• ES Español
• EN English

5.65. Topics in Artificial Intelligence (Mandatory)

5.65.1. Justification ↑ Back to top

Artificial Intelligence is a rapidly evolving field. This course explores advanced and contemporary topics that are currently shaping the industry and research. Focus is placed on Generative AI, Large Language Models (LLMs), and Advanced Reinforcement Learning. Students will analyze the architectures behind modern AI systems, such as Transformers, and discuss the technical and societal challenges of deploying autonomous intelligent agents at scale.

5.65.2. Generales Goals ↑ Back to top

1. Understand the architecture and training process of Large Language Models.
2. Master the principles of Generative AI for text and image synthesis.
3. Implement advanced Reinforcement Learning algorithms for decision making.
4. Apply prompt engineering and fine-tuning techniques to existing models.
5. Critically evaluate the latest research papers and trends in AI.

5.65.3. Contribution to Outcomes ↑ Back to top

AG-C06) Life-long Learning: Recognizes the importance of continuous learning and adaptation to new technologies. (Usage)

5.65.4. Content ↑ Back to top

5.65.4.1. Machine Learning Fundamentals (6 hours) [Skills AG-C06] ↑ Back to top

Bibliography: (Goodfellow et al., 2016a; Foster, 2022)

Topics

1. Definition and examples of a broad variety of machine learning tasks:
2. Supervised learning:
3. Classification
4. Regression enumerate
5. Reinforcement learning
6. Unsupervised learning:
7. Clustering enumerate enumerate
8. Fundamental ideas:
9. No free lunch theorem: no one learner can solve all problems; representational design decisions have consequences.
10. Sources of error and undecidability in machine learning enumerate
11. A simple statistical-based supervised learning such as linear regression or decision trees:
12. Focus on how they work without going into mathematical or optimization details; enough to understand and use existing implementations correctly enumerate
13. The overfitting problem/controlling solution complexity (regularization, pruning - intuition only):
14. The bias (underfitting) - variance (overfitting) tradeoff enumerate

Learning Outcomes

1. Describe the differences among the three main styles of learning (supervised, reinforcement, and unsupervised) and determine which is appropriate to a particular problem domain [Explain]
2. Differentiate the terms of AI, machine learning, and deep learning [Evaluate]
3. Frame an application as a classification problem, including the available input features and output to be predicted (e.g., identifying alphabetic characters from pixel grid input) [Apply]
4. Identify overfitting in the context of a problem and learning curves and describe solutions to overfitting [Evaluate]
5. Explain how machine learning works as an optimization/search process [Explain]

5.65.4.2. Machine Learning Data, Representations and Evaluation (5 hours) [Skills AG-C06] ↑ Back to top

Bibliography: (Goodfellow et al., 2016a; Foster, 2022)

Topics

1. Working with Data:
2. Data preprocessing:
3. Importance and pitfalls of preprocessing choices enumerate
4. Handling missing values (imputing, flag-as-missing):
5. Implications of imputing vs flag-as-missing enumerate
6. Encoding categorical variables, encoding real-valued data
7. Normalization/standardization
8. Emphasis on real data, not textbook examples enumerate
9. Representations:
10. Hypothesis spaces and complexity
11. Simple basis feature expansion, such as squaring univariate features
12. Learned feature representations enumerate
13. Machine learning evaluation:
14. Separation of train, validation, and test sets
15. Performance metrics for classifiers
16. Estimation of test performance on held-out data
17. Tuning the parameters of a machine learning model with a validation set
18. Importance of understanding what a model is doing, where its pitfalls/shortcomings are, and the implications of its decisions enumerate

Learning Outcomes

1. Apply two or more simple statistical learning algorithms to a classification task and measure the classifiers' accuracy [Apply]
2. Implement a statistical learning algorithm and the corresponding optimization process to train the classifier and obtain a prediction on new data [Apply]
3. Explain proper ML evaluation procedures, including the differences between training and testing performance, and what can go wrong with the evaluation process leading to inaccurate reporting of ML performance [Explain]
4. Compare two machine learning algorithms on a dataset, implementing the data preprocessing and evaluation methodology (e.g., metrics and handling of train/test splits) from scratch [Apply]

