AI Summer Program
AI Summer Program
The Iribe Initiative for Inclusion and Diversity is pleased to offer the TRAILS AI Summer Academy at the University of Maryland.
About
The TRAILS AI Summer Academy is a two-week long, nonresidential computer programming and artificial intelligence (AI) summer camp at the University of Maryland. Students will come away from the camp knowing how AI can be used to help people and an idea of what kinds of careers there are in AI. Accepted students will have to complete ~25 hours of asynchronous content prior to the start date.
The camp will be offered to rising 10th, 11th, and 12th graders. Students will be exposed to personal growth, education, and hands-on experiences presented by faculty, guest lecturers, and University of Maryland students.
This program intends to create a more inclusive and diverse field of artificial intelligence by targeting and serving underrepresented communities. Students will be given the opportunity to use artificial intelligence to address problems of a probabilistic and numeric nature. Participants will explore the field of AI through team projects, industry field trips, and presentations from guest speakers. There will also be opportunities to engage with faculty, staff and researchers who have been leaders in AI. Students will be exposed to a breadth of knowledge in the field with the goal of leveraging AI for social good.
The program will focus on three aspects:
- AI education and inspiration
- Personal growth
- Hands-on research experience.
Typical Camp Day
- 9:00-12:00pm Classroom Instruction*
12:15-1:15pm Lunch
1:30-5:15pm Classroom Instruction
*Field trips and guest speakers are scheduled during Classroom Instruction time blocks.
Lab Meetings will take place in the Brendan Iribe building.
AI Education
- Formal AI education curriculum instruction by a local AI high school teacher
- Guest Speakers by UMD professors and industry professionals
- In-depth introduction to ongoing research projects from faculty
- Field trip to AI industry leaders where students are introduced to people, topics and career opportunities
Personal Growth
- Discussions lead by experts in career and personal development
- Small group mentoring with AI faculty and graduate students
- Social events with peers
Hands-on Experience
- Small-group research project led by faculty or graduate students; projects focus on using AI for societal good
- Group presentations showcasing work at the end of the program
- Applicant must be able to attend both weeks.
- Applicant must be a rising 10th, 11th, or 12th grader.
- Applicant will be required to submit family and student information.
- Financial Assistance is available for those with displayed need by completing our Scholarship Application.
- Must submit academic transcripts (unofficial transcripts are welcome for application review).
- Emails for teacher recommendations are required
- I4C's AI Summer Program is no longer affiliated with AI4ALL as of 2023. For more information, please visit: https://medium.com/ai4allorg/changes-at-ai4all-a-message-from-ai4alls-ce...
Dates & Links
Program Date and Quick Info
Dates: July 8 - July 19 (25 hours of asynchronous content will be completed prior to the start date)
2-week nonresidential program experience
Target Student: Rising 10th, 11th, and 12th graders (focus on the DC, MD, and VA areas)
2024 I4C Summer Academy applications are closed.
Most recent grade report (transcript or report card): https://go.umd.edu/sum24Grades
Projects
2023 Projects (2024 Projects will be posted in May 2024)
Deep neural networks have become popular in many fields achieving state-of-the-art results in many domains and tasks. However, they also bring up certain questions - chiefly, how do neural nets do what they do? The building blocks of neural nets are very simple, but they give rise to complex behavior which cannot be easily understood. One way to make progress on this question is to modify the input to the neural network in some specific way and observe the corresponding output. This gives insight into what parts of the input are relevant to the neural net. Using this principle, we proceed to use adversarial attacks as a tool to understand neural networks. An adversarial attack automatically finds what parts of the input need to be changed in order to maximize the change in the output. Using this as a tool, we study the behavior of normal and adversarially robust neural networks. We will find that neural networks can be surprisingly sensitive to minute changes in the input. For example, an image of the dog can be modified so slightly that humans can’t even tell the difference, but the neural net thinks that the image is actually a traffic light. We will then study how to make the networks robust to such changes in the input using a technique called adversarial training. This involves training the AI to specifically ignore small changes in the input as these do not change the image content significantly. However, this comes at the cost of accuracy - a tradeoff which is not yet completely understood. Interestingly, since adversarially robust models are not sensitive to small changes, adversarially attacking these models generates some new, surreal images. This technique of visualizing what the AI is focused on via adversarial attack is very powerful and gives new insight into how neural networks operate.
