High school and college-age Changemakers in AI collaborate with industry mentors, completing innovative research and outreach projects. Header image caption: Changemakers and AI4ALL staff at UC Berkeley AI4ALL 2018.
Launched in 2018, Changemakers in AI builds on educational foundations laid in AI4ALL college and high school programs, building community and offering students continued opportunities for leadership and technical growth. Now comprising over 750 college and high school age youth, Changemakers in AI supports AI4ALL students at all points of their learning trajectories, understanding that multiple touchpoints are necessary to influence long-term student outcomes and change the direction of the AI field. In Special Interest Groups, Changemakers explore research topics, working to create deliverables under the guidance of mentors from AI academia and industry. These deliverables take the form of research papers, websites, walkthroughs, and calls to action. Special Interest Groups are always student-led.
This winter and spring, 63 Changemakers formed 8 research groups, building tools to teach AI understanding and literacy and models to study AI applications in health care, sustainability, and policy. Learning from mentors that include researchers from the University of California, San Francisco, Georgetown University, New York University, the Ohio State University, and Princeton University and industry leaders from Workday, the American Civil Liberties Union, and the AI4ALL curriculum team, Changemakers investigated applications, limitations, and ethical considerations of current AI models. Read on to check out Changemakers’ full deliverables.
“This experience increased my confidence in leading a group in discussions and reaching out to mentors. It helped me figure out that I want to pursue a career in something related to AI ethics…I feel like I have a clear direction I want to take my career in that I am really passionate about.”
— Fall 2020 Special Interest Group Leader, AI4ALL High School Student
AI & Health
A Comprehensive Review of the Efficacy of Various Machine Learning Algorithms on the Diagnosis of Psychiatric Disorders
As mental illnesses are increasingly diagnosed in patients of all ages, Changemakers investigate the efficacy and ethics of using machine learning in the prediction and treatment of psychotic, mood, eating, and personality disorders.
In recent years, machine learning (ML), a subset of AI, has been introduced to mental health providers, such as psychologists, psychiatrists, and therapists, in the healthcare and criminal justice fields, for its decision-making prowess based on patient data — medical records, survey responses, social media usage, family history, and behavioral data. With big data generated in healthcare, ML algorithms have been increasingly employed for prediction and diagnosis due to their flexibility, scalability, and ability to analyze diverse data types. In this study, Changemakers review the current research on the implementation of various ML algorithms in the diagnosis of a subset of mental disorders — psychotic, mood, eating, and personality — and evaluate the efficacy of different ML algorithms in establishing early preventative models for diagnosis. Supervised learning models like support vector machines (SVMs) are widely used for image classification, and natural language processing (NLP) techniques are used to analyze electronic health records (EHRs) and other text notes, while neural networks including CNNs and recurrent neural network (RNNs) are suitable for pattern recognition in medical time-series data analysis. Changemakers then discuss the limitations of each algorithm. The most significant challenges faced by ML algorithms for the diagnosis of mental health disorders stem from the heterogeneity of symptoms, limited sample sizes, and lack of diversity within datasets, leading to bias and overfitting. The majority of the data collected in the studies Changemakers surveyed were skewed with respect to socioeconomic status, gender, and race. In developing new AI-based systems to diagnose mental illnesses, it is imperative that researchers examine the underlying biases in the data, methods, and algorithms.
Changemakers: Isha K (Group Leader). Alyssa T (Group Leader). Raunak A. Rahi D. Diya J. Camilla M. Adithi R. Alina S.
Mentor(s): Julienne LaChance, PhD candidate, Dynamics and Control, Princeton University
Comparing the Architecture and Performance of AlexNet, Faster R-CNN, and YOLOv4 in the Multiclass Classification of Alzheimer Brain MRI Scans
In this study, Changemakers apply convolutional neural networks to analyze MRI images, introducing methods to diagnose Alzheimer’s Disease more accurately and increasingly efficiently than medical professionals.
Changemakers created four classifications from the Kaggle brain MRI dataset to train their models: (a) Non-Demented (b) Very Mild Demented © Mild Demented (d) Moderate Demented.
Alzheimer’s Disease (AD) is the sixth leading cause of death worldwide and the most common form of dementia, accounting for up to 80% of dementia diagnoses. Advancements in diagnostic imaging such as magnetic resonance imaging (MRIs) have led to a greater understanding of the diagnosis and treatment of AD; however, analyzing large and complex MRI datasets is time-consuming and vulnerable to errors, presenting a challenge for clinicians. In this study, Changemakers introduce three applications of convolutional neural networks (CNNs) to perform multiclass classification of brain MRI scans of AD patients: AlexNet, Faster R-CNN, and YOLO. CNNs are a class of neural networks that are most commonly applied to computer vision and represent an analogy to the neuron connectivity pattern in the brain. Changemakers hand-labeled their training images, obtained from a publicly available database on Kaggle. For practicing radiologists, algorithmic prediction can supplement professional judgment of a patient’s AD condition. The biggest limitation, for both our Changemakers conducting this research and in widespread CNN application in MRI diagnostics, is the lack of representation of images in MRI datasets. To function at their highest capability, CNNs need data representative of a diverse array of patients at all levels of disease progression. For the full potential of AI to be realized, the data must include all of us.
Changemakers: Ria M (Group Leader). Caroline Y (Group Leader). Supraja G. Archita K. Alyssia N. Anupama N. Sonica P.
