Deep learning has transformed artificial intelligence by enabling models to learn hierarchical representations directly from raw data—dramatically outperforming traditional hand-engineered approaches across vision, language, and scientific domains. You will build the conceptual and practical vocabulary the entire course depends on: how neural networks are constructed, how training proceeds through forward and backward passes, and why deep learning is particularly suited to unstructured, high-dimensional data.

Convolutional Neural Networks are the architectural backbone of modern computer vision and a component you will encounter repeatedly throughout this course—inside autoencoders, GANs, and diffusion model U-Nets. You will develop the ability to read, design, and reason about CNN architectures from filter-level convolution operations through landmark designs like VGG and ResNet, and learn how pretrained models can be adapted to new tasks through transfer learning.

Computer vision is the field that enables machines to perceive and interpret visual information—the domain where deep learning first achieved superhuman performance. You will survey its core tasks, from image classification and object detection to semantic segmentation, then work through the full detection pipeline from the R-CNN family to YOLOv8, gaining enough architectural depth to understand how these systems are extended and fine-tuned for new domains.

The models you studied in earlier modules treat inputs as fixed-size, spatially arranged structures. Many real-world problems involve sequences where order matters and context accumulates over time: text, speech, time-series data, financial signals. You will learn how RNNs process sequences through a hidden state, how LSTMs and GRUs address the vanishing gradient problem, and why these architectures—and their failure modes—directly motivated the attention mechanism covered in the Transformer module.

This module marks the course's inflection point: the shift from discriminative models that learn decision boundaries to generative models that learn to synthesize new data. You will survey the full generative landscape—VAEs, GANs, autoregressive models, normalizing flows, diffusion models, and energy-based models—before diving into the autoencoder and its probabilistic extension, the Variational Autoencoder.

Generative Adversarial Networks take a fundamentally different approach to generative modeling: rather than maximizing a likelihood objective, two networks train in competition. You will work through the full GAN toolkit—from Deep Convolutional GANs and training stabilization techniques to Wasserstein distance, gradient penalty, conditional generation, and cycle-consistent domain translation.

Introduced in "Attention Is All You Need" (Vaswani et al., 2017), the Transformer is arguably the most consequential architectural development in deep learning since the CNN. You will derive the attention mechanism from first principles—Query, Key, Value, scaled dot-product, multi-head attention—assemble the full architecture with positional encoding and causal masking, and see it applied in a GPT-style language model.