Researchers
Assistant Professor, UMD Department of Computer Science
Graduate Student
AI has generated tremendous excitement in several high-stakes applications, e.g., hiring, lending, etc. that directly influence people’s lives. With the growing use of machine learning in high-stakes decision-making, there is a growing interest in understanding how these models make their decisions. For example, if a loan gets denied, one might be interested in knowing which features were the most important in that decision. In this project, students will get to implement machine learning models, and then apply some explainability techniques, such as, SHAP to quantify the contribution of different features to the overall decision.
The next part of the project would deal with the question of substitute features. Once the most important features have been identified, an interesting experiment would be to drop the most important features, retrain another model, and then see which features are the most important now. Interestingly, features that were not at all important in the first experiment
Researchers
Assistant Professor, UMD Department of Electrical & Computer Engineering
Recent years have seen the tremendous successes of machine learning, especially reinforcement learning (RL), where an agent makes sequential decisions by interacting with the environment via trial and error. Prominent and high-profiled examples include AlphaGo, an intelligent Go game-playing agent that has beaten the human champions, autonomous driving, and more recently the training of large-language models such as ChatGPT. Many of the successful stories are concerning the scenarios where there may exist “multiple” decision-makers/learning agents, and they make individual and strategic decisions, with possibly misaligned objectives. An example is autonomous driving, where each agent (self-driving car) has its own goal, while they are coupled with each other by interacting with each other on the road, which may cause congestion if the fleet of cars are not scheduled properly. Thus, it is natural to study the theme of “multi-agent” reinforcement learning, and the behavior of multiple RL agents when they coexist in a common environment. Our goal is to first get familiar with the concept of reinforcement learning and sequential decision-making, and then the concept of multi-agent RL. Further, we aim to develop new multi-agent RL algorithms that may be useful for settings beyond game-playing, e.g., video games and Go games, which mostly focused on “competition” among agents, but can be used to “encourage” the cooperation among them for social good, despite the fact that they may have very different objectives. Along the way, the students will also get familiar with the use of Python, a useful programming language, as well as the basic mathematical concepts in related areas such as optimization, statistics, and game theory.
Researchers
Assistant Professor, UMD Department of Electrical & Computer Engineering
Graduate Student, UMD Department of Computer Science
Graduate Student, UMD Department of Computer Science
People are pretty good at getting around. When moving in a crowd, we’ve figured out how to stay close to the people in our group, and move away from those that are not. Our behavior depends on the type of interaction. We can also infer others' objective by implicit interaction with each other through motion. Robots, on the other hand, rely on explicit communication or instructions to avoid collisions or getting stuck while moving towards a goal. The ability to model these interactions can help robots predict the uncertain behavior of pedestrians in the absence of explicit communication. That way a food delivery robot, for example, could autonomously and safely navigate through crowds without human help.
Researchers
Assistant Professor, UMD Department of Mechanical Engineering
Riddles have been around in the West for around 2500 years, representing a cognitively demanding task that requires people to reason carefully about the world. In this project, we will explore how AI systems can be used for generating and answering riddles, as a specific example of the broader question of how AI systems are built in general. Students will have the opportunity to test AI’s understanding of a fun riddles data set, interactively playing with a chatbot such as ChatGPT and training their own models for answering riddles. Specifically, we will walk students through a reasoning task such as having the AI assess the answer to the riddle and create its own riddles. We will learn how to prompt large language models to perform tasks, how to use crowdsourcing to collect human judgments, and how to train our own natural language processing models to answer riddles.
Researchers
Volpi-Cupal Family Endowed Professor, UMD Department of Computer Science
Graduate Student
Graduate Student
Graduate Student