Research Focus: Week of May 22, 2023

Microsoft Research
Research Focus 16 | Week of May 22nd, 2023

Welcome to Research Focus, a series of blog posts that highlights notable publications, events, code/datasets, new hires and other milestones from across the research community at Microsoft.


Causal Reasoning and Large Language Models: Opening a New Frontier for Causality

Emre Kıcıman, Robert Ness, Amit Sharma, Chenhao Tan

Recent advances in scaling large language models (LLMs) have led to breakthroughs in AI capabilities, including writing code in programming languages, generating stories, poems, essays, and other texts, and strong performance in certain reasoning tasks. LLMs can even create plausible explanations for their outputs, and update their conclusions given new evidence.

At the same time, LLMs can make absurd claims and basic errors of logic, mathematics, and complex reasoning, which raises questions about their applicability in societally impactful domains such as medicine, science, law, and policy.

In a new paper: Causal Reasoning and Large Language Models: Opening a New Frontier for Causality, researchers from Microsoft examine the causal capabilities of LLMs. They find that LLMs, on average, can outperform state-of-the-art causal algorithms in graph discovery and counterfactual inference, and can systematize nebulous concepts like necessity and sufficiency of cause by operating solely on natural language input. They show that by capturing commonsense and domain knowledge about causal mechanisms, LLMs open new frontiers for advancing the research, practice, and adoption of causality. The researchers envision pairing LLMs alongside existing causal methods to reduce the required manual effort that has been a major impediment to widespread adoption of causal analysis. 

Spotlight: Microsoft Research Podcast

AI Frontiers: The Physics of AI with Sébastien Bubeck

What is intelligence? How does it emerge and how do we measure it? Ashley Llorens and machine learning theorist Sébastian Bubeck discuss accelerating progress in large-scale AI and early experiments with GPT-4.


DNA storage in thermoresponsive microcapsules for repeated random multiplexed data access

As the world generates more and more data, data storage capacity has not kept pace. Traditional long-term storage media such as hard disks or magnetic tape have limited durability and storage density. But DNA has an intrinsic capacity for information storage, durability, and high information density.

In DNA data storage, a large amount of data is stored together, and it is important to perform random access – selective retrieval of individual data files. This is achieved using polymerase chain reaction (PCR), a molecular process that can exponentially amplify a target file. However, this process can damage the data and cause errors. PCR amplification of multiple files simultaneously creates serious undesired DNA crosstalk. Currently one can only read one file at a time, but not a subset of files in a larger set.

In a recent paper: DNA storage in thermoresponsive microcapsules for repeated random multiplexed data access, researchers from Microsoft and external colleagues report on their work to develop a microcapsule-based PCR random access. By encapsulating individual files in each capsule, DNA files were physically separated, reducing undesired crosstalk. This enabled the simultaneous reading of all 25 files in the pool, without significant errors. The use of microcapsules also allowed DNA files to be recovered after random access, addressing the destructive reads problem and potentially making DNA data storage more economical.


Human-centered AI with Ben Shneiderman, Distinguished University Professor—University of Maryland Department of Computer Science

A new synthesis is emerging that integrates AI technologies with human-computer interaction (HCI) to produce human-centered AI (HCAI). Advocates of HCAI seek to amplify, augment, and enhance human abilities, so as to empower people, build their self-efficacy, support creativity, recognize responsibility, and promote social connections. Researchers, developers, business leaders, policy makers, and others are expanding the technology-centered scope of AI to include HCAI ways of thinking.

In this recent Microsoft Research Talk: Human-Centered AI: Ensuring Human Control While Increasing Automation Ben Shneiderman discusses his HCAI framework, design metaphors, and governance structures and other ideas drawn from his award-winning new book Human-Centered AI. The talk by Shneiderman, a Distinguished University Professor in the University of Maryland Department of Computer Science, is hosted by Mary Czerwinski, Partner Researcher and Research Manager with Microsoft Research.


AI and the New Future of Work – call for proposals

The Microsoft New Future of Work Initiative is now accepting proposals to fund academic projects that help maximize the impact of LLMs and related AI systems on how work gets done. This call for proposals targets work that specifically supports the use of LLMs in productivity scenarios. The program plans to distribute five $50,000 USD unrestricted awards to support creative research that redefines what work might mean in various contexts. 

For example: how can we ensure these new technologies truly accelerate productivity rather than having effects on the margins; how can LLMs achieve these gains by augmenting human labor; what is the future of a ‘document’ in a world where natural language can be so easily remixed and repurposed.  

Proposals will be accepted through June 5, 2023.

The post Research Focus: Week of May 22, 2023 appeared first on Microsoft Research.

Read More