MAIDAP

Offering new graduates the opportunity to collaborate with product groups across Microsoft by providing AI as a service in solving some of our most exciting and challenging AI problems.

Join us on the leading edge of AI

To foster Microsoft’s leadership in the field of AI, we created this ground-breaking program to develop the next generation of AI leaders. The Microsoft AI Development Acceleration Program (MAIDAP), located at the Microsoft New England Research and Development Center (NERD), is an early-in-career program for recent BS/MS/PhD graduates to gain exposure to a diverse set of AI opportunities at Microsoft. Every summer, a new cohort begins a two-year rotation program.

Candidate Qualifications

Graduates with a PhD/MS/BS degree in Electrical Engineering, Computer Science, Data Science, Statistics or other relevant fields, and any of the following experience:

  • Deep learning frameworks
  • C/C++/Java/C# and 1+ scripting languages
  • Program management and customer design
  • Designing and developing high-scale distributed systems

We are currently hiring for the following positions:

FAQs

Who is eligible to enter the program?    

The MAIDAP Program is targeted at graduating students from BS, MS, and PhD programs. Students who graduated more than a year ago are outside the target candidate pool.

When does the program start and what is its duration?

The FY21 cohort started August 10th, 2020 and the program lasts for two years. A new cohort starts every summer and the program lasts for two years.

I have a visa-related need to start early. Is that feasible?

We will process such requests on a case-by-case basis.

Where is the program located? 

Our entire team sits together at our Cambridge, MA location. We overlook the Charles River with a wonderful view of the Boston skyline.

Will the cohort members be placed in the sponsoring teams, including teams in Redmond, during the program? 

No, the cohort will be based at NERD for the duration of the program and will work with the sponsors in a client-customer (mostly remote) relationship. The team members will visit remote customers at least once per project.

What happens at the end of the program?  

The cohort members are incorporated into one of the sponsoring teams. This is accomplished via a mutual matching of interests and needs.

What happens if a participant really likes a sponsor’s team and wants to stay and focus on the current product instead of continuing with the rotational program? 

We are open to that possibility and would consider this on a case-by-case basis. However, we expect all participants to at least complete 2 project rotations.

Can someone also rotate amongst roles (e.g., PM and DS) during the program? 

We may consider this on a case-by-case basis. However, we expect most of such interests to be covered by the vast latitude in the roles that allows one to learn, explore, and develop new skills.

Is the focus on delivering quality AI solutions to be pushed to production or on educating the participants and giving them the opportunity to explore AI applications?  

While our preference is for projects that ship to external customers, we also consider projects that extend the state-of-art in AI in ways that are likely to impact our customers in the future. The learning happens as a byproduct of work. We interact with diverse data and tool stacks on a variety of product applications.

What is the ratio between the projects related to internal tools and the ones that are for Microsoft's clients? 

Most of our projects are for Microsoft’s customers. We project the ratio next year to be 20/25:80/75

Will mentoring be available?  

Each participant will be assigned a mentor, who is a Microsoft employee of the same discipline.

Can I attend AI conferences?    

Yes, Microsoft has a strong commitment to continued learning.

Will I still be able to write papers? 

Yes, we encourage a strong presence in the technical community with publishing and presenting our work, provided the project’s contents are not proprietary.

Projects

Here are some of the public-facing MAIDAP projects our cohort has been working on.

MFA Project. GAN 

Microsoft, MIT, and the Metropolitan Museum of Art (the MET) collaborated on a project to bring the MET’s recently open sourced art collection to a wider audience by developing interactive engagements using AI methods. This collaboration is best described in The MET’s articleThe MET’s blog, and Microsoft’s blog.

The MAIDAP team saw a fantastic opportunity to leverage our information extraction pipeline, improve it by experiencing it as a user, and prove its efficiency with an extremely rich dataset. In the course of a hackathon, we created and trained a generative adversarial network capable of transitioning smoothly between visual states closely resembling artwork of the MET.

NLP Repo

MAIDAP, along with the AI CAT and AzureML teams, has recently launched an open-source GitHub repository on Natural Language Processing (NLP). This repository received over 1,300 stars and was forked over 100 times in its first two weeks of availability. The included tutorials and examples ensure that our customers can use state-of-the-art algorithms for common NLP scenarios and easily build AI solutions on Azure. The repository contains a set of Python notebooks grouped by scenario, including Sentence Similarity, Question Answering, Text Classification, and Entailment. These scenarios use open-source libraries such as BERT and AllenNLP to teach common NLP concepts and best practices for implementing NLP solutions on the Azure ML infrastructure.

XLNet Starter Kit

NLP is one of the main domains in Machine Learning and with its ever-changing landscape it is often difficult to keep up with the state-of-the-art models. We created an easy to use template for one such model, XLNet. XLNet is a new unsupervised language representation learning method based on a novel generalized permutation language modeling objective. Additionally, XLNet employs Transformer-XL as the backbone model, exhibiting excellent performance for language tasks involving long context. We created an easy to use Jupyter notebook that anyone can use to perform text classification using XLNet. The project was added to NLP Best Practices repository and the boilerplate notebook now lives here.

Bike Parking Predictions

Our project is a website that predicts the number of bikes in the parking garage at NERD at any time. We experimented with multiple methods for measuring the number of bikes, including a camera running a bike detection model. Eventually, we decided to use RFID data from the building to count the number of entrances and exits from the garage. In addition to RFID data, we used weather data, including precipitation, humidity, temperature, and other factors to train different models. Our best model is a Lasso Regression model that achieves an R squared of 88%.

Project Buzz Aldrin: AI as your partner in gaming

The goal of the project was to build a cooperative AI which assists people with disabilities to play XBOX games thereby enhancing the gaming experience of the consumer. Consequently, the human and AI work together to win the game using the best possible strategy. The team used Reinforcement Learning (RL) to develop six possible Human-AI collaborative environments. A user study was conducted to determine which of the six methods is the best approach for human-AI collaboration. This project strongly ties into the accessibility element of the Microsoft culture.

BERT Notebook

Our project goal was to accelerate the ability of data scientists without a strong natural language processing (NLP) background to learn NLP best practices. Microsoft’s new open source NLP repo contains examples for building state-of-the-art NLP systems, many of which were built by MAIDAP. Using these notebooks as a starting point, we investigated whether we could effectively train BERT (Bidirectional Encoder Representation from Transformers) for sentence similarity.

DeepMurals

DeepMurals is inspired by a popular technique called NeuralStyle Transfer first introduced in a paper by Gayts and his team. This project aims to recreate an image in the style of another image of one’s choosing. Developed using an enhanced version of the original implementation, the project extended the scope from style transferring images to style transferring videos in near real-time. By pretraining Style Transfer algorithms on GPUs, Neural Style Transfer becomes accessible to everyday scenarios at the click of a button. Accessible Style Transfer can lead to innovations in interactive artistic experiences, with the potential to be integrated into museums and exhibits.

From our cohort

See what’s been happening through these stories and updates.

Microsoft AI Development Acceleration Program: Transforming Microsoft teams

From Cortana to CRISPR to XiaoIce, Microsoft’s commitment to artificial intelligence is clear. As part of our dedication to providing the best-in-class technology to our teams as well as our consumers, we’re constantly looking for ways to use AI to improve and empower lives.

READ MORE

MAIDAP Cohort Reflects on One Year of AI Excellence

This summer, we were excited to welcome our second cohort of the MAIDAP program, as our first cohort kicked off their second year with us. We’re reflecting on our first year with four members of our first cohort.

READ MORE