Introduction to Deep Learning

Overview: This foundational lesson introduces the core building blocks of deep learning. We begin by examining linear classifiers, specifically focusing on the Softmax function and the use of Cross-Entropy loss to quantify error. Building upon this, we define the structure of a basic feedforward neural network (Multi-Layer Perceptron), detailing the role of weights, biases, and non-linear activation functions (e.g., ReLU). The major focus shifts to the optimization process required to train these highly parameterized models. We will introduce Gradient Descent (GD) as the core optimization algorithm, contrasting its computational requirements with those of Stochastic Gradient Descent (SGD) and Mini-batch GD. Crucially, the lesson culminates in a detailed explanation of the backpropagation algorithm, showing how the chain rule from calculus is applied efficiently via computation graphs to calculate gradients necessary for weight updates across all layers. Learning Outcomes:

• Define the structure of a basic feedforward neural network and explain the necessity of non-linear activation functions (e.g., ReLU).
• Formulate classification loss functions (e.g., Softmax and Cross-Entropy) and understand how they quantify model error.
• Explain the mechanics of Gradient Descent (GD) and differentiate between its variants (SGD, Mini-batch GD) in terms of convergence and computational efficiency.
• Derive the backpropagation algorithm using the chain rule and demonstrate its implementation via computational graphs for calculating gradients.
• Identify the key mathematical prerequisites (linear algebra and multivariate calculus) required to understand neural network optimization.

Overview: This lesson transitions from theoretical concepts to production-ready deep learning implementation using PyTorch, the core library for this course. We begin with PyTorch fundamentals, detailing the Tensor structure, utilizing CUDA for GPU acceleration, and understanding automatic differentiation through the dynamic computation graph. A critical focus will be placed on efficient data handling: introducing the PyTorch Dataset class for data abstraction and the DataLoader for managing large datasets, enabling batching, shuffling, and multi-process data loading. Finally, we will address practical considerations for scaling training, covering memory management optimization, techniques like gradient accumulation, and introducing the core concepts behind distributed training (e.g., Data Parallelism) necessary for working with models that exceed single-GPU capacity. Learning Outcomes:

• Implement core deep learning operations using PyTorch Tensors and utilize its automatic differentiation features for gradient calculation.
• Design and implement efficient data pipelines using PyTorch Dataset and DataLoader abstractions to handle large-scale, batched data inputs.
• Configure models and data for training on CUDA-enabled GPUs to significantly accelerate the training and inference process.
• Explain the role of memory optimization techniques, such as gradient accumulation, and understand the conceptual basics of distributed training for scalability.

Overview: This lesson introduces Convolutional Neural Networks (CNNs), the cornerstone of modern computer vision. We will deeply explore the foundational modules: the Convolutional Layer and the Pooling Layer. For the Convolutional Layer, we will cover the operation mathematics, including the roles of kernels (filters), stride, and padding, and discuss key concepts like local connectivity and parameter sharing that make CNNs efficient for high-dimensional image data. We will differentiate between Max Pooling and Average Pooling and explain their critical role in downsampling and inducing translation invariance. Finally, we will synthesize these layers into complete, basic CNN architectures, illustrating the common sequential transition from raw pixel data through hierarchical feature extraction stacks to fully connected layers for final classification, using classic models like LeNet-5 as representative examples. Learning Outcomes:

• Explain the mathematical operation of 2D convolution, including how filter size, stride, and padding affect the output feature map dimensions.
• Articulate the concepts of local connectivity and parameter sharing and how they contribute to the efficiency and effectiveness of CNNs compared to fully connected networks for image data.
• Differentiate between Max Pooling and Average Pooling, and describe the primary purpose of pooling layers in feature map downsampling and achieving translation invariance.
• Design and analyze a basic sequential CNN architecture composed of interleaved convolution, activation (ReLU), pooling, and fully connected layers.

Overview: This lesson advances beyond foundational CNNs (like AlexNet) to explore sophisticated and highly influential deep learning architectures used in state-of-the-art computer vision tasks. We will analyze the design principles and innovations behind key models, including the streamlined depth of VGG networks, the multi-scale feature aggregation of Inception (GoogLeNet), and the critical use of residual connections in ResNet to overcome the vanishing gradient problem in extremely deep networks. The second half of the lesson focuses on the vital topic of model interpretability and explainable AI (XAI). Students will learn visualization techniques, such as inspecting feature map activations, and delve into gradient-based localization methods. Specifically, we will cover the mechanics and implementation of Class Activation Mapping (CAM) and its gradient-based generalization, Grad-CAM, which visually explains network decisions by highlighting salient regions in the input image. Learning Outcomes:

• Compare and contrast the core architectural innovations (e.g., residual connections, inception modules) of VGG, GoogLeNet, and ResNet models.
• Explain the role and challenges of scaling up network depth, specifically addressing the degradation problem and how ResNet mitigates it.
• Detail fundamental methods for feature visualization, including inspecting intermediate layer activations and learned filters.
• Outline the theoretical mechanism of Class Activation Mapping (CAM) and Grad-CAM for generating visual explanations based on gradient flow.
• Apply interpretability techniques to analyze and diagnose the decision-making process of advanced CNNs in classification tasks.

