Assistant Professor (starting in Fall 2020)
Department of Mechanical Engineering
Massachusetts Institute of Technology
Email: ''faez'' at ''mit'' dot ''edu''
Google Scholar | LinkedIn


Dissertation Research: Design for everyone, by everyone

I develop machine learning and optimization methods to study Engineering Design problems. I am interested in transforming the way we design products by democratizing design — where teams of globally distributed people work together to design physical products. To achieve this, I work on principled methods for representation, learning, and optimization of discrete problems occurring in design. This will enable us to democratize design by allowing designers to express complex ideas from anywhere in the world and firms to process them efficiently. Particularly, I have answered three questions through my research:

1. How does one reliably measure the creativity of ideas?

A wide variety of engineering applications such as designing self-driving cars, medical devices, or manufacturing equipment require searching for creative solutions among many possible options. However, giving a numerical score to measure creativity is difficult due to two main reasons: 1) creativity measurement requires human insight, which is hard to quantify; and 2) mathematical functions to compute scores cannot be applied to designs, which are not in a vector space a priori. As a result, most existing creativity measures are criticized for being subjective and not generalizable to new domains.

To overcome these problems, I proposed a new framework to study creativity using modern machine learning methods. First, I find design embedding in 2-D Euclidean space, such that the embedding captures human intuition of relationships between items. I find these relationships by asking pairwise comparisons from people. Next, I calculate an embedding using kernel learning methods and then use a family of novelty detection functions (like one-class classification) to measure creativity.

Recent Publications:

2. How does one form teams to evaluate design ideas?

A classical and long-standing problem in computer science and economics, with widespread application in health-care, education, advertising, and general resource allocation is the bipartite matching problem. In bipartite matching, items on one side of a market are matched to items on the other. For example, allocating conference papers to reviewers requires allocating multiple reviewers to each paper and each paper should be matched to three or more reviewers

When forming teams to review designs, we often allocate reviewers based on their expertise. However, having multiple reviewers who are all from the same department or are similar to each other may not be desirable. To address this, I proposed a new algorithm for diverse matching, using a quadratic programming based approach to solve a super-modular minimization problem that balances diversity and quality of the solution. I demonstrated the efficacy of our methods on three real-world datasets and showed that, in practice, encouraging diverse teams does not sacrifice performance.

I also proposed a method for the online version of this problem, where we do not know the availability of people beforehand. In such cases, we need to decide at the moment whether to assign a newly arrived person to a team or not. The online matching algorithms are relevant to crowd markets, reviewer assignment for journal papers as well as domains where discrete items arrive sequentially.

Recent Publications:
  • Ahmed, F., Dickerson, J. & Fuge, M. (International Joint Conference of Artificial Intelligence 2017).
    Diverse Weighted Bipartite b-Matching
  • Ahmadi, S., Ahmed, F., Dickerson, J., Fuge, M. & Khuller, S. (Under review).
    On Diverse Bipartite b-Matchings
  • Ahmed, F., Dickerson, J. & Fuge, M. (Under review).
    Forming Diverse Teams from Sequentially Arriving People

3. How does one select the most diverse ideas out of hundreds of submissions?

Subset selection is a common problem in many domains like network analysis, pattern classification, knowledge discovery, document summarization, and online contests. For example, in many online contests, the end goal is to fund the implementation of a small subset of novel and useful ideas. This problem can be formulated as constrained submodular optimization, which is NP-Hard. I proposed and successfully implemented an optimization method to filter out a small subset of ideas which has both high quality and coverage (ideas are different from each other).

I also proposed a diverse ranking method to rank order all ideas by maximizing quality and coverage. We showed the applicability of our algorithms to information retrieval and finding design analogies. To rank ideas, I showed how to capture the similarity between items, modeled coverage with Determinantal Point Processes and proposed an efficient combinatorial optimization method to find the solution. To test diverse ranking algorithms in the real world, I deployed it on a web-platform with the aim of searching for best ideas using crowd voting. Our results on two real-world problems (OpenIDEO and UNICEF) showed how diverse ranking leads to a large increase in filtering efficiency.

Recent Publications: