My work focuses on applying reinforcement learning to the routing process of VLSI Physical design. I am working on designing several parts of the routing problem as reinforcement learning problems to get a better routing solution for the bottleneck nets in comparison with current optimization and heuristic methods. The motivation to use reinforcement learning generates from insufficient routing data available in the academic community. Using reinforcement learning algorithms enables me to train an RL agent without any external data.  My research mainly has two components. First to train an agent on how to route a net and second to train an agent to identify bad nets and re-route (called rip-up and re-route approach) them with the current S-O-A router. Specifically, global routing is considered, and the objective is to optimize total wirelength, total Vertical Interconnect Access (VIA), short violations, and runtime. For the first part of the research, I have finished working on the routing part and proved as a proof-of-concept that reinforcement learning can be used to route 2-pin nets to successfully solve short violations in small circuits. For the second part, a collaborative framework has been developed with a routing part and a reinforcement learning part. This is ongoing research work. 

Basically, I have been doing Machine Learning research applied to EDA for routability, more specifically, DRV (Design Rule Violation) prediction. During the physical synthesis in an EDA flow, you have a netlist (logic instances + macros, etc, and their connections) that needs to be placed and routed. Nowadays, routing is one of the most computationally expensive steps because you have millions of logic gates to be performed using, for instance a 3D grid graph computation model. Whenever a router finds a problem, it needs to perform rip-up-and-reroute, or even worse; the placement needs to be redone. As you can see, placement and routing are challenging steps that need to be constantly evolved to be able to handle current design technology. Therefore a routing assessment could be performed at former steps to prevent some of those routing problems. In my work, I am trying to apply machine learning to identify (predict) which parts of netlist are prone to routing problems, so the designer could try to address it before it happens.

The underrepresentation of women in STEM industries is a multifaceted problem, and gendered language in advertisements for these jobs plays a key role in this issue. Gendered wording commonly employed in job recruitment materials can maintain gender inequality in traditionally male dominated jobs. In Gaucher’s paper [?], evidence that gendered wording in job advertisements exists and sustains gender inequality, Gaucher has shown that the specific language used in the adverts for job postings has an impact on professionals’ perceptions on if they are eligible for a job or not. Gendered language used in these advertisements is believed to be one factor that impacts the decision-making process of male or female prospects on applying for the job. Thus, this project aims to identify, minimize, and predict gender-oriented language and its impact on the STEM job industry.   

The first step has mainly been completed. It was to scrape a variety of job adverts (this study has used Indeed) and determine the trends that are currently happening with both STEM (Science, Technology, Engineering and Math), which is more male-dominated, and HEAL jobs (Health, Education, Administration, Literacy) which are more female-dominated. Data has been collected for the years 2019, 2020 and 2022 (pre, during and after covid). In 2019, the majority of the adverts were coded feminine, however the qualifications for the adverts were coded more masculine. As we continue throughout the years, wording has shifted to much more feminine wording for both STEM and HEAL jobs.   

Kirill is a part-time master's student working on applying ML and AI to human learning as of January 16, 2023.

He says:

I am currently at Microsoft/Microsoft Research working on robotics and applying Reinforcement learning to control systems problems.

For my research, I have been trying to use Large Language Models (like GPT-3) to label exam questions with their levels of understanding (these levels/classes are from an educational system called Bloom's Taxonomy).  My initial experiments last year failed to get reasonable results so I am still looking into alternate projects I can work on. One is re-trying the experiments with newer LLMs (GPT-3.5), and the other thing currently I'm looking into using generative models to create explanations of content (I'm still looking into what would be the best domains in this context, but the most likely candidate would be to create educational visualizations from text).

An engineer’s scope of responsibility used to be limited to calculations of safety factors, but there is now an awareness of the damage this limited scope caused. Climate change, indigenous boil water advisories, and bias in AI are all prime examples of how decisions made by engineers have had devastating and traumatic impacts on the natural world, indigenous communities, and minoritized identities. Through engineering education, we need to consider the narratives we are teaching students, and we need to prepare engineering students to take a community-based approach to their responsibilities. In this work, I explore how critical pedagogical interventions can build new narratives to positively change the norms and culture of engineering education. 

I developed and implemented three inventions at a variety of time scales: 

· 15-minute mental wellness modules in first-year engineering 

· 4-week bioengineering physics program to provide broader access to engineering  

· 4-month integrated learning stream in second-year electrical engineering  

I am about to start my second Ph.D. year. Currently, I am working on a literature review of Detailed Routing in Electronic Design Automation to help guide my research. As of recently, the state-of-the-art works have transitioned into doing both Global Routing and Detailed Routing together, which achieves better results than separately. Some other works also have been backtracking all the way to placement, running Detailed Placement, Global Routing and Detailed Routing in the same framework. 

My current plan is to explore the subject of Detailed Routing in isolation and see if there are new techniques that can either improve the result further than the state-of-the-art or match the quality with improved run-time. Some of these ideas could be simply reimplementing the state-of-the-art algorithm and then changing parts of it, post-processing the routing or even proposing an entirely new algorithm.