Project: Tufts TTS Network Diagnostic Mobile App

The Problem: A Tufts student is very sad about the performance of Internet in dorm room. TTS needs information to diagnose the problem: SSID, IP address, MAC address of device, RSSI, etc.

Project Sponsor: David DeChellis, Asst Director Network Services, David.DeChellis@tufts.edu

Project: PDF Accessibility

Project Description: I'm a member of an organization (AccessSIGCHI) dedicated to improving the accessibility of ACM HCI conferences (and in the long term, all ACM conferences), and one of our big problems right now is finding a way to build accessible PDFs (usable with screen readers). Despite the fact that PDF is an open file format, the only tool that can be used to make a fully accessible PDF is the Pro version of Adobe Acrobat (which is very expensive and the process is not nearly as automated as it could be). I'd like to have some students dig into the PDF format and build a tool (ideally online) for making accessible PDFs. This would also be something that could generally benefit the academic and student community, by making it easier for faculty to provide accessible PDFs to students.

Let me know if there's anyone interested in working on such a project; you can see more about me and my work at my lab website (http://aabl.cs.tufts.edu) and more about AccessSIGCHI at our (under construction) website here: https://accesssigchi.com

Project Sponsor: Elaine Schaertl Short, elaine.short@tufts.edu

Projects with STR Research (https://www.stresearch.com/)

Project Idea #1

Building algorithms to correlate documents (pdf) with source code. Notionally, build a topic model over a corpus of algorithm description documents (and other sources). Build similar type models over software (or binaries). Learn correlations to link algorithms and documentations to part of code. Could you link a description of an algorithm (like FFT) to functions (embedded in a large project) that implement an FFT.

Project Sponsor Mark Stevens, mark.r.stevens@stresearch.com

Project Idea #2

Scrape Github to learn machine translations of source code and comments. When a function is documented, learn associations of the code comments to the function signatures. Given new function signatures generate possible documentation. Other potential options include blind source separating developers based on coding style. Or learn to cluster code (like the kernel) based on developer patterns.

Project Sponsor Mark Stevens, mark.r.stevens@stresearch.com

Project Idea #3

Given a CVE description of a vulnerability and a corresponding binary released to fix that vulnerability, use differential analysis over previous versions to identify the part of the binary that was patched. Extensions could include fuzzing to exercise the vulnerability.

Project Sponsor Mark Stevens, mark.r.stevens@stresearch.com

Project Idea #4

Retroactively add security features to existing binaries. An example would be to use PIN to add a shadow stack to existing binary and check return addresses. Evaluate what could be stopped given these security bolt-ons.

Project Sponsor: Mark Stevens, mark.r.stevens@stresearch.com

Projects with Fidelity Investments (https://www.fidelity.com/)

Project Idea #1:

We believe that current employment models do not adequately factor in the potential for major disruptions to the labor force (ex. the recent COVID-19 pandemic). Therefore, in order to better understand the economic fallout and recovery from such a disruption, we would like to develop a new model factoring in more recent datapoints.

1. The main goal of the project is to develop a model (preferably in Python) that can accurately predict the civilian unemployment rate based on data that is available prior to the result, which is published monthly. Another body of work associated with building a model like this is to determine how to measure the accuracy of the model/candidate models.

2. Additionally, please provide any insight into different pockets of the workforce that may be affected differently by the current economic state and potential recovery. Examples of this include but are not limited to measuring impact across various geographies or
demographic categories.

Project Sponsor: Simon Chao, Head of Emerging Technology / Asset Management, simon.chao@fmr.com

Project Idea #2: Reinforcement Learning for ChatBot Optimization

A customer service chatbot usually has some predefined intents for specific business functions such as opening an account, which involves several rounds of interactions between the bot and the customer in order to get all the relevant information before the bot can successfully fulfill the customer’s request for action.

For example, in order for Fidelity to open an account for a potential customer, quite some information needs to be gathered such as identity, contact info, employment, risk assessment etc. This is definitely more complicated than ordering a move ticket, which Microsoft AI researchers did a study using
reinforcement learning back in 2017. The research found out that for a "goal oriented chatbot", we can leverage reinforcement learning to find out an optimized way to organize chat conversation such that there is least resistance from the customer in completing the task. Intuitively, the chatbot intent design and experience optimization process bear similarity to learning the best way to pass through various stages of online video game. Both involves the simulation of the agent and the environment as well as figuring out the reward function as feedback loop.

The project deliverables are described below:

1. A reinforcement learning model to optimize the conversation flow for Fidelity’s account opening chatbot;
2. A general applicable process for simulating the natural language conversation environment so that other "goal oriented chatbot" can apply the same approach to tune and optimize;

There are two final deliverables for the project:

1. A report that contains a summary of the research completed and the findings. The report should contain an executive summary (one page) of the overall findings, followed by the details of findings and research. The report will serve as documentation for the project if it were to be continued in the future. The audience for the executive summary is a non-technical user, the audience for the sections containing detail should be tailored for a technical/data science savvy audience.
2. A presentation of the research completed and findings.

Project Sponsor: Simon Chao, Head of Emerging Technology / Asset Management, simon.chao@fmr.com

Projects with MIT Lincoln Laboratory

Project: Explainable AI (XAI)

Explainable AI (XAI) seeks to illuminate the black box models of machine learning (ML) and provide reasoning for model decisions. As XAI is a new and emerging area of research, many different open source techniques that aim to explain ML models are being developed and released. The optimal solution for a given use case will depend on many factors, including but not limited to:

- Whether a ML model is supervised or unsupervised
- The problem domain (e.g. image recognition, cyber security)
- Type of explanations desired (explain a decision, a particular data point, or an entire dataset)
- Feature engineering/selection (determining which features contribute the most/least to a ML model decision)

We would like to sponsor a capstone project where students will focus on developing a tool/application to generate visualizations for a desired use case and domain using XAI techniques.

This work could involve:

- Selection of a domain, data type, and models (e.g., face ID, images)
- Literature review of existing XAI techniques for the data and model types
- Implementation of several techniques, with potential customization towards the domain
- Evaluation on a publicly available dataset
- Tool/application/dashboard prototyping to display multiple XAI outputs within one view
- Further steps/research questions:
  - Is there a better way to visualize the outputs of the XAI techniques in a more unified/synergistic fashion? Are there ways to synthesize information from each technique and display as one output?
  - Identify a second domain or data type and evaluate how well the tools generalize/transfer to that domain. Which parts of the pipeline need to be changed to do this? What information might be lost when applying techniques to other domains?

Project Sponsors: Kenny Alperin Kenneth.Alperin@ll.mit.edu; Pedro Torres-Carrasquillo ptorres@ll.mit.edu