【Introduction】

We are developing a suite of smaller language models (SLMs) that are similar to LLMs but use less computing power. This project focuses on studying advanced training techniques that can better align the capabilities of SLMs with various aspects of different product scenarios, including but not limited to instruction following and task planning.

【Research Areas】

Language Model, Machine Learning

【Qualifications】

• Passionate about research in language models
• Solid coding and communication skills
• Prior experience in language model and/or machine learning research is preferred

【Introduction】

Large language models (LLMs) have revolutionized various fields with technologies like Retrieval Augmented Generation (RAG), In-Context Learning (ICL), Chain of Thought (CoT), and Agent-based models. These advancements, while groundbreaking, often result in lengthy prompts that lead to increased computational and financial costs, higher latencies, and added redundancy. Moreover, the intrinsic position bias of LLMs and redundancy within prompt will impact their performance, leading to the "lost in the middle" issue.

Previous studies have introduced prompt compression methods such as LLMLingua and LongLLMLingua, which address these issues and show promising results in generic scenarios. This project aims to explore research questions around complex scenarios, such as agent-related prompts and the compression of LLM responses. Furthermore, it seeks to investigate the effects of such compression techniques on adversarial attacks, security, and other critical aspects.

【Research Areas】

Large Language Models, Agent-based, Efficient Method

https://www.microsoft.com/en-us/research/project/llmlingua/overview/

【Qualifications】

• Passionate about research in LLMs and AGI
• Solid coding and communication skills
• Prior experience in LLM and/or Machine Learning research is preferred

【Introduction】

While there have been significant efforts on leveraging LLMs as an evaluator, it is not quite there yet. It is only useful in English and in certain tasks, which severely limits its useability and trustworthiness across language. Join a groundbreaking project at Microsoft Research Asia, focusing on answering fundamental questions around LLM-based evaluation, but with direct production impact. This project aims to surpass the current capabilities of LLMs in certain tasks, emphasizing accuracy, reliability, robustness, and generalizability. The intern will be instrumental in creating a production-deployed system that adapts to needs serving hundreds of millions of users; and answer fundamental questions around the capabilities, limitations, and usages of LLMs and beyond.

【Research Areas】

Large Language Models, LLM-based Evaluation, LLM for low-resource language

https://www.microsoft.com/en-us/research/group/natural-language-computing/

【Qualifications】

• Passionate about research in LLMs and AGI
• Solid coding and communication skills
• Graduate student working towards a Ph.D
• Prior experience in LLM and/or Machine Learning research is preferred

【Introduction】

Retrieval-augmented generation (RAG) is a technique for enhancing the quality of responses generated by large language models (LLMs) by using external sources of knowledge to supplement the LLM's internal representation of information. RAG allows LLMs to access the most up-to-date and reliable facts from a knowledge base or internal information storage. It can be used for various natural language generation tasks, such as question answering, summarization, and chat. However, the documents retrieved might be redundant and noisy. This project aims to develop efficient and robust RAG methods, which leverage shorter context length by removing contradiction and redundancy, reduce hallucination and are robust in different domains.

【Research Areas】

Large Language Models, Retrieval-Augmented Generation

https://www.microsoft.com/en-us/research/group/natural-language-computing/

【Qualifications】

• Passionate about research in LLMs and AGI
• Solid coding and communication skills
• Prior experience in LLM and/or Machine Learning research is preferred

【Introduction】

For the new Designer app and Designer in Edge, we need to resize templates to different sizes, since different social media platforms require different target dimensions of the media, e.g., Facebook Timeline Post for personal accounts and business pages (1200 x 628), LinkedIn timeline post (1200 x 1200), Twitter timeline post (1600 x 900), etc. Image is the center of a template design. We need an ML-powered technique to automatically resize (including aspect ratio change, crop, zoom in/out) an image and put it into a resized template (more specifically speaking, resized image placeholder) for the target platform, so that the image placement looks good (i.e., maintaining the aesthetic values).

【Research Areas】

Computer Vision and Machine Learning

【Qualifications】

• Ph.D. students majoring in computer science, applied mathematics, electrical engineering or related technical discipline
• Relevant experience in the development and application of computer vision and/or machine learning algorithms to solve challenging image understanding problems
• Strong scientific programming skills, including C/C++, MATLAB, Python
• Independent analytical problem-solving skills
• Experience collaborating within research teams to develop advanced research concepts, prototypes, and systems
• Strong communication skills

【Introduction】

Pretrained language models such as BERT and UniLM have achieved huge success in many natural language processing scenarios. In many recommendation scenarios such as news recommendation, video recommendation, and ads CTR/CVR prediction, user models are very important to infer user interest and intent from user behaviors. Previously, user models are trained in a supervised task-specific way, which cannot achieve a global and universal understanding of users and may limit they capacities in serving personalized applications.

