6 ECTS; 3º Ano, 1º Semestre, 28,0 T + 28,0 TP + 5,0 OT , Cód. 814359.
Lecturer
- Sandra Maria Gonçalves Vilas Boas Jardim (1)(2)
(1) Lead Professor
(2) Teaching Professor
Prerequisites
It is essential that students have a solid understanding of:
- Linear Algebra and Calculus
- Python programming language
- Machine Learning
Objectives
The central objectives are to provide students with a solid theoretical understanding and practical skills in the area of ??Deep Learning.
Students should:
1) Understand the mathematical and algorithmic fundamentals of neural networks (Perceptrons, Backpropagation, Optimization);
2) Design and implement complex architectures, including Convolutional Networks (CNNs) for vision, Recurrent Networks (RNNs) and Transformers for sequences, and Reinforcement Learning (RL) algorithms;
3) Diagnose training problems (overfitting/underfitting) and apply regularization strategies;
4) Use modern frameworks (PyTorch) to solve real-world problems.
Program
1. History of AI/ML/DL: Biological vs. Artificial Neuron,
Perceptron and Activation Functions (Sigmoid, Tanh, ReLU)
2. Feedforward Networks: MLPs, Loss Functions (MSE, Cross-Entropy), Backpropagation and Gradient Descent; 1. Gradient Variants (SGD, Adam, RMSProp), Learning Rate, Dropout, Early Stopping, and Batch/Mini-batch
2. Convolutions, Pooling, Feature Maps, and Modern Architectures (VGG, ResNet, Inception, DenseNet)
3. Limitations of MLPs, RNNs, LSTMs, GRUs, and Introduction to Attention Mechanisms and Transformers (ViT)
4. Agents, Environments, Rewards, Q-Learning, Deep Q-Networks (DQN), and Policy Gradients
5. Neural Architecture (NAS) Research and Multimodal Fusion.
Evaluation Methodology
Assessment Period
- Written Exam: 50%
- Practical Work: 50%
Exam Period
- Written Exam: 50%
- Practical Work: 50%
In all assessment periods, the evaluation of the practical work involves its presentation and discussion, with this discussion accounting for 50% of the practical work grade.
The final grade results from the weighted average of the grades obtained in the assessment components defined for each assessment period. Students must obtain a grade of 10 or higher to pass the course.
To pass the course, students cannot have a grade lower than 8 in any of the defined assessment components.
Bibliography
- Goodfellow, I. e Bengio, Y. e Courville, A. (2016). Deep Learning. (Vol. 1). (pp. 1---). USA: MIT Press
Teaching Method
Theoretical typology classes: presentation of concepts inherent to the program content; discussion of examples and practical cases.
Theoretical-practical classes: focused on solving practical problems.
Software used in class
Google Colab interactive programming platform
