Summary of Study | Summary of Branches | All Subject Groups | All Subjects | List of Roles | Explanatory Notes               Instructions

---

B3B33UROB	Robot Learning
Roles:	PV	Extent of teaching:	2P+2C
Department:	13133	Language of teaching:	CS
Guarantors:	Zimmermann K.	Completion:	Z,ZK
Lecturers:	Zimmermann K.	Credits:	6
Tutors:	Too many persons	Semester:	Z

Web page:

https://cw.fel.cvut.cz/wiki/courses/b3b33urob/start

Anotation:

The course teaches deep learning methods on known robotic problems, such as semantic segmenation or reactive motion control. The overall goal is timeless universal knowledge rather than listing all known deep learning architectures. Students are assumed to have working prior knowledge of mathematics (gradient, jacobian, hessian, gradient descend, taylor polynomial) and machine learning (bayes risk minimization, linear classifier). The labs are divided into two parts, in the first one, the students will solve elementary deep ML tasks from scratch (including the reimplementation of autograd backpropagation), in the second one, students will build on existing templates in order to solve complex tasks including RL, tranformers and generative networks.

Course outlines:

Machine learning 101: model, loss, learning, issues, regression, classification Under the hood of a linear classifier: two-class and multi-class linear classifier on RGB images Under the hood of auto-differentiation: Computational graph of fully connected NN, Vector-Jacobian-Product (VJP) vs chainrule and multiplication of Jacobians. The story of the cat's brain surgery: cortex + convolutional layer and its Vector-Jacobian-Product (VJP) Where the hell does the loss come from? MAP and ML estimate, KL divergence and losses. Why is learning prone to fail? - Structural issues: layers + issues, batch-norm, drop-out Why is learning prone to fail? - Optimization issues: optimization vs learning, KL divergence, SGD, momentum, convergence rate, Adagrad, RMSProp, AdamOptimizer, diminishing/exploding gradient, oscillation, double descent What can('t) we do with a deep net?: Classification (ResNet, Squeeze and Excitation Nets), Segmentation (DeepLab), Detection (Yolo, fast-RCNN), Regression (OpenPose), Spatial Transformer Nets, Reinforcement learning: Approximated Q-learning, DQN, DDPG, Derivation of the policy gradient (REINFORCE), A2C, TRPO, PPO, Reward shaping, Inverse RL, Applications, Memory and attention: recurrent nets, Image transformers with attention module Generative models: GANs and diffusion models Implicit layers: Backpropagation through unconstrained and constrained optimization problems, ODE solvers, roots, fixed points) + existing end-to-end differentiable modules cvxpy, gradSLAM, gradMPC, gradODE, pytorch3d

Exercises outline:

The labs are divided into two parts, in the first one, the students will solve elementary deep ML tasks from scratch (including the reimplementation of autograd backpropagation), in the second one, students will build on existing templates in order to solve complex tasks including RL, tranformers and generative networks.

Literature:

Ian Goodfellow and Yoshua Bengio and Aaron Courville, Deep learning, MIT press, 2016 http://www.deeplearningbook.org

Requirements:

Keywords:

machine learning, deep neural networks

Subject is included into these academic programs:

Program	Branch	Role	Recommended semester
BPKYR_2021	Common courses	PV	5

---

Page updated 26.4.2024 12:51:24, semester: Z/2024-5, Z,L/2023-4, Send comments about the content to the Administrators of the Academic Programs

Proposal and Realization: I. Halaška (K336), J. Novák (K336)