Overview: This lesson introduces the challenges of modeling structured data, specifically sequences (e.g., text, time series), which violate the assumption of independence common in feedforward networks. We will define sequence modeling tasks, such as machine translation, speech recognition, and time series prediction, emphasizing the need for a mechanism to maintain state information. The core focus will be on the architecture of traditional Recurrent Neural Networks (RNNs). Key concepts covered include the shared weight mechanism, unfolding computation graphs across time steps, calculating hidden state updates (h_t), and handling variable-length input sequences. We will also examine the primary limitations of basic RNNs, namely the failure to capture long-term dependencies due to the vanishing and exploding gradient problems encountered during backpropagation through time (BPTT). Learning Outcomes:

• Define structured data (sequences) and explain why standard Feedforward Networks (FNNs) are inadequate for modeling sequential dependencies.
• Describe the fundamental architecture of a basic Recurrent Neural Network (RNN), identifying the components like the hidden state and shared weight matrices.
• Illustrate the process of 'unfolding' an RNN computation graph over time steps and discuss how input sequences of variable length are handled.
• Explain the mechanism of Backpropagation Through Time (BPTT) and analyze the vanishing and exploding gradient problems inherent in training traditional RNNs.

Overview: This lesson provides a deep dive into the paradigm shift introduced by the "Attention Is All You Need" paper, moving sequence modeling beyond Recurrent Neural Networks (RNNs) by eliminating recurrence and solely relying on attention. We will first establish the mathematical foundation of the Attention Mechanism, specifically focusing on the Scaled Dot-Product Attention using Query (Q), Key (K), and Value (V) vectors. The lecture then expands this concept into the Multi-Head Attention mechanism, explaining its role in capturing diverse contextual dependencies. The core focus will be on the complete Transformer architecture, analyzing the structure of both the Encoder and Decoder stacks, including crucial elements like Residual Connections, Layer Normalization, and the essential Positional Encoding required to maintain sequential information. Finally, we examine how the Transformer enables significant parallelization and its revolutionary impact on fields like Neural Machine Translation and pre-trained language models. Learning Outcomes:

• Define the purpose of attention mechanisms and explain how they resolve the limitations (e.g., long-range dependencies, sequential processing bottleneck) of Recurrent Neural Networks.
• Detail the mathematical operation of Scaled Dot-Product Attention, accurately identifying the roles of Query, Key, and Value vectors.
• Describe the overall structure of the Transformer model, differentiating between the Encoder and Decoder stacks and explaining the function of Multi-Head Attention and Feed-Forward Networks.
• Explain the necessity and mathematical implementation of Positional Encoding within the permutation-invariant Transformer architecture.
• Analyze the computational benefits (parallelization) and widespread applicability of the Transformer architecture in modern Deep Learning tasks, referencing models like BERT and GPT.

Overview: This lecture dives into foundational and applied aspects of Deep Learning for Natural Language Processing (NLP). We begin by addressing the crucial need for effective word representations, transitioning from sparse methods to dense, learned word embeddings. The core mechanisms of Word2Vec (Skip-gram and CBOW) will be explained, highlighting how context generates rich vector representations that capture semantic meaning. We then apply these foundational concepts to two major NLP tasks: Neural Machine Translation (NMT), utilizing sequence-to-sequence encoder-decoder architectures and the critical role of Attention Mechanisms in handling long dependencies and alignment; and Automated Speech Recognition (ASR), exploring how deep models handle temporal sequences of acoustic data to generate textual output. The discussion will emphasize how embeddings and sequential deep learning architectures form the backbone of modern commercial NLP systems. Learning Outcomes:

• Explain the limitations of sparse word representations (e.g., one-hot encoding) and justify the necessity of dense word vector embeddings.
• Describe the fundamental principles and architecture of models like Word2Vec (Skip-gram/CBOW) used for learning distributed representations.
• Outline the core components (Encoder, Decoder, Attention) of a modern Neural Machine Translation system, contrasting it with traditional methods.
• Analyze the challenges inherent in sequence-to-sequence tasks like NMT and Automated Speech Recognition (ASR), particularly concerning variable input/output lengths.
• Identify how neural architectures are adapted to handle audio input in the context of Automated Speech Recognition.

