• ES Español
• EN English

5.41. Deep Learning (Mandatory)

5.41.1. Justification ↑ Back to top

This course covers modern deep learning fundamentals, including convolutional networks, transformer architectures, and training techniques for advanced AI systems.

5.41.2. Generales Goals ↑ Back to top

1. Implement deep neural networks using modern frameworks (PyTorch/TensorFlow)
2. Design architectures for specific domains (vision, language, etc.)
3. Analyze ethical limitations of deep models

5.41.3. Contribution to Outcomes ↑ Back to top

AG-C09) Design and Development of Solutions: Designs, implements, and evaluates solutions for complex computing problems. (Usage)

AG-C11) Use of Tools: Applies modern computing tools in problem solving. (Assessment)

5.41.4. Content ↑ Back to top

5.41.4.1. Neural Network Training and Optimization (10 hours) [Skills AG-C09,AG-C11] ↑ Back to top

Bibliography: (Goodfellow et al., 2016b)

Topics

1. Modern backpropagation: computational graphs and automatic differentiation
2. Advanced optimizers: momentum, RMSProp, Adam, and learning-rate schedules

Learning Outcomes

1. Implement a deep neural network from scratch, including its backpropagation pass [Apply]
2. Select and tune an optimizer and its learning-rate schedule for training stability [Apply]

5.41.4.2. Regularization Techniques for Deep Networks (5 hours) [Skills AG-C09,AG-C11] ↑ Back to top

Bibliography: (Goodfellow et al., 2016b)

Topics

1. Dropout regularization
2. Batch normalization

Learning Outcomes

1. Diagnose overfitting in a training run using regularization diagnostics [Analyze]
2. Apply dropout and batch normalization to improve a model's generalization [Apply]

5.41.4.3. Convolutional and Residual Network Architectures (9 hours) [Skills AG-C09,AG-C11] ↑ Back to top

Bibliography: (He et al., 2016)

Topics

1. Deep convolutional neural network architectures
2. Residual connections and ResNets

Learning Outcomes

1. Design a deep CNN architecture for a given vision task [Apply]
2. Adapt a residual architecture to a specific vision domain [Apply]

5.41.4.4. Transformer Architectures and Generative Models (16 hours) [Skills AG-C09,AG-C11] ↑ Back to top

Bibliography: (Vaswani et al., 2017)

Topics

1. Transformers and the attention mechanism
2. Generative models: variational autoencoders (VAEs) and generative adversarial networks (GANs)

Learning Outcomes

1. Train a transformer model on a real dataset [Apply]
2. Analyze and contrast VAE and GAN generative-model architectures [Analyze]

5.41.5. Bibliography ↑ Back to top

Goodfellow, I., Bengio, Y., and Courville, A. (2016b). Deep Learning. MIT Press.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep residual learning for image recognition. CVPR, pages 770–778.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., and Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.