Exam DLNN 2019/2020: INSTRUCTIONS Wojtek Kowalczyk |Deep Learning

w.j.kowalczyk@liacs.leidenuniv.nl Your (individual!) task is to prepare 14 “multiple choice – single answer” quiz questions (or small problems), each with 4 potentialy correct answers. Next, you have to provide correct answers to your questions. To edit your quiz, use the attached template. You can get an idea of how such a test should look like by studying the text of exams with answers that have been used in the earlier editions of this course (years 2018 and 2019). In total you should prepare 14 questions/problems with solutions (2 questions per topic – during the course we have covered 7 topics that are specified below). Evaluation Criteria: When grading your exams, we will take into account the following 4 criteria:  completeness: are all 7 topics tested?  clarity: are questions well formulated?  correctness: are the offered answers/solutions correct?  difficulty: aren’t the questions too easy or too difficult? As a guideline, assume that a good student should be able to solve your quiz in about 90 minutes. How shell we grade your work? For each question-answers-solution you will get:  9.0 if your solution is correct and the “question-choice answers” part can be directly used in a real exam,  8.0-7.0 if your “question-choice answers” part is suitable for a real exam, but the solution part requires some extra work/rephrasing/corrections,  6.0 suitable for a real exam after essential corrections,  <6.0 otherwise (e.g., your solution part is wrong, questions are vague, etc.). The final grade for this exam is the average all 14 grades, rounded to a multiple of 0.5. Exceptionally good submissions we will be rewarded by a bonus of 0.5 or 1.0. Format You may use any format that is compatible with LibreOffice or Microsoft Word (no pdf!). Your work (exam), together with the signed “Declaration of originality” should be submitted via the Blackboard in the same way you submit your assignments. Originality Your work should be original and individual! To get an inspiration, you are allowed to search the internet, study textbooks, blogs, etc., but you have to come up with your original questions, answers and solutions! Keep in mind that according to general rules of the Leiden University, we are obliged to report the suspected cases of plagiarism (e.g., direct copies from textbooks/internet/exams of other students, etc.) to the Board of Examiners. Additionally, when submitting your exam, you should also attach the signed “Declaration of originality” (attached). The deadline for submitting your exam: Tuesday, June, 2nd, 23:59. TOPICS 1. Single Layer Perceptron (slides lecture 2): single perceptron, linear separability, perceptron trainig algorithm,Cover’s Theorem, multiclass perceptron, overfitting and regularization. 2. MLP, SGD, Backpropagation, alternative loss functions (slides lecture 3) 3. Backpropagation in depth (Chapter 11 and slides lecture 4) – vanishing & exploding gradients – initialization strategies – alternative activation functions – optimization algorithms – batch normalization 4. CNNs for Computer Vision (Chapter 14, slides Lecture 5, Kaltura) – key concepts and terminology (receptive field, convolution, filter, pooling, padding, strides), – key CNNs architectures (LeNet, AlexNet(!), GoogleNet, VGGNet, ResNet(!), – related tasks: transfer learning, classification and localization, object detection, fully convolutional networks, semantic segmentation (Chapter 14, pp. 481-496 (not covered during the course!)) 5. Recurrent Neural Networks (Slides lecture 9, Kaltura, Chapter 15) – modelling sequential data: typical tasks and scenario’s – plain RNN and backpropagation through time – LSTM and GRU networks – word2vec embeddings (slides) 6. Autoencoders and GANs (Chapter 17, A3, slides lectures 11, 12, Kaltura) – the key idea behind autoencoders – autoencoders (deep, convolutional, sparse autoencoders, variational) – applications – GANs: the key idea and applications 7. Deep Learning for Reinforcement Learning (Slides lecture 10, Kaltura, Chapter 18 (related parts)) – the key concepts – details of DQN for Atari games

Leave a Reply

Your email address will not be published. Required fields are marked *