BSc: Practical Machine Learning Deep Learning

Practical Machine Learning and Deep Learning

• Course name: Practical Machine Learning and Deep Learning
• Course number: PMLDL-04

Course Characteristics

Key concepts of the class

• Practical aspects of deep learning (DL)
• Practical applications of DL in Natural Language Processing, Computer Vision and generation.

What is the purpose of this course?

The course is about the practical aspects of deep learning. In addition to frontal lectures, the flipped classes and student project presentations will be organized. During lab sessions the working language is Python. The primary framework for deep learning is PyTorch. Usage of TensorFlow and Keras is possible, usage of Docker is highly appreciated.

Prerequisites

• CSE202 — Analytical Geometry and Linear Algebra I / []: Manifolds "Linear Alg./Calculus: Manifolds
• CSE203 — Mathematical Analysis II: Basics of optimisation
• CSE201 — Mathematical Analysis I: integration and differentiation.
• CSE103 — Theoretical Computer Science: Graph theory basics, Spectral decomposition.
• CSE206 — Probability And Statistics: Multivariate normal dist.
• CSE504 — Digital Signal Processing: convolution, cross-correlation"

Course objectives based on Bloom’s taxonomy

The course focuses on the following outcomes:

• Students should be able to apply deep learning methods to effectively solve practical (real-world) problems;
• Students should be able to work in data science team; to understand of principles and a lifecycle of data science projects.

- What should a student remember at the end of the course?

By the end of the course, the students should be able

• to apply deep learning methods to effectively solve practical (real-world) problems;
• to work in data science team;
• to understand of principles and a lifecycle of data science projects.

- What should a student be able to understand at the end of the course?

By the end of the course, the students should be able

• to understand modern deep NN architectures;
• to compare modern deep NN architectures;
• to create a prototype of a data-driven product.

- What should a student be able to apply at the end of the course?

By the end of the course, the students should be able

• to apply techniques for efficient training of deep NNs;
• to apply methods for data science team organisation;
• to apply deep NNs in NLP and computer vision.

Course evaluation

Course grade breakdown

		Proposed points
Labs/seminar classes	20
Interim performance assessment	30
Exams	50

If necessary, please indicate freely your course’s features in terms of students’ performance assessment.

Grades range

Course grading range

		Proposed range
A. Excellent	90-100
B. Good	75-89
C. Satisfactory	60-74
D. Poor	0-59

Resources and reference material

• Goodfellow et al. Deep Learning, MIT Press. 2017
• Géron, Aurélien. Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems. 2017.
• Osinga, Douwe. Deep Learning Cookbook: Practical Recipes to Get Started Quickly. O’Reilly Media, 2018.

Course Sections

The main sections of the course and approximate hour distribution between them is as follows:

Course Sections

Section	Section Title	Teaching Hours
1	Review. CNNs and RNNs	12 hours of lectures, 12 hours of labs
2	Team Data Science Processes	6 hours of lectures, 6 hours of labs
3	VAE, GANs	8 hours of lectures, 8 hours of labs

Section 1

Section title:

Review. CNNs and RNNs

Topics covered in this section:

• Image processing, FFNs, CNNs
• Training Deep NNs
• RNNs, LSTM, GRU, Embeddings
• Bidirectional RNNs
• Seq2seq
• Encoder-Decoder Networks
• Attention
• Memory Networks

What forms of evaluation were used to test students’ performance in this section?

|a|c| & Yes/No
Development of individual parts of software product code & 1
Homework and group projects & 1
Midterm evaluation & 1
Testing (written or computer based) & 1
Reports & 0
Essays & 0
Oral polls & 0
Discussions & 1

Typical questions for ongoing performance evaluation within this section

1. Suppose you use Batch Gradient Descent and you plot the validation error at every epoch. If you notice that the validation error consistently goes up, what is likely going on? How can you fix this?
2. Is it a good idea to stop Mini-batch Gradient Descent immediately when the validation error goes up?
3. List the optimizers that you know (except SGD) and explain one of them
4. Describe Xavier (or Glorot) initialization. Why do you need it?

Typical questions for seminar classes (labs) within this section

1. Name advantages of the ELU activation function over ReLU.
2. Can you name the main innovations in AlexNet, compared to LeNet-5? What about the main innovations in GoogLeNet and ResNet?
3. What is the difference between LSTM and GRU cells?

Test questions for final assessment in this section

1. Explain what the Teacher Forcing is.
2. Why do people use encoder–decoder RNNs rather than plain sequence-to-sequence RNNs for automatic translation?
3. How could you combine a convolutional neural network with an RNN to classify videos?

Section 2

Section title:

Team Data Science Processes

Topics covered in this section:

• Team Data Science Processes
• Team Data Science Roles
• Team Data Science Tools (MLFlow, KubeFlow)
• CRISP-DM
• Productionizing ML systems

What forms of evaluation were used to test students’ performance in this section?

|a|c| & Yes/No
Development of individual parts of software product code & 1
Homework and group projects & 1
Midterm evaluation & 1
Testing (written or computer based) & 1
Reports & 0
Essays & 0
Oral polls & 0
Discussions & 1

Typical questions for ongoing performance evaluation within this section

1. What is CRISP-DM?
2. What is TDSP?
3. How to use MLflow?
4. What is TensorBoard?
5. How to apply Kubeflow in practice?

Typical questions for seminar classes (labs) within this section

1. Explain issues in distributed learning of deep NNs.
2. How do you organize your data science project?
3. Recall a checklist for organization of a typical data science project.

Test questions for final assessment in this section

1. Can you explain what it means for a company to be ML-ready?
2. What a company can do to become ML-ready / Data driven?
3. Can you list approaches to structure DS-teams? Discuss their advantages and disadvantages.
4. Can you list and define typical roles in a DS team?
5. What do you think about practical aspects of processes and roles in Data Science projects/teams?

Section 3

Section title:

VAEs, GANs

Topics covered in this section:

• Autoencoders
• Variational Autoencoders
• GANs, DCGAN

What forms of evaluation were used to test students’ performance in this section?

|a|c| & Yes/No
Development of individual parts of software product code & 1
Homework and group projects & 1
Midterm evaluation & 1
Testing (written or computer based) & 1
Reports & 0
Essays & 0
Oral polls & 0
Discussions & 1

Typical questions for ongoing performance evaluation within this section

1. What is an Autoencoder? Can you list the structure and types of Autoencoders?
2. Can you describe ways to train Stacked AEs?
3. What is Denoising AE? Can you describe what is sparsity loss and why it can be useful?
4. Can you make a distinction between AE and VAE?

Typical questions for seminar classes (labs) within this section

• If an autoencoder perfectly reconstructs the inputs, is it necessarily a good autoencoder? How can you evaluate the performance of an autoencoder?
• How do you tie weights in a stacked autoencoder? What is the point of doing so?
• What about the main risk of an overcomplete autoencoder?
• How the loss function for VAE is defined? What is ELBO?
• Can you list the structure and types of a GAN?
• How would you train a GAN?
• How would you estimate the quality of a GAN?
• Can you describe cost function of a Discriminator?

Test questions for final assessment in this section

• Can you make a distinction between Variational approximation of density and MCMC methods for density estimation?
• What is DCGAN? What is its purpose? What are main features of DCGAN?
• What is your opinion about Word Embeddings? What types do you know? Why are they useful?
• How would you classify different CNN architectures?
• How would you classify different RNN architectures?
• Explain attention mechanism. What is self-attention?
• Explain the Transformer architecture. What is BERT?