Responsible Generative AI
Programma
This course is organized into three modules, supported by student seminars to foster critical engagement. The first module introduces foundational concepts of generative AI, including a historical overview of natural language processing (NLP) development, tracing its evolution from rule-based systems to contemporary deep learning approaches. The second module focuses primarily on advanced technical concepts related to text generation models, emphasizing the architectures and algorithms driving state-of-the-art generative AI systems. The final module addresses the responsible design, deployment, and governance of generative AI technologies. Throughout the course (in all modules), discussions are grounded in a trustworthiness framework, emphasizing robustness, transparency, and human oversight in generative AI systems.
Module 1: Introductory Module
This module provides an overview of generative AI, covering its fundamental concepts and key applications. It also includes a historical perspective on NLP and sets the foundation for understanding the development and impact of generative AI technologies.
Module 2: Advanced Concepts on Generative AI
This module delves into advanced technical topics in generative AI, with a focus on cutting-edge model architectures and large-scale computing techniques. It begins with an in-depth exploration of text generation models, including the progression towards foundation
models and the transformative impact of the Transformer architecture. The module further covers transfer learning strategies and the computational challenges associated with training and deploying large language models (LLMs).
Additionally, it introduces image generation techniques, highlighting models such as Stable Diffusion and Vision-Language Large Models (VLLMs), as well as the emerging area of multimodal generative AI.
Module 3: Responsible Generative AI
This module focuses on the ethical, societal, and governance aspects of generative AI. It covers foundational principles and best practices for building trustworthy generative AI systems, emphasizing transparency, fairness, robustness, human oversight as well as sustainability.
Students will engage in case study seminars applying these concepts in real-world contexts, including the development of a trustworthy generative AI chatbot for public administration. An interactive seminar will explore red teaming techniques in LLMs, highlighting strategies to identify and mitigate risks and vulnerabilities in generative AI systems.
Obiettivi formativi
The rapid development and deployment of generative AI models and applications has the potential to revolutionise various domains which brings about the urgency to use these models in a responsible manner.
Generative AI refers to creating new content in different modalities of digital text, images, audio, code, and other artifacts based on already existing content. Text generator models such as GPT-4, and its chat version,
ChatGPT as well as text-to-image models such as DALL-E 3 and Stable Diffusion are popular generative AI models. Although these models have significant implications for a wide spectrum of domains, there are several ethical
and social considerations associated with generative AI models and applications. These concerns include the existence of bias, lack of interpretability, privacy, fake and misleading content such as hallucinations, and lack of accountability.
Thus, it is very crucial to discuss these risks with their corresponding potential safeguards (if any) in addition to the technical details of these powerful models.
The course is organized as follows in three modules:
- i) an introductory one presenting main motivations as well as main concepts in NLP and further generative AI;
- ii) an advanced module providing technical details about the building blocks of foundation models with the main focus on the text generation models which also includes seminars;
- iii) an advanced module providing the main risks of generative AI with the corresponding technical considerations and challenges to establish responsible generative AI models in practice which includes seminars and a hands-on session as well.
Each module rigorously examines core principles of trustworthiness emphasizing human agency and oversight, technical robustness and safety, transparency, and the advancement of societal and environmental well-being in accordance with sustainability frameworks.
Target Audience: This course is addressed to the students of PhD in AI (uniPI) and Computational methods (SNS). The students of other PhD programs in SNS are welcome to enroll in the course if they are interested.
Riferimenti bibliografici
- Vaswani, Ashish, et al. "Attention is all you need." Advances in neural information processing systems 30 (2017).
- Paaß, G., & Giesselbach, S. (2023). Foundation models for natural language processing: Pre-trained language models integrating media (p. 436). Springer Nature.
- Cho, Kyunghyun, et al. "Learning phrase representations using RNN encoder-decoder for statistical machine translation." arXiv preprint arXiv:1406.1078 (2014).
- Sutskever, Ilya, Oriol Vinyals, and Quoc V. Le. "Sequence to Sequence Learning with Neural Networks."
- Bahdanau, Dzmitry, Kyunghyun Cho, and Yoshua Bengio. "Neural machine translation by jointly learning to align and translate." arXiv preprint arXiv:1409.0473 (2014).
- Luong, Minh-Thang, Hieu Pham, and Christopher D. Manning. "Effective approaches to attention-based neural machine translation." arXiv preprint arXiv:1508.04025 (2015).
- Peters, Matthew E., et al. "Dissecting contextual word embeddings: Architecture and representation." arXiv preprint arXiv:1808.08949 (2018).
- Mikolov, Tomas, et al. "Efficient estimation of word representations in vector space." arXiv preprint arXiv:1301.3781 (2013).
- Devlin, Jacob, et al. "Bert: Pre-training of deep bidirectional transformers for language understanding." Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). 2019.
- Radford, Alec, et al. "Improving language understanding by generative pre-training." (2018).
- Brown, Tom, et al. "Language models are few-shot learners." Advances in neural information processing systems 33 (2020): 1877-1901.
- Jiang, Albert Q., et al. "Mixtral of experts." arXiv preprint arXiv:2401.04088 (2024).
- Schick, Timo, and Hinrich Schütze. "Exploiting cloze questions for few shot text classification and natural language inference." arXiv preprint arXiv:2001.07676 (2020).
- Scao, Teven Le, and Alexander M. Rush. "How many data points is a prompt worth?." arXiv preprint arXiv:2103.08493 (2021).
- Dong, Qingxiu, et al. "A survey on in-context learning." arXiv preprint arXiv:2301.00234 (2022).
- Liu, Jiachang, et al. "What Makes Good In-Context Examples for GPT-$3 $?." arXiv preprint arXiv:2101.06804 (2021).
- Wei, Jason, et al. "Chain-of-thought prompting elicits reasoning in large language models." Advances in neural information processing systems 35 (2022): 24824-24837.
- Wang, Xuezhi, et al. "Self-consistency improves chain of thought reasoning in language models." arXiv preprint arXiv:2203.11171 (2022).
- Yao, Shunyu, et al. "Tree of thoughts: Deliberate problem solving with large language models." Advances in neural information processing systems 36 (2023): 11809-11822.
- Kojima, Takeshi, et al. "Large language models are zero-shot reasoners." Advances in neural information processing systems 35 (2022): 22199-22213.
- Ouyang, Long, et al. "Training language models to follow instructions with human feedback." Advances in neural information processing systems 35 (2022): 27730-27744.
- Gao, Yunfan, et al. "Retrieval-augmented generation for large language models: A survey." arXiv preprint arXiv:2312.10997 2 (2023): 1.
- 21. Sparck Jones, Karen. "A statistical interpretation of term specificity and its application in retrieval." Journal of documentation 28.1 (1972): 11-21.
- Robertson, Stephen E. "Overview of the okapi projects." Journal of documentation 53.1 (1997): 3-7.