This module invites you to explore how probability theory and sequential modeling power modern AI systems. You’ll begin by examining how conditional and joint probabilities shape predictions in language and image models, and how the chain rule enables structured generative processes. Then, you’ll transition to recurrent neural networks (RNNs), learning how they handle sequential data through hidden states and feedback loops. You’ll compare RNNs to feedforward models, explore architectures like one-to-many and sequence-to-sequence, and address challenges like vanishing gradients. By the end, you’ll understand how probabilistic reasoning and temporal modeling combine to support tasks ranging from text generation to autoregressive image synthesis.

This module explores how attention-based architectures have reshaped the landscape of deep learning for both language and vision. You’ll begin by unpacking the mechanics of the Transformer, including self-attention, multi-head attention, and the encoder-decoder structure that enables parallel sequence modeling. Then, you’ll transition to Vision Transformers (ViTs), where images are tokenized and processed using the same principles that revolutionized NLP. Along the way, you’ll examine how normalization, positional encoding, and projection layers contribute to model performance. By the end, you’ll understand how Transformers and ViTs unify sequence and spatial reasoning in modern AI systems.

In this module, you’ll explore two transformative approaches in multimodal and generative AI. First, you’ll dive into CLIP, a model that learns a shared embedding space for images and text using contrastive pre-training. You’ll see how CLIP enables zero-shot classification by comparing image embeddings to textual descriptions, without needing labeled training data. Then, you’ll shift to diffusion models, which generate images through a gradual denoising process. You’ll learn how noise prediction, time conditioning, and reverse diffusion combine to produce high-quality samples. This module highlights how foundational models can bridge modalities and synthesize data with remarkable flexibility.