In this project, inspired by the success of pretrained language models, we plan to pretrain universal user models from large-scale unlabeled user behaviors using self-supervision tasks. The pretrained user models aim to better understand the characteristics, interest and intent of users, and can empower different downstream recommendation tasks by finetuning on their labeled data. Our recent work can be found at

https://scholar.google.co.jp/citations?hl=zh-CN&user=0SZVO0sAAAAJ&view_op=list_works&sortby=pubdate

【Research Areas】

Recommender Systems and Natural Language Processing

【Qualifications】

• Ph.D. students majoring in computer science, electronic engineering, or related areas
• Self-motivated and passionate in research
• Solid coding skills
• Experienced in Recommender Systems and Natural Language Processing

【Introduction】

Learning visual representation by vision-language pair data has shown highly competitive compared to previous supervised and self-supervised approaches, pioneered by CLIP and DALL-E. Such vision-language learning approaches have also demonstrated strong performance on some pure vision and vision-language applications. The aim of this project is to continually push forward the boundary of this research direction.

【Research Areas 】

Computer vision

https://www.microsoft.com/en-us/research/group/visual-computing/

https://www.microsoft.com/en-us/research/people/hanhu/

【Qualifications】

• Currently enrolled oversea Ph. D. students with promised or deferred offer, and is now staying in China
• Major in computer vision, natural language processing, or machine learning

【Introduction】

Are you excited to apply deep neural networks to solve practical problems? Would you like to help secure enterprise computer systems and users across the globe? Cyber-attacks on enterprises are proliferating and oftentimes causing damage to essential business operations. Adversaries may steal credentials of valid users and use their accounts to conduct malicious activities, which abruptly deviate from valid user behavior. We aim to prevent such attacks by detecting abrupt user behavior changes.

In this project, you will leverage deep neural networks to model behaviors of a large number of users, detect abrupt behavior changes of individual users, and determine if changed behaviors are malicious or not. You will be part of a joint initiative between Microsoft Research and the Microsoft Defender for Endpoint (MDE). During your internship, you will get to collaborate with some of the world’s best researchers in security and machine learning.

You would be expected to:

• Closely work with researchers in China and Israel towards the research goals of the project.
• Develop and implement research ideas and conduct experiments to validate them.
• Report and present findings.

Microsoft is an equal opportunity employer.

【Research Areas】

Software Analytics, MSR Asia

https://www.microsoft.com/en-us/research/group/software-analytics/

Microsoft Defender for Endpoint (MDE)

This is a Microsoft engineering and research group that develops the Microsoft Defender for Endpoint, an enterprise endpoint security platform designed to help enterprise networks prevent, detect, investigate, and respond to advanced threats

https://www.microsoft.com/en-us/security/business/threat-protection/endpoint-defender

【Qualifications】

• Must have at least 1 year of experience applying machine learning/deep learning to real world/ research problems
• Demonstrated hands on the experience with Python through previous projects
• Familiarity with Deep Learning frameworks like PyTorch, Tensorflow, etc
• Keen ability for attention to detail and a strong analytical mindset
• Excellent in English reading and reasonably good in English communications
• Advisor’s permission

Those with the following conditions are preferred:

• Prior experience in behavior modeling
• Prior experience in anomaly detection
• Security knowledge a plus

【Introduction】

While PLMs have been widely used to generate high-quality texts in a supervised manner (by imitating texts written by humans), they lack a mechanism for generating texts that directly optimize a given reward, e.g., given user feedback like user clicks or a criterion that cannot be directly optimized by using gradient descent. In real-world applications, we usually wish to achieve more than just imitating existing texts. For example, we may wish to generate more attractive texts that lead to increased user clicks, more diversified texts to improve user experience, and more personalized texts that are better tailored to user tastes. Combing RL with PLMs provides a unified solution for all these scenarios, and is the core for machines to achieve human parity in text generation. Such a method has the potential to be applied in a wide range of products, e.g., Microsoft Advertising (text ad generation), Microsoft News (news headline generation), and Microsoft Stores and Xbox (optimizing the description for recommended items).