5.65.4.3. Neural Networks and Ethics (5 hours) [Skills AG-C06] ↑ Back to top

Bibliography: (Goodfellow et al., 2016a; Foster, 2022)

Topics

1. Basic neural networks:
2. Fundamentals of understanding how neural networks work and their training process, without details of the calculations
3. Basic introduction to generative neural networks (e.g., large language models) enumerate
4. Ethics for Machine Learning:
5. Focus on real data, real scenarios, and case studies
6. Dataset/algorithmic/evaluation bias and unintended consequences enumerate

Learning Outcomes

1. Describe the neural network training process and resulting learned representations [Explain]

5.65.4.4. Machine Learning Formulation and Evaluation (3 hours) [Skills ] ↑ Back to top

Bibliography: (Goodfellow et al., 2016a)

Topics

1. Formulation of simple machine learning as an optimization problem, such as least squares linear regression or logistic regression:
2. Objective function
3. Gradient descent
4. Regularization to avoid overfitting (mathematical formulation) enumerate
5. Ensembles of models:
6. Simple weighted majority combination enumerate
7. Performance evaluation:
8. Other metrics for classification (e.g., error, precision, recall)
9. Performance metrics for regressors
10. Confusion matrix
11. Cross-validation:
12. Parameter tuning (grid/random search, via cross-validation) enumerate enumerate

Learning Outcomes

1. Compare and contrast several learning techniques (e.g., decision trees, logistic regression, naive Bayes, neural networks, and belief networks), providing examples of when each strategy is superior [Compare]
2. Evaluate the performance of a simple learning system on a real-world dataset [Evaluate]
3. Characterize the state of the art in learning theory, including its achievements and shortcomings [Explain]
4. Explain the problem of overfitting, along with techniques for detecting and managing the problem [Explain]
5. Explain the triple tradeoff among the size of a hypothesis space, the size of the training set, and performance accuracy [Explain]

5.65.4.5. Deep Learning, RL and Ethics (3 hours) [Skills ] ↑ Back to top

Bibliography: (Goodfellow et al., 2016a)

Topics

1. Deep learning:
2. Deep feed-forward networks (intuition only, no mathematics)
3. Convolutional neural networks (intuition only, no mathematics)
4. Visualization of learned feature representations from deep nets
5. Other architectures (generative NN, recurrent NN, transformers, etc.) enumerate
6. Overview of reinforcement learning methods
7. Two or more applications of machine learning algorithms:
8. E.g., medicine and health, economics, vision, natural language, robotics, game play enumerate
9. Ethics for Machine Learning:
10. Continued focus on real data, real scenarios, and case studies
11. Privacy
12. Fairness
13. Intellectual property
14. Explainability enumerate

Learning Outcomes

1. Visualize the training progress of a neural network through learning curves in a well-established toolkit (e.g., TensorBoard) and visualize the learned features of the network [Apply]
2. Given a real-world application of machine learning, describe ethical issues regarding the choices of data, preprocessing steps, algorithm selection, and visualization/presentation of results [Apply]

5.65.4.6. Natural Language Processing (20 hours) [Skills AG-C06] ↑ Back to top

Bibliography: (Tunstall et al., 2022; Russell and Norvig, 2020)

Topics

1. Deterministic and stochastic grammars
2. Parsing algorithms:
3. CFGs and chart parsers (e.g., CYK)
4. Probabilistic CFGs and weighted CYK enumerate
5. Representing meaning/Semantics:
6. Logic-based knowledge representations
7. Semantic roles
8. Temporal representations
9. Beliefs, desires, and intentions enumerate
10. Corpus-based methods
11. N-grams and HMMs
12. Smoothing and backoff
13. Examples of use: POS tagging and morphology
14. Information retrieval:
15. Vector space model:
16. TF & IDF enumerate
17. Precision and recall enumerate
18. Information extraction
19. Language translation
20. Text classification, categorization:
21. Bag of words model enumerate
22. Deep learning for NLP:
23. RNNs
24. Transformers
25. Multi-modal embeddings (e.g., images + text)
26. Generative language models enumerate

