Convolutional Neural Networks

Introduction

1. Lecture notes

2. Jupyter notebook

Convolutions and computer vision

1. Lecture notes

Basic classification architectures

Object Detection

Semantic segmentation

21st of April: there will be no lecture, just your coding and my help if necessary

Metric Learning

May 5th: there will be no lecture, just your coding and my help if necessary

May 12th: no lecture, no coding: this is a free day at the university

May 19th: lecture on Plotly + Dash with examples on geographic data

1. Notebook

2. Data

May 26th: there will be no lecture, just your coding and my help if necessary

June 2nd: Introduction to AWS SageMaker

June 9th: there will be no lecture, just your coding and my help if necessary

Image Classification
Use one of the architectures from point 3. above to train a model on an own set of images. Change the last layer so that it classifies 5 classes (or another small number of images in your dataset), and freeze all other layers. Observe the final results, illustrate error rates for example in the form of a confusion matrix.

Object Detection

1. Use a ready Yolo model but implement your own version of its loss function.

2. Implement a version of the IoU metric for bounding boxes giving match or no match results for a predefined threshold.

3. Use this implementation to track model results for different thresholds between 0.5 and 1.0.

4. Track (through TensorBoard or simple logging) all parts of the loss function.

Metric Learning

1. Apply either Triplet Loss or SupCon Loss to the MNIST dataset

2. Consider all pairs of images in the test set, label them as either positive or negative and:

   1. Calculate False Match Rate and False Non-Match Rate

   2. Calculate the distance (or similarity for cosine function) distribution and illustrate it as either 2 histograms or 2 pairs of expected values and standard deviations

   3. Calculate the Detection Error Tradeoff (DET) curve