In this project, we aim to study how pretrained language models (PLMs) can be enhanced by using deep reinforcement learning (RL) to generate attractive and high-quality text ads. While finetuning PLMs have been shown to be able to generate high-quality texts, RL additionally provides a principled way to directly optimize user feedback (e.g., user clicks) for improving attractiveness. Our initial RL method UMPG is deployed in Dynamic Search Ads and published in KDD 2021. We wish to extend the method so that it can work for all pretrained language models (in addition to UNILM) and study how the technique can benefit other important Microsoft Advertising products and international markets.

【Research Areas】

Social Computing (SC), MSR Asia

https://www.microsoft.com/en-us/research/group/social-computing-beijing/

Microsoft Advertising, Microsoft Redmond

【Qualifications】

• Ph.D. students majoring in computer science, electrical engineering, or equivalent areas
• Experience with deep NLP and Transformers a strong plus
• Background knowledge of language model pre-training and/or reinforcement learning
• Capable of system implementing based on academic papers in English

Those with the following conditions are preferred:

• Good English reading and writing ability and communication skills, capable of writing English papers and documents
• Active on GitHub, used or participated in well-known open source projects

【Introduction】

The parallel and distributed systems are the solution to address the ever-increasing complexity problem of deep learning trainings. However, existing solutions still leave efficiency and scalability on the table by missing optimization opportunities on various environments at industrial scale.

In this project, we’ll work with scientists who are at the forefront of system and network research, leveraging the world-leading platforms to solve system and networking problems in parallel and distributed deep learning area. The current project team members, from both MSR Asia and MSR Redmond labs, have rich experience contributing to both industry and academic community through transferring innovations that support production systems and publications at top conferences.

【Research Areas】

System and Networking, MSR Asia

https://www.microsoft.com/en-us/research/group/systems-and-networking-research-group-asia/

Research in Software Engineering, MSR Redmond

https://www.microsoft.com/en-us/research/group/research-software-engineering-rise/

【Qualifications】

• Major in computer science, electrical engineering, or equivalent field
• Solid knowledge of data structure/algorithm
• Familiarity with Python, C/C++ and other programming languages, familiar with Linux and development on Linux platform
• Good communication and presentation skills
• Good English reading and writing ability, capable of system implementing based on academic papers in English, capable of writing English documents

Those with the following conditions are preferred:

• Familiarity with deep learning systems, e.g., PyTorch TensorFlow, GPU programming and networking
• Familiarity with NCCL, MPI communication protocols such as OpenMPI and MVAPICH
• Rich knowledge of machine learning and machine learning models
• Familiarity with engineering process as a strong plus
• Active on GitHub, used or participated in well-known open-source projects

【Introduction】

Tabular data such as Excel spreadsheets and databases are one of the most important assets in large enterprises today, which however are often plagued with data quality issues. Intelligent data cleansing focuses on novel ways to detect and fix data quality issues in tabular data, which can assist the large class of less-technical and non-technical users in enterprises.

We are interested in a variety of topics in this area, including data-driven and intelligent techniques to detect data quality issues and suggest possible fixes, leveraging inferred constraints and statistical properties based on existing data assets and software artifacts.

【Research Areas】

Data, Knowledge, and Intelligence (DKI), MSR Asia

https://www.microsoft.com/en-us/research/group/data-knowledge-intelligence/

Exploration and Mining (DMX), MSR Redmond

https://www.microsoft.com/en-us/research/group/data-management-exploration-and-mining-dmx

【Qualifications】

• Graduate-level students in Computer Science or related STEM fields. PhD students are preferred
• Students with research background in database, data mining, statistics, software engineering, and visualization are preferred

【Introduction】

As one of the world-leading cloud service providers, Microsoft Azure manages tens of millions of virtual machines every day. Within such a large-scale cloud system, how to efficiently allocate virtual machines on servers is critical and has been a hot research topic for years. Previously, teams from MSR-Asia and MSR-Redmond have made significant contributions in this area that resulted in production impact and publication of academic papers at top-tier conferences (e.g., IJCAI, AAAI, OSDI, NSDI). In this project we intend to unify the strength of MSR-Asia and MSR-Redmond for performing forward-looking and collaborative research on power management in datacenters, including power-aware virtual machine allocation. The project involves developing power prediction models by leveraging the start-of-the-art machine learning methods, as well as building efficient and reliable allocation systems in large-scale distributed environments.

【Research Areas】

Data, Knowledge, and Intelligence (DKI), MSR Asia

https://www.microsoft.com/en-us/research/group/data-knowledge-intelligence/

System, MSR Redmond

https://www.microsoft.com/en-us/research/group/systems-research-group-redmond/

【Qualifications】

• Currently enrolled in a graduate program in computer science or equivalent field
• Good research track record in related areas
• Able to carry out research tasks with high quality
• Good communication and presentation skills in written and oral English
• Knowledge and experience in machine learning, data mining and data analytics are preferred
• Familiarity with AIOps or AI for systems is a strong plus

【Introduction】

In today’s real-time video applications, a key component for optimizing the user’s quality of experience is bandwidth estimation and rate control. It estimates the network capacity based on congestion signals observed on the path and adapts the video bitrate accordingly through the codec. However, existing handcrafted bandwidth estimators have failed to accommodate a wide range of complex network conditions, calling for a data-driven approach.

Motivated by the recent success in applying reinforcement learning (RL) to video streaming and congestion control, we have made an initial attempt at designing an RL-based bandwidth estimator for one-on-one video calls. Going forward, we are working to optimize the performance of our current neural network model, as well as extending the research study of bandwidth estimation and rate control to multiparty videoconferencing.

【Research Areas】

System and Networking, MSR Asia

https://www.microsoft.com/en-us/research/group/systems-and-networking-research-group-asia/

Mobility and Networking, MSR Redmond

https://www.microsoft.com/en-us/research/group/mobility-and-networking-research/

【Qualifications】

• Major in computer science or a related field
• Strong programming skills in Python or C++
• Excellent English communication skills
• Experience with deep reinforcement learning or related areas is preferred
• Knowledge of computer networks is preferred
• Background in AI for systems and networking is a strong plus
• Track record of publications in top systems, networking, or AI conferences is strongly preferred

【Introduction】

Our cross-lab, inter-disciplinary research team develops AI technology for interactive coding assistance for data science, data analytics, and business process automation. It allows the user to specify their data processing intent in the middle of their workflow using a combination of natural language, input-output examples, and multi-modal UX – and translates that intent into the desired source code. The underlying AI technology integrates our state-of-the-art research in program synthesis, semantic parsing, and structure-grounded natural language understanding. It has the potential to improve productivity of millions of data scientists and software developers, as well as establish new scientific milestones for deep learning over structured data, grounded language understanding, and neuro-symbolic AI.

The research project involves collecting and establishing a novel benchmark dataset for data science program generation, developing novel neuro-symbolic semantic parsing models to tackle this challenge, adapting large-scale pretrained language models to new domains and knowledge bases, as well as publishing in top-tier AI/NLP conferences. We expect the benchmark dataset and the new models to be used in academia as well as in Microsoft products.

【Research Areas】

Natural Language Computing, MSR Asia

https://www.microsoft.com/en-us/research/group/natural-language-computing

Neuro-Symbolic Learning, MSR Redmond

【Qualifications】

• Masters or Ph.D. students, majoring in computer science or equivalent areas
• Background in deep NLP, semantic parsing, sequence-to-sequence learning, Transformers required
• Experience with PyTorch and HuggingFace Transformers
• Fluent English speaking, listening, and writing skills
• Background in deep learning over structured data (graphs/trees/programs) and program synthesis preferred
• Students with papers published at top-tier AI/NLP conferences are preferred

【Introduction】

The goal of this project is to develop game-changing techniques for next-gen large pre-trained language models, including：

(1) Beyond UniLM/InfoXLM: novel pre-training frameworks and self-supervised tasks for monolingual and multilingual pre-training to support language understanding, generation and translation tasks;

(2) Beyond Transformers: new model architectures and optimization algorithms for improving training effectiveness and efficiency of extremely large language models;

(3) Knowledge Fusion: new modeling frameworks to fuse massive pre-compiled knowledge into pre-trained models;

(4) Lifelong Self-supervised Learning: mechanisms and algorithms for lifelong (incremental) pre-training. This project extends our existing research and aims to advance SOTA on NLP and AI in general.

【Research Areas】

Natural Language Computing, MSR Asia

https://www.microsoft.com/en-us/research/group/natural-language-computing

Deep Learning, MSR Redmond

https://www.microsoft.com/en-us/research/group/deep-learning-group

【Qualifications】

• Major in computer science or equivalent areas
• One+ year research experience in deep learning for NLP, CV or related areas
• Experience with open-source tools such as PyTorch, Tensorflow, etc.
• Background knowledge of language model pre-training is preferred
• Track record of publications in related top conferences (e.g., ACL, EMNLP, NAACL, ICML, NeurIPS, ICLR) is preferred
• Excellent communication and writing skills