Mehdi Afsar
Welcome! I have more than 10 years of experience in programming, software development and research. I received my PhD degree in Software Engineering from University of Calgary. During my PhD studies, I worked on recommender systems, machine learning, NLP, and intelligent agents/chatbots. Prior to joining UofC, I completed my MSc and BSc degrees in Computer Engieering in Iran.
Experience
ML Researcher
- Developed various recommender systems using ML techniques, all in Python.
- Led a team of grad and undergrad students to develop a healthcare chatbot.
- Developed a web scraping tool in Python to crawl large websites, like Wikipedia and cancer forums.
- Mined various projects from millions of repositories on GitHub using a big data software repository mining infrastructure based on MapReduce and Hadoop.
Instructor
- Data Structure and Algorithms (winter 2022) and Java Programming (winter 2020).
- Designed and marked course materials and slides, labs, tutorials, assignments, quizzes, and exams.
- Maintained regular scheduled office hours to advise and assist students.
Intern
- Developed, trained, and evaluated the first GAN model to synthesize diverse, realistic laughter.
- Built a CNN classifier of laughter spectrograms (images) with more than 90% accuracy.
- Performed GAN inversion and laughter interpolation (e.g., changing male to female, child to adult laughter).
Software Developer & Counselor
- Developed several discrete-event network simulators using Java, C++, and MATLAB.
- Implemented an automatic screen sensitive to brightness (unregistered patent).
- Built various commercial applications and modules, like document management, registration system, and contact manager, using Python, Java, and C++..
Education
University of Calgary
Projects
ALAN, a healthcare chatbot
We designed and implemented two prototypes of a chatbot to play the role of a patient decision aid. We used a multi-layered, multi-agent architecture for the design of ALAN, demonstrated in the figure below.
In the first prototype, we developed a web application containing that could communicate with pateints and physicians, understand the language, and provide some services, like fetching the most recent survey filled by the patient or visit summarization. We used SQL Server for the database management. The second prototype consisted of impelentation of ALAN in JADE (a Java platform for multi-agent systems). The architecture of the second prototype is as follow.
We used a simple GUI for the second prototype, depicted below.
ALAN could provide various services for both pateints and practitioners and show the response to the user in the most logical way. For example, a sample of retrieved data by ALAN is as follows.
Sound of Laughter
During my internship at Mila, I worked on an art project offered by National Film Board of Canada. We aimed at developing an interactive web platform that connects people through synthetic laughter. We wanted to explore the phenomenon of empathy triggered by laughter, the relationship between individual memory and laughter, and how the sound of laughter evolves over a lifetime. To this end, we trained GAN models in order to synthesize real, diverse laughter. We trained our model on the spectrograms of our laughter audio dataset. The followings are two spectrogram samples of the data; see how similar they look like.
Also, you may also listen to some samples generated by our models. For more details, please check our paper presented at NIPS, MLCD workshop.
Reinforcement Learning based Recommender Systems
Recommender systems (RSs) have become an inseparable part of our everyday lives. Traditionally, the recommendation problem was considered to be a classification or prediction problem, but it is now widely agreed that formulating it as a sequential decision problem can better reflect the user-system interaction. Therefore, it can be formulated as a Markov decision process (MDP) and be solved by reinforcement learning (RL) algorithms. In this research project, a survey on reinforcement learning based recommender systems (RLRSs) is presented. We propose an RLRS framework with four components, i.e., state representation, policy optimization, reward formulation, and environment building, and survey RLRS algorithms accordingly. This work reviews 97 RLRS algorithms, a summary of which is presented in the following figure.
