Toronto Human-AI Interaction
Summer Research School
June 6 - August 12, 2022
June 13 - August 5
Design a Study
Collect and Analyze Data
Write and Present
Intensive Week 1
Intensive Week 2
Supervised Group Work on a Research Project
Daily Research Talks
Daily Research Skills Workshops
Time for Project Work
Applications are OPEN until April 27, 2022
Daily Research Talks
Daily Research Skills Workshops
Time for Project Work
Due to potential complications with traveling, the THAI RS 2022 will take place online.
Welcome to the second Toronto Human-AI Interaction summer research school!
The THAI RS program includes two weeks of intensives, separated by eight weeks dedicated to flexible work on group research projects.
Group Research Projects
Two Weeks of Intensives
Each day during the Weeks of Intensives includes a research keynote from one of the leading researchers in the area of human-AI interaction, a hands-on research skills workshop, and time for group work.
Accepted students are expected to attend and participate in all the daily activities.
The daily participation during the Weeks of Intensives (June 6-10 and August 8-12) includes attendance of research talks (approximately an hour), participation in workshop activities (approximately 3 hours), and participation in group work sessions (approximately 2 hours).
By the end of the first Week of Intensives, each project group is expected to complete the research proposal, which includes the design of their study and an initial review of the related work. The groups will have a chance to work on their proposals during the workshops activities and dedicated group work sessions.
During the last Week of Intensives, each project is expected to be presented at the final symposium.
Eight Weeks of Flexible Group Work
During the eight-week period between the Weeks of Intensives, each project group will work on its own schedule. Each group is expected to have bi-weekly meetings with their assigned research advisor for check-in and consultation. During this period, each group is expected to follow the progress schedule, determined by their research advisor.
The specific workload depends on each project group's decisions but can be expected to be approximately 4-5 hours a week. By the end of this period, each project group is expected to complete the data collection (when applicable), analysis of the collected data, and synthesis of the results.
Eligibility & Application
Eligibility: We invite applications from students who are currently enrolled (or were officially accepted) in a Master's program at any Canadian university. The types of invited programs include, but not limited to the areas of Information Science, Computer Science, Data Science, Artificial Intelligence, Psychology, Systems Design Engineering, Human-Computer Interaction. This year, the applications are also open for those who have a Master's degree and currently work in a relevant industry field.
I'm interested! How to Apply? Please see more details here.
Each accepted student will receive a Participant Honoraria of 1,000 CAD.
Upon completion of the program, each student will receive a certificate of participation.
Master of Information students from the University of Toronto will be able to additionally receive a reading course credit (upon request).
Application Deadline: April 27, 23:59 pm EST
Each project is performed in a group of 2-4 fellow summer school students. The groups are formed according to their topics of interest, chosen by each student during the application process, and based on the complementarity of the skills of group members. The groups are formed following the acceptance of a student to the THAI RS and prior to the first week of Intensives.
Each project is assigned a research advisor from our team of University of Toronto professors.
Each project is expected to result in a written report and will be presented by the project group at the symposium on the final week of intensives. Particularly successful projects might be invited for academic publication.
Topic 1: Identifying and Mitigating Confounds During Mobile EMAs
Description: Ecological momentary assessments (EMAs) sent through mobile devices make it possible to gather a quick snapshot about a person's behavior and context during everyday living. In recent years, there has been interest in distributing validated cognitive tests via mobile EMAs to probe a person's mental capabilities throughout the day (e.g., NIH Toolbox). Although mobile phones are convenient, the environments in which they are used can lead to numerous confounds that impact how people perform on these tests. We will explore various mechanisms for identifying and mitigating such confounds. These mechanisms may include but are not limited to, conversational agents or passive sensing.
Skills/Experiences/Interests (in any combination):
Mobile app development (Android and/or iOS)
Automated messaging service (e.g., Twilio)
Small amount of signal processing and statistical analysis (e.g., Python or R)
Communication/conversation analysis and/or design
Topic 2: Justified AI: Improving Human-AI Interaction through Justification
Description: Artificial Intelligence (AI) applications often work as a 'black box', and the users do not understand why the applications are making a certain action. While there have been several efforts made to generate explanations for AI applications, they did not work well because they did not take into account the user's background. Moreover, not every explanation is justified and acceptable to every user. This problem has kept many AI applications untrusted by users. To address this problem, we will design, develop, and evaluate a new kind of AI application that will produce a justification of its action based on the background of the users. We will develop this system based on two AI software systems that we have developed to minimize the hate speech and misinformation on social media.