Learning Outcomes

1. Define and contrast deterministic and stochastic grammars, providing examples to show the adequacy of each [Define]
2. Simulate, apply, or implement classic and stochastic algorithms for parsing natural language [Apply]
3. Identify the challenges of representing meaning [Analyze]
4. List the advantages of using standard corpora. Identify examples of current corpora for a variety of NLP tasks [List]
5. Identify techniques for information retrieval, language translation, and text classification [Analyze]
6. Implement a TF/IDF transform, use it to extract features from a corpus, and train an off-the-shelf machine learning algorithm using those features to do text classification [Apply]

5.65.4.7. Machine Learning Fundamentals (6 hours) [Skills AG-C06] ↑ Back to top

Bibliography: (Sutton and Barto, 2018)

Topics

1. Definition and examples of a broad variety of machine learning tasks:
2. Supervised learning:
3. Classification
4. Regression enumerate
5. Reinforcement learning
6. Unsupervised learning:
7. Clustering enumerate enumerate
8. Fundamental ideas:
9. No free lunch theorem: no one learner can solve all problems; representational design decisions have consequences.
10. Sources of error and undecidability in machine learning enumerate
11. A simple statistical-based supervised learning such as linear regression or decision trees:
12. Focus on how they work without going into mathematical or optimization details; enough to understand and use existing implementations correctly enumerate
13. The overfitting problem/controlling solution complexity (regularization, pruning - intuition only):
14. The bias (underfitting) - variance (overfitting) tradeoff enumerate

Learning Outcomes

1. Describe the differences among the three main styles of learning (supervised, reinforcement, and unsupervised) and determine which is appropriate to a particular problem domain [Explain]
2. Differentiate the terms of AI, machine learning, and deep learning [Evaluate]
3. Frame an application as a classification problem, including the available input features and output to be predicted (e.g., identifying alphabetic characters from pixel grid input) [Apply]
4. Identify overfitting in the context of a problem and learning curves and describe solutions to overfitting [Evaluate]
5. Explain how machine learning works as an optimization/search process [Explain]

5.65.4.8. Machine Learning Data, Representations and Evaluation (5 hours) [Skills AG-C06] ↑ Back to top

Bibliography: (Sutton and Barto, 2018)

Topics

1. Working with Data:
2. Data preprocessing:
3. Importance and pitfalls of preprocessing choices enumerate
4. Handling missing values (imputing, flag-as-missing):
5. Implications of imputing vs flag-as-missing enumerate
6. Encoding categorical variables, encoding real-valued data
7. Normalization/standardization
8. Emphasis on real data, not textbook examples enumerate
9. Representations:
10. Hypothesis spaces and complexity
11. Simple basis feature expansion, such as squaring univariate features
12. Learned feature representations enumerate
13. Machine learning evaluation:
14. Separation of train, validation, and test sets
15. Performance metrics for classifiers
16. Estimation of test performance on held-out data
17. Tuning the parameters of a machine learning model with a validation set
18. Importance of understanding what a model is doing, where its pitfalls/shortcomings are, and the implications of its decisions enumerate

Learning Outcomes

1. Apply two or more simple statistical learning algorithms to a classification task and measure the classifiers' accuracy [Apply]
2. Implement a statistical learning algorithm and the corresponding optimization process to train the classifier and obtain a prediction on new data [Apply]
3. Explain proper ML evaluation procedures, including the differences between training and testing performance, and what can go wrong with the evaluation process leading to inaccurate reporting of ML performance [Explain]
4. Compare two machine learning algorithms on a dataset, implementing the data preprocessing and evaluation methodology (e.g., metrics and handling of train/test splits) from scratch [Apply]

5.65.4.9. Neural Networks and Ethics (5 hours) [Skills AG-C06] ↑ Back to top

Bibliography: (Sutton and Barto, 2018)

