Google DeepMind's new paradigm for distributed AI training

Hello, fellow human! Sophia here with some virtual talks we have lined up.

In this Newsletter:

Virtual Talk: Google DeepMind presents a new way to solve distributed training problems.

Video Lecture: Can AI solve Sophie's Choice? A video lecture on AI alignment with pluralistic human values.

A New Way to Solve Distributed Training Problems by Google DeepMind

Circle June 13th on your calendar to join us for a virtual talk with Arthur Douillard, a Senior Research Scientist at Google DeepMind. Arthur will introduce Distributed Paths Composition (DiPaCo) – a modular ML paradigm. The high-level idea is to distribute computation by path, which works better since paths are small relative to the entire model and require only a handful of tightly connected devices to train or evaluate.

Learn details of the talk and register here.

Can AI solve Sophie's Choice?

We are exploring the connection between AI and human values. Taylor Sorensen from the University of Washington recently shared his team's insightful research on aligning AI with diverse human values. He discusses the complexities of integrating AI with ethical dilemmas like the trolley problem and cases like Sophie's Choice.

Taylor introduced the ValuePrism dataset, a rich collection of over 200,000 values from 30,000 human-written situations. This dataset prompted GPT-4 to generate values, rights, and duties, which were then evaluated by humans to ensure quality and completeness.

The research team conducted a study with over 600 people from diverse backgrounds to ensure representation. Participants were asked if they agreed with the suggested values, rights, and duties and if their perspective was missing. While most participants agreed with the provided values, there were disagreements, especially on political questions like wealth redistribution. Liberals tended to agree more with certain statements than conservatives, indicating a cultural bias in the responses.

Using this feedback, the researchers developed the Kaleido System, a framework to align AI with pluralistic human values. This system aims to create customizable AI systems for users from different backgrounds, enabling AI to represent human diversity better.

In complex ethical dilemmas like the trolley problem, there is no single "right answer." The response of AI depends on its training and the perspectives it represents. Aligning human values is crucial for training AI systems accordingly.

Approaches to Align AI with Ethical Dilemmas:

• Overton Pluralism: Different schools of thought provide different answers to the ethical situation, leading to a broad spectrum of reasonable responses.

• Steerable Pluralism: The model can be customized according to requirements and guided to a certain value.

• Distributional Pluralism: An AI system should randomly sample an answer that a human might give, proportional to the population that would provide that particular answer.

It's an ongoing process to align AI with human values. There's no definitive solution, but rather an opportunity to collaborate on defining ethical behaviors for AI systems in various situations.

Catch the full lecture on the BuzzRobot YouTube channel and join the conversation.