Natural Language Processing (NLP)
Topic 3: Developing Predictive Models to Characterize Covid Long Haul Patient Symptoms
Description: COVID-19 long-haul (CLH) patients are defined as individuals who have tested positive for COVID-19, both mild and severe infections, and report symptoms that linger for at least 4 weeks beyond the initial infectious state. The typical recovery time for mild COVID-19 cases is 11.1 days and severe cases 28.6 days. However, CLH patients have reported chronic, intermittent, and new symptoms for four or more weeks after their initial diagnosis, even patients who were initially asymptomatic. Research on CLH patients as a domain is in its infancy. The CDC estimates that one third of COVID-19 patients who were never acutely ill enough to be hospitalized will become a CLH patient. While patient-reported outcomes are important to research, the collection of objective data is critical within the clinical realm. A common way of assessing objective health data is by the use of activity trackers. Consumer off-the-shelf activity trackers like Fitbits have proven to be effective in providing real-time data for movement/exercise, heartrate, and sleep. Sleep is an especially critical component in understanding CLH patients due to the frequency of symptoms like brain fog, impaired concentration, and fatigue. This project is a collaborative project with Parkview Health where the objective is to establish a predictive model framework to understand CLH symptoms from electronic health records, activity tracking and patient reported outcome measures
Topic 4: Inclusive Interactions with Smart Speakers for Marginalized Populations
Description: Voice interfaces open exciting opportunities for addressing needs of diverse marginalized populations. For instance, voice interfaces can increase the inclusiveness of AI systems for older adults by helping address visual, motor, and cognitive challenges present in other types of interfaces. But, if voice interfaces are not designed carefully with full understanding of the unique needs of older adults, more barriers to access may occur. This project will investigate how older adults perceive the voices of the current smart speakers (e.g. Alexa, Google Home, etc.), including the users' preferences of voice characteristics and specifics and underlying mechanisms of interaction preferences. In particular, the aim of this project is to investigate older adults’ perceptions and preferences in the personalities and conversation styles of current smart speakers in order to better understand how to design such systems to meet older adults’ needs.
Conversational/voice user interfaces
Qualitative and/or quantitative analysis
Topic 5: Developing An Automated Conversational Agent for Motivational Interviewing
Description: A team from the University of Toronto and the Centre for Addiction and Mental Health (CAMH) is developing an automated conversational agent to help smokers move toward the decision to quit smoking. The agent uses a known clinical technique, called Motivational interviewing, which encourages clients to reflect on the habit. The current version of the conversational agent, called MIBot, uses state-of-the-art AI Transformer-based text generation (based on the GPT-2 transformer) and text classification methods. It has been trained using expert labelled data, and those experts are part of the broader team. The goal of the project is to analyze transcripts of MIBot-client interactions using a variety of qualitative and quantitative methods and evaluate their usefulness in understanding and predicting client impacts. The team will have the opportunity to influence the evolving design of the MIBot chatbot.
Select and explore quantitative text analysis techniques including classification, relationship analysis, clustering, sentiment, topic modeling, and cosign analysis
Carry out select techniques
Provide evidence-based recommendations for use of specific quantitative techniques, and suggestions for improved conversational strategies
Topic 6: Predicting How Risk, Protective and Systemic Factors Lead to Biases within Child Welfare Systems Using Deep Learning Language Models
Description: Recent work within the HCI and AI research communities has highlighted the significant impact of uncertainties on decision-making in any high-stakes domain. There is an irreducible degree of uncertainty associated with any predicted decision which has important implications, especially for algorithmic decision-making in the public sector where mistakes can have serious consequences for the lives of citizens [1,2,3]. Consequently, researchers have highlighted the need to uncover important sources of uncertainties because it is at these points that human discretion is most needed as well as allow for a closer inspection of confounding factors that can lead to erroneous decisions from both algorithmic systems and human decision-makers . In the child-welfare system, caseworkers are trained to write detailed casenotes which encapsulate critical decisions, circumstances surrounding these decisions, families’ interactions in their social ecosystem, uncertainties in a case, as well as competing factors that confound caseworkers’ judgments. As illustrated by our recent study on computational narrative analysis of casenotes, these contextual factors can be derived from this unstructured text and can help inform decision-making. In this next study, we will employ transformer-based NLP models such as BERT to inspect casenotes and assess how different risk, protective, and systemic factors mediate one another through the life of a child-welfare case, how uncertainties arise (at the intersection of these factors), and finally, how these critical insights can be meaningfully presented to a child-welfare team such that it informs a collaborative decision-making process
Topic 7: Designing Inclusive Risk Messaging for People with Visual Disabilities
Description: People with visual disabilities face significant barriers to receiving timely and instructive information on risk and safety during severe weather events. Such information is necessary for “protective decision-making” – decisions that individuals and communities make in advance of and during disasters that impact their safety with regard to these events. In this project, students will use inclusive design methods to design and prototype a risk messaging authoring tool for this audience. The project will consist of analysing existing datasets of weather risks, training a lightweight AI application that will transform these datasets into more accessible forms, and creating inclusive interfaces to present them to people with visual disabilities.
Accessibility and Inclusive Design
CAD Participation Honoraria to Each Student
Timely Project Topics on Human-AI Interaction
Weeks of Group Work on Research Projects
Weeks of Research Talks and Hands-On Workshops
Research Project Advisors:
Proud to bring outstanding researchers from the University of Toronto
Speakers and Workshop Instructors:
April 1st - 27th
Week 1 of Intensive
June 6th - June 10th
Week 2 of Intensive
August 8th - August 12th
Please prepare the following for your application process:
1) Your most recent CV/resume. Please submit as a .pdf file with the file name in format CV_LastnameFirstname.pdf
2) Your choice of at least 3 project topics (you will be prompted by the application form). Your preferences and experience will help us to form project groups. However, please note, that we cannot guarantee that, if accepted, you will be invited to one of the topics you identify.
3) A one-page research statement, describing your motivation to participate in the THAI RS, relevant experiences, interest in the chosen research school topics, and any other information you would like us to consider. The research statement should be submitted as a .pdf file with the file name in format RS_LastnameFirstname.pdf