Topics

1. Basic neural networks:
2. Fundamentals of understanding how neural networks work and their training process, without details of the calculations
3. Basic introduction to generative neural networks (e.g., large language models) enumerate
4. Ethics for Machine Learning:
5. Focus on real data, real scenarios, and case studies
6. Dataset/algorithmic/evaluation bias and unintended consequences enumerate

Learning Outcomes

1. Describe the neural network training process and resulting learned representations [Explain]

5.65.4.10. AI Applications and Societal Impact (6 hours) [Skills AG-C06] ↑ Back to top

Bibliography: (Russell and Norvig, 2020; Bostrom, 2014)

Topics

1. At least one application of AI to a specific problem and field, such as medicine, health, sustainability, social media, economics, education, robotics, etc. (choose at least one for the CS Core):
2. Formulating and evaluating a specific application as an AI problem:
3. How to deal with underspecified or ill-posed problems enumerate
4. Data availability/scarcity and cleanliness:
5. Basic data cleaning and preprocessing
6. Data set bias enumerate
7. Algorithmic bias
8. Evaluation bias
9. Assessment of societal implications of the application enumerate
10. Deployed deep generative models:
11. High-level overview of deep image generative models (e.g., as of 2023, DALL-E, Midjourney, Stable Diffusion, etc.), their uses, and their shortcomings/pitfalls.
12. High-level overview of large language models (e.g., as of 2023, ChatGPT, Bard, etc.), their uses, and their shortcomings/pitfalls. enumerate
13. Overview of societal impact of AI:
14. Ethics
15. Fairness
16. Trust/explainability
17. Privacy and usage of training data
18. Human autonomy and oversight/regulations/legal requirements
19. Sustainability enumerate

Learning Outcomes

1. Given a real-world application domain and problem, formulate an AI solution to it, identifying proper data/input, preprocessing, representations, AI techniques, and evaluation metrics/methodology [Apply]

5.65.4.11. Additional AI Applications and Societal Impact (6 hours) [Skills AG-C06] ↑ Back to top

Bibliography: (Russell and Norvig, 2020; Bostrom, 2014)

Topics

1. One or more additional applications of AI to a broad set of problems and diverse fields, such as medicine, health, sustainability, social media, economics, education, robotics, etc. (choose a different area from that chosen for the CS Core):
2. Formulating and evaluating a specific application as an AI problem:
3. How to deal with underspecified or ill-posed problems enumerate
4. Data availability/scarcity and cleanliness:
5. Basic data cleaning and preprocessing
6. Data set bias enumerate
7. Algorithmic bias
8. Evaluation bias
9. Assessment of societal implications of the application enumerate
10. Additional depth on deployed deep generative models:
11. Introduction to how deep image generative models work, (e.g., as of 2023, DALL-E, Midjourney, Stable Diffusion) including discussion of attention
12. Introduction to how large language models work, (e.g., as of 2023, ChatGPT, Bard) including discussion of attention
13. Idea of foundational models, how to use them, and the benefits/issues with training them from big data enumerate
14. Analysis and discussion of the societal impact of AI:
15. Ethics
16. Fairness
17. Trust/explainability
18. Privacy and usage of training data
19. Human autonomy and oversight/regulations/legal requirements
20. Sustainability enumerate

Learning Outcomes

1. Analyze the societal impact of one or more specific real-world AI applications, identifying issues regarding ethics, fairness, bias, trust, and explainability [Evaluate]
2. Describe some of the failure modes of current deep generative models for language or images, and how this could affect their use in an application [Describe]

5.65.5. Bibliography ↑ Back to top

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

Foster, D. (2022). Generative Deep Learning. O'Reilly Media, 2nd edition.

Tunstall, L., von Werra, L., and Wolf, T. (2022). Natural Language Processing with Transformers. O'Reilly Media.

Russell, S. and Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Pearson, 4th edition.

Sutton, R. S. and Barto, A. G. (2018). Reinforcement Learning: An Introduction. MIT Press, 2nd edition.

Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.