Halley Young

Maintaining Diversity in the Age of AI

Postdoctoral Researcher at Microsoft Research NYC focused on diversifying AI outputs and preserving user autonomy in both creative and technical domains.

About Me

Combining technical expertise with a focus on AI diversity and human agency

I'm a postdoctoral researcher at Microsoft Research in NYC, with a focus on diversifying the output of AI in both creative and technical domains. With a background in computer science and expertise in deep learning, generative models, and music AI, I approach artificial intelligence from a unique perspective: how can we ensure AI systems enhance human creativity and autonomy rather than diminish them?

My research addresses two related but distinct dangers in current AI development:

  1. Output homogenization: The tendency of AI systems to converge toward a narrow range of outputs that represent a statistical average rather than the full spectrum of human expression
  2. Deprivation of user autonomy: The risk that AI tools replace human decision-making rather than augmenting it
Diversity Preservation

Developing techniques to maintain diversity in AI outputs across domains

User Agency

Creating systems that enhance rather than replace human decision-making

Technical Expertise

Deep knowledge of generative models, formal methods, and creative AI

Research Focus

Exploring the technical and social aspects of AI diversity

Diversity as a Technical and Social Imperative

For me, diversity is both a technical term (maximizing expected pairwise distance between two answers conditioned on them both being "good") and a concept with profound creative, economic, and social implications.

In a recent study (paper in submission), I found that 249 out of 250 poems generated by ChatGPT without special prompting were all identified by readers as being derived from the same poet (John Keats). This striking homogeneity illustrates how large language models tend to converge toward a "safe" mean statisticala problem that extends beyond poetry to code, design, business ideas, and other domains where innovation requires deviation from the norm.

My research develops methods to maintain diversity in AI outputs in cases where individuals, companies, or society would benefit from exploring a range of good but different possibilities rather than a single "optimal" solution.

Counter-Egos, Not Replacements

Collaborating with professional poets

Working with professional poets to develop methods for creating "Counter-Egos" rather than AI replacements. This approach focuses on reinterpreting poetic styles through the lens of Webpage Poetry, allowing for new forms of expression.

Identity Matters

IUI (HAI-GEN) 2024

Enabling people to explore their identity using LLMs with sophisticated algorithms that lead to real-world usable outputs, providing tools for personal exploration and discovery guided by user agency.

Controllable Music Generation

70,000+ listens on SoundCloud

Developing music generation models that enhance musician creativity while preserving artistic control and avoiding output homogenization.

Projects

Research initiatives focused on diversity in AI

AI for Creativity

aiforcreativity.net

A hub for research, tools, and essays on maintaining creative diversity in the age of AI. Provides resources for artists, developers, and researchers interested in building AI systems that enhance rather than constrain human creativity.

Music Generation Research

SoundCloud (70,000+ listens)

Showcasing research on autonomy in music generation with practical demonstrations that have reached a substantial audience.

Diversifying LLM Outputs

Current Research

Developing techniques to ensure large language models can produce diverse, creative outputs rather than converging on a statistical mean.

Selected Publications

Research contributions to AI diversity and controllability

Neurosymbolic Deep Generative Models for Sequence Data with Relational Constraints (2022)
Joint work with Osbert Bastani and Maxwell Du
Published in NEURIPS 2022

Developed a method for measuring and producing meaningful global coherence in domains such as music and poetry, enabling generated outputs that maintain both local quality and structural integrity.

Compositional Steering of Music Transformers (2021)
Joint work with Jesse Engel, Vincent Dumoulin, Pablo Castro, and Anna Huang
Published in HAI-GEN 2022

Developed a novel approach to steer music transformer models without manipulating their parameters, giving users greater control over generated music while preserving model capabilities.

Learning Neurosymbolic Generative Models via Program Synthesis (2019)
Joint work with Osbert Bastani and Mayur Naik
Published in ICML 2019 proceedings

Developed generative models based on latent structures detectable via program synthesis, laying groundwork for more interpretable and controllable generative AI.

Experience

Academic and industry roles focused on AI diversity

2023-Present

Microsoft Research, NYC

Postdoctoral Researcher

Leading research on diversifying AI outputs in creative and technical domains, developing new algorithms and evaluation metrics for measuring and promoting diversity in generative models.

2021-2023

Google Brain/Magenta

Research Intern

Developed controllable generative music models to avoid both user de-agentification and output homogeny. Created tools that allow musicians to guide AI systems rather than be replaced by them.

2019-2020

Riffit

Software Engineer/Researcher

Performed research in automated music production for therapeutic purposes leading to a patent.

2020-2021

University of Pennsylvania

Teaching Assistant

TA in Graduate Artificial Intelligence, Head TA in Graduate Applied Machine Learning. Designed curriculum that emphasized ethical considerations alongside technical skills.

Education

2017-2023

PhD in Computer Science

University of Pennsylvania

Awarded the Chateaubriand Fellowship for Humanities and Social Sciences (declined due to health issues). Dissertation focused on maintaining controllability in generative AI systems.

2014-2017

B.A. magna cum laude, Computer Science

New York University

Awarded the Faculty Memorial Prize, the Computer Science Prize, and the Computer Science Prize for Outstanding Performance in Computer Science.

2013-2014

B.A. Candidate

University of Chicago

Awarded the Lilian Gertrude Selz Prize.

Skills

Programming Languages

Python (PyTorch, JAX)
Julia
C
Matlab
Haskell

Research Areas

Deep Learning
Generative Models
Steerable AI
Music AI
Formal Methods
Human-AI Interaction
AI Diversity
Neurosymbolic Models

Contact

Get in touch to discuss research collaboration or speaking opportunities