Advanced Applied Deep Learning

Convolutional Neural Networks (CNNs) and Fundamentals

Week 1: Introduction to Convolutional Neural Networks

• Basics of deep learning and neural networks

• Introduction to CNNs and their application in image processing

• Mathematical foundations: Convolutions, padding, stride

• Practical exercise: Implement a basic CNN for image classification

Week 2: CNN Architecture in Detail

• Deeper look into common layers: Convolution, pooling, activation functions

• Understanding feature maps and how CNNs learn spatial hierarchies

• Visualization of CNN feature maps

• Hands-on: Building a deeper CNN and visualizing feature activations

Week 3: Transfer Learning and Pre-trained Models

• Overview of transfer learning and its significance

• Pre-trained models: VGG, ResNet, Inception, and DenseNet

• Fine-tuning strategies and their real-world application

• Hands-on: Using a pre-trained model for fine-tuning on a new dataset

Week 4: Regularization Techniques for CNNs

• Dropout, weight decay, and data augmentation

• Batch normalization and its impact on training

• Hands-on: Implementing dropout and batch normalization for improved model performance

Week 5: Optimization Techniques

• Advanced optimizers: Adam, RMSprop, learning rate schedules

• Training strategies: Early stopping, gradient clipping

• Hands-on: Exploring different optimizers and tuning for CNN training

Week 6: Hyperparameter Tuning

• Tuning convolution layers, number of filters, and kernel sizes

• Tuning depth of CNNs and layer configurations

• Hands-on: Hyperparameter search using grid search and random search

Week 7: Object Detection Fundamentals

• Overview of object detection: Introduction to bounding boxes and localization

• Intersection over Union (IoU), anchor boxes, and region proposals

• Hands-on: Implementing a basic object detection model using CNNs

Mid-Semester Project

• Select a problem from classification or object detection

• Build a full pipeline using CNNs, preprocess data, train, validate, and report results

• Presentation and report submission

Week 8: Advanced CNN Architectures

• Inception networks, ResNeXt, EfficientNet, and feature extraction

• Use of deeper and more complex architectures for better performance

• Hands-on: Implement an advanced CNN architecture

Week 9: Single Shot Detectors (SSD)

• Difference between R-CNN and SSD approaches

• Understanding anchor boxes and multi-scale feature maps in SSD

• Hands-on: Implementing SSD for object detection on a dataset

Week 10: You Only Look Once (YOLO) Framework

• YOLO architecture: End-to-end object detection without region proposals

• YOLO versions (YOLOv3, YOLOv4, YOLOv5) and improvements

• Hands-on: Using YOLOv5 for real-time object detection

Week 11: Mask R-CNN and Instance Segmentation

• Introduction to Mask R-CNN: Adding segmentation to object detection

• Role of region proposal networks in pixel-level predictions

• Hands-on: Implementing Mask R-CNN for instance segmentation tasks

Week 12: Real-time Object Detection and Tracking

• Techniques for real-time detection in videos and streaming data

• Tracking objects across frames: Optical flow, SORT, and DeepSORT

• Hands-on: Real-time object detection and tracking in a video stream

Week 13: Explainability and Visualization Techniques

• Techniques to interpret CNN decisions: Grad-CAM, LIME, SHAP

• Visualizing activations and gradients in object detection models

• Hands-on: Implementing Grad-CAM to explain object detection results

Week 14: Model Optimization – Quantization, Pruning, and Knowledge Distillation