In this module, you will learn how to examine a vision dataset systematically before training a model. You will analyze class distribution, image statistics, data quality, and deployment gaps to understand what your dataset supports and where it may fail in production. You will use those findings to choose an appropriate model family and define a preprocessing pipeline grounded in dataset size, image properties, and quality issues rather than assumptions. By the end of the module, you will be able to turn dataset analysis into concrete modeling decisions that reduce debugging time and improve downstream performance.

In this module, you will learn how to use augmentation as a strategic tool for expanding dataset diversity and improving model generalization. You will explore core augmentation techniques across geometric, color, noise, blur, and composition-based transformations, and you will evaluate each one through the lens of semantic validity. You will learn how to select and combine augmentations based on dataset gaps, class imbalance, and real deployment conditions, while correctly scoping augmentation to the training set only. By the end of the module, you will be able to design an augmentation pipeline that is purposeful, domain-aware, and aligned with what your model needs to learn.

You’ll turn a trained vision model into a usable service. You’ll standardize inputs/outputs, containerize the app, and expose /predict that returns class names and confidence scores as JSON. By the end, you’ll have a reproducible, testable inference pipeline aligned with real engineering needs.

You will evaluate deployed vision models using metrics and error analysis. You will compute task-specific measures such as mean Average Precision (mAP) and segment errors by condition (e.g., low-light vs. daytime). You will apply this analysis to diagnose failure modes, document causes, and recommend next steps—strengthening your ability to balance performance reporting with actionable insight. By the end, you will know how to turn raw metrics into meaningful narratives that guide improvement and communicate reliability.

You’ll explore the fundamentals of deep learning stability, why models diverge, overfit, or fail to converge, and how to fix them. You’ll practice using weight initialization, normalization, and regularization to stabilize a segmentation model. Along the way, you’ll use TensorBoard to interpret gradient norms and identify vanishing gradients before they derail your training.

You will explore how gradients behave during deep neural network training. You will analyze gradient-norm plots, activation distributions, and loss curves to diagnose issues like vanishing and exploding gradients. Through videos, discussions, and a hands-on lab, you will learn to interpret training signals and apply architectural and activation-based fixes. By the end, you will be able to identify instability in training and recommend targeted solutions to stabilize model performance.

You will explore how to design, run, and interpret ablation studies that isolate the real impact of design decisions in AI models. You will practice structuring controlled experiments, evaluating model variations, and interpreting results statistically to distinguish meaningful improvements from noise. Through guided reflection, readings, videos, and hands-on experimentation, you will develop the discipline of evidence-based model evaluation.

You will focus on reproducibility in AI research—ensuring that results are not just impressive once, but repeatable by anyone, anywhere. You will design end-to-end workflows that lock randomness, manage configurations, version data, and document experiments clearly. Instead of a traditional lab, you will complete a Final Project, combining everything from both lessons—running controlled experiments and implementing a reproducible pipeline.