningful vector representations that capture semantic relationships.

Model sequential dependencies in text and time-series data with memory-aware architectures.

Learn how models learn to focus on the most relevant parts of an input sequence when making predictions.

Build the encoder-decoder architecture that revolutionized NLP and underpins modern foundation models.

Adapt powerful pretrained language models to downstream tasks with just a few lines of PyTorch-compatible code.

Dynamically adjust the learning rate during training to converge faster and escape local minima.

Prevent overfitting and stabilize training using the most widely used regularization techniques in deep learning.

Safeguard your training runs from exploding gradients and floating-point instability.

Cut memory usage and double training speed by combining 16-bit and 32-bit floating point operations.

Scale training across multiple GPUs or machines to handle larger models and datasets.

Visualize loss curves, metrics, and model outputs in real time to monitor training health.

Diagnose and fix the most common training failure modes using gradient inspection techniques.

Identify CPU/GPU bottlenecks in your training pipeline down to the individual operator level.

Ensure your experiments produce consistent results so findings can be trusted and shared.

Automate the search for optimal learning rates, architectures, and training configurations.

Train two competing networks to synthesize realistic images, data, or other high-dimensional outputs.

Learn a compact, structured latent space that enables controlled generation and data interpolation.

Understand the generative approach behind state-of-the-art image synthesis models like Stable Diffusion.

Extend deep learning to non-Euclidean data like social networks, molecules, and knowledge graphs.

Train agents to make sequential decisions by learning from rewards through direct environment interaction.

Choose the right export format for your deployment target, from Python servers to mobile and edge devices.

Shrink model size and boost inference speed with minimal impact on accuracy.

Expose your model as a scalable REST API that can handle real-world production traffic.

Package your model, dependencies, and serving logic into a portable, reproducible container.

Automate testing, validation, and deployment of new model versions using modern DevOps practices.

Detect when your model's real-world performance degrades and set up alerts before users are impacted.