COGNITIVE COMPUTING AND ARTIFICIAL INTELLIGENCE

Nota: Questo insegnamento è erogato in lingua inglese - This course is offered in English

The course provides an integrated and modern approach to the design and development of intelligent systems, by resorting to state-of-the-art technologies and methods from the fields of machine learning, knowledge representation, natural computation, logic, and automated reasoning to solve typical and topical problems in application scenarios such as business intelligence, decision-making support, human-computer interaction, and human-robot interaction. The course provides the theoretical foundations of artificial cognitive systems but is practical and application-oriented. The students will gather hands-on experience with cutting-edge techniques of neuro-symbolic AI preparing them for research or industry roles in AI.

Learning objectives:

Knowledge and understanding

• Understand the key tenets of modern cognitive architectures
• Understand the mechanisms underlying classical, neural, and reinforcement learning methods
• Understand differences in approaches to performing reasoning and inference
• Understand the scope of applicability of pre-trained foundation models, advanced deep learning architectures, and hybrid solutions
  

Applying knowledge and understanding

• Be able to design, develop, and validate systems that learn from heterogeneous data (text, images, audio, temporal sequences)
  
• Be able to design, develop, and query a knowledge graph
  
• Be able to prototype autonomous agents' behaviors capable of interacting adaptively with humans.

Making judgements

• Recognize when a problem is best solved by a machine learning approach or by a symbolic AI  approach
• Recognize the limitations, risks, and ethical implications of deploying Artificial Intelligence technologies

Communication skills

• Present orally in English AI methods and architectures to solve domain-specific problems
• Document software choices in commented Python notebooks
• Explain the motivations behind specific design choices

Learning skills

• Learning to read effectively scientific papers written in English to focus on the key idea and recognize the article's contribution
• Learning to repurpose creatively software artifacts
• Learning to collaborate to solve a problem and prepare a joint report

The course involves frontal lessons, laboratories, and seminars.

Should teaching be carried out in mixed mode or remotely, it might be necessary to introduce changes with respect to previous statements, in line with the programme planned and outlined in the syllabus.

• Proficiency in Python programming
• Basic knowledge of calculus, linear algebra, and probability
• Understanding of basic machine learning concepts
• Familiarity with machine learning libraries (e.g., PyTorch, TensorFlow) is not mandatory but is an asset.
• Attendance in the Deep Learning course in the LM32 curriculum is strongly recommended
  

Strongly recommended. Attending and actively participating in classroom activities will contribute positively to the overall assessment of the oral exam.

Part 1: Knowledge Representation, Reasoning, and Semantic Technologies

• Rational agents and their environments: from reactive agents to autonomous agents. The general SOAR cognitive architecture. Developmental cognitive architectures. The concept of embodied cognition. The problem of knowledge representation. 

• Reasoning: deductive and inductive reasoning; reasoning with uncertainty; case-based reasoning.  Causal inference. Argumentation and explanation.

• Problem-solving: search strategies and optimization. Problem-solving by search vs. Problem-solving by description. Semantic nets and similarity metrics
  

• The Logic approach: First-order logic and logic programming. Other logics: fuzzy and temporal logic. Strengths and limitations.

• Semantic technologies. Knowledge graphs. Ontologies. Query languages for knowledge graphs.

Part 2: Machine learning paradigms

• Deep neural models:  Autoencoders and Variational autoencoders; Generative Adversarial Networks (GANs). Recurrent Neural Networks (RNN);

• The Transformers Architecture and Attention Mechanism. Embedding methods. Pre-trained Large Language Models (LLMs) and the revolution of multimodal foundation models: GPT-3, BERT, CLIP, DALL-E, SORA. Pre-training objectives: masked language modeling, autoregressive modeling.

• Prompting and fine-tuning. Pre-processing text, audio/speech, image/video, and biosignals for use with foundational models. Evaluation metrics for foundation models. Applications in Natural Language Processing: Text generation, summarization, translation, and question answering.  Applications in Computer Vision: Vision transformers in medical imaging. 

• Learning paradigms: Self-supervised Learning, Meta-learning. Continual Learning. Federated learning.

• Reinforcement learning: Markov Decision Processes (MDP). Dynamic programming. Basic RL Algorithms. Deep reinforcement learning (DRL) concepts. Key algorithms: Deep Q-Networks (DQN), Policy Gradient Methods. Proximal Policy Optimization (PPO).
  

• The challenge of Language understanding and knowledge integration. Model hallucinations. Chatbot development with LLMs, Retrieval Augmented Generation, Knowledge Graphs, and Reinforcement Learning.

• Explainable AI. Methods for making the models interpretable.

• Integrating Symbolic and Sub-symbolic AI.
  

• Ethical considerations in AI. Addressing bias, fairness, and accountability. Case studies on ethical issues in AI.  Bias in AI systems: Identification and mitigation strategies.
  

Part 3: Autonomous agents and the Reachy humanoid robotic platform

• Theories of perception, action, and interaction. Interactive autonomous agents. Human-robot interaction

• The Reachy anthropomorphic robot  and simulation environment

• Applications. Augmenting the Reachy perceptual and cognitive system.

Selected chapters from the following resources:

1. Artificial Cognitive Systems: A Primer.  David Vernon, MIT Press, 2014
2. Artificial intelligence: a modern approach. Stuart Russell, Peter Norvig, 3rd edition, 2010
3. Deep Learning. I. Goodfellow, Y. Bengio and A. Courville, MIT Press, 2016
4. Teaching materials provided by the instructor

The competencies to be developed by the students will be tested by an oral exam consisting of the discussion of project work (60% of the final grade), presentation of a research article (10% of the grade), and 3 questions on key concepts and methodologies covered in the course (30% of the grade).  Assessment criteria of the project work include depth of analysis, adequacy, correctness, and originality. Assessment criteria of the oral include the ability to justify and critically evaluate the technological solutions adopted in the project/homework, and clarity.

Examples of questions and projects are available in Studium