Overview: This lesson introduces the two cornerstone modern deep generative models: Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). We begin with VAEs, detailing their architecture—an encoder mapping data to a parameterized latent distribution and a decoder generating samples. A strong emphasis will be placed on the underlying mathematics, particularly the Evidence Lower Bound (ELBO) objective function, dissecting the roles of the reconstruction loss and the KL divergence term for regularization. The critical Reparameterization Trick, necessary for enabling gradient flow through the sampling process, will be explained thoroughly. We then transition to GANs, defining the adversarial, zero-sum game between the generator (G) and discriminator (D). The lecture covers the theoretical minimax value function, explores how the optimal discriminator maximizes the objective, and discusses major practical challenges such as mode collapse and training instability. Finally, we provide a qualitative comparison, contrasting VAEs' interpretable latent space with GANs' generally superior sample fidelity. Learning Outcomes:

• Differentiate between discriminative and generative modeling and explain the mathematical goal of learning complex data distributions.
• Explain the architecture of a Variational Autoencoder (VAE) and derive the Evidence Lower Bound (ELBO) objective function.
• Analyze the necessity and function of the reparameterization trick in VAE training to ensure effective backpropagation.
• Describe the training process of a Generative Adversarial Network (GAN) as a minimax game between the Generator and Discriminator.
• Compare and contrast VAEs and GANs based on sample quality, latent space interpretability, and common training challenges like mode collapse.

Overview: This lesson introduces Deep Reinforcement Learning (DRL) by establishing the foundational decision-making framework, the Markov Decision Process (MDP). We will define the agent-environment loop, state and action spaces, and the goal of maximizing expected discounted return. Core concepts of traditional RL will be covered, including Value Functions and the Bellman Optimality Equation. The lesson then transitions to DRL, exploring the challenges of large state spaces and how Deep Q-Networks (DQN) overcome this by using neural networks for Q-function approximation. We will detail stability techniques essential for DQN, such as experience replay and target networks. Finally, we contrast value-based methods with Policy Gradient techniques, detailing the mathematical intuition behind the REINFORCE algorithm for direct policy optimization and setting the stage for more advanced Actor-Critic architectures. Learning Outcomes:

• Formalize sequential decision-making problems using the Markov Decision Process (MDP) framework, including definitions of state, action, reward, and the value function.
• Explain the transition from tabular Q-learning to Deep Q-Networks (DQN) and identify the critical techniques (experience replay, target networks) used to stabilize DRL training.
• Differentiate fundamentally between value-based methods (like DQN) and policy-based methods (like REINFORCE).
• Describe the objective function and mathematical intuition behind the Policy Gradient Theorem and its implementation in the REINFORCE algorithm.
• Contrast the applications of value-based versus policy-based approaches in modern Deep RL scenarios.

Overview: This lesson introduces advanced deep learning paradigms necessary for robust deployment and addresses critical societal implications. We first explore the theoretical foundations and practical applications of Unsupervised Deep Learning, focusing on models like Autoencoders and Generative Models when used for representation learning and anomaly detection. Subsequently, we delve into Semi-Supervised Learning (SSL) techniques, such as pseudo-labeling and consistency regularization (e.g., \Pi-Model, MixMatch), which are crucial for leveraging large amounts of unlabeled data alongside scarce labeled examples. The second major part of the lesson critically examines Ethical AI, detailing how data curation and architectural choices introduce Algorithmic Bias. We define and analyze key Fairness metrics (e.g., Equal Opportunity Difference, Demographic Parity) and discuss effective mitigation strategies, emphasizing the importance of model interpretability (XAI) and accountability in high-stakes deep learning systems. Learning Outcomes:

• Distinguish between unsupervised, semi-supervised, and standard supervised learning and identify real-world scenarios appropriate for each paradigm.
• Describe the function and architecture of key unsupervised models, such as Autoencoders and their use in dimensionality reduction or representation learning.
• Explain the methodology of modern semi-supervised techniques, including the concepts of pseudo-labeling and consistency regularization.
• Identify and categorize the primary sources of algorithmic bias introduced during the deep learning lifecycle (data acquisition, modeling, deployment).
• Define and compare common algorithmic fairness metrics (e.g., Equalized Odds) and discuss the trade-offs inherent in bias mitigation strategies.