Mentor(s): Ina Chen, PhD candidate, Bioinformatics, the University of California, San Francisco
AI & Sustainability
Utilizing Machine Learning Algorithms to Model Water Quality Based on Indian Environmental Data
Using publicly available Indian Water Quality data, Changemakers developed ML algorithms that accurately predict water quality by geography, empowering users to mitigate outbreaks of preventable disease and protect themselves, both in health and from financial loss.
Changemakers used 6 indicators of water quality to calculate and plot India’s overall Water Quality Index score by year.
According to the World Health Organization, at least 2 billion people globally According to the World Health Organization, at least 2 billion people globally use a contaminated drinking water source. Absent, inadequate, or inappropriately managed water and sanitation services expose individuals to preventable health risks, and are linked to the transmission of diseases such as cholera, diarrhea, dysentery, hepatitis A, typhoid, and polio. Water pollution is a major crisis in India, where around 70 percent of water sources are contaminated. ML algorithms have shown considerable promise in the classification and prediction of fluctuations in water quality, based on key environmental factors: dissolved oxygen, pH, conductivity, biochemical oxygen demand, nitrate amount, and total coliform. In this project, Changemakers use multiple ML-based techniques to create a model that can analyze Indian water quality data and accurately predict water quality. The team used the publicly available “Indian Water Quality Data” dataset available in Kaggle for their analysis. The team’s Random Forest Regressor model most accurately predicted water quality and can be applied to make predictions based on any water quality data in India or utilized for datasets from geographically diverse locations.
Changemakers: Ayush R (Group Leader). Hiya S (Group Leader). Isita T (Group Leader). Archika D. Anakha G. Aroshi G. Audrey K. Ayush R. Hannah Z.
Mentor(s): Sam Shah, Software Engineer, Workday; Beth McBride, AI Education & Instruction Manager, AI4ALL Summer Programs
Predicting a Country’s Sustainability Using Artificial Neural Networks and Public Datasets
Focusing on concrete indicators of climate change like carbon emissions and declining water quality, Changemakers built an artificial neural network that can analyze sustainability data by country, and assign a “numerical value of consumption” to guide development decisions.
To prevent unsustainable practices, data and analytical models can be used to guide planning and decision-making. Standard statistical models need many data entries and very clean data to accurately predict future sustainability metrics. Unfortunately, collecting sustainability data often involves surveying various entities and individuals, which are expensive and time-consuming processes, with oversights, logistical issues, and uncooperative subjects, leading to missing training data. In this proposal, Changemakers use artificial neural networks (ANNs) to predict whether a country is sustainable or unsustainable, as ANNs can learn hidden correlations between features, inferring whether an entity is sustainable or not sustainable based on fewer indicators. The “World Development Indicators by Countries” dataset provided by the World Bank was compared to the United Nations’ sustainability goals to identify which criteria were indicators of a sustainable society, like carbon emissions and freshwater usage. The main limitation of this project was that the world development dataset was small, unsystematic, and contained many missing values. In the future, this project can be improved by using larger, more complete, datasets.
Changemakers: Aniket R (Group Leader). Michelle W (Group Leader). Ashna K. Elena P L. Chidozie N. Yaquelyn R. Emily W.
Mentor(s): Dr. Tanya Berger-Wolf, Computational Ecologist, the Ohio State University
AI & Ethics
4DataPrivacy
4DataPrivacy is a website created by Changemakers that brings public awareness to data rights and privacy laws and gives individuals concrete ways to advocate for the passing of equitable consumer data laws on the local, state, and federal levels.
As AI becomes integrated into every facet of society, public awareness of data rights and privacy issues remains concerningly low. Outlining two user rights, the “Right to be Informed” and the “Right to Contest,” Changemakers share how the technology we use every day breaches these user rights. Changemakers then share national and international precedents for regulation, such as the European Union’s right of access and Vermont’s data broker law, and steps users can take today to protect their data and privacy online. Understanding that public awareness is crucial to the fight for data rights, the website concludes by proposing a letter-writing campaign, offering constituents a template to send to their representatives. The letter advocates for the passing of three Changemaker-developed rights, the Right to Be Informed, the Right of Access, and the Right to Object/Contest.
Changemakers: Ezra F (Group Leader). Ines K (Group Leader). Shivani A. Esha G. Chaira H. Miseok K. Rachana K. Kaitlyn T.
Mentor(s): Albert Fox Cahn, Executive Director, the Surveillance Technology Oversight Project
The State of Facial Recognition in America
Joined by policy experts from Georgetown Law and the American Civil Liberties Union, Changemakers moderated a freely available webinar on the current use of facial recognition technology, data privacy, and user rights.
Held on March 1, 2021, the State of Facial Recognition in America webinar explored how what the growth of facial recognition technology means for governance and our everyday lives.
Despite the increasing prevalence of facial recognition in our day-to-day lives — from Snapchat filters to unlocking our phones — few truly understand the full implications of facial recognition’s use. Changemakers conducted research on the current ways facial recognition technology is used by the United States government, uncovering how unethical applications erode the privacy of the everyday citizen and reinforces the biases present in our society. Recent advancements in data storage and processing have given rise to the ability for organizations to collect, store, and process vast amounts of data, while effectively eliminating the need to ever delete user information. As data is collected for legitimate purposes, large amounts of personally identifiable information (PII) are inevitably collected as well. PII is described by the US Department of Labor as “any representation of information that permits the identity of an individual to whom the information applies to be reasonably inferred, by either direct or indirect means.” Data privacy pertains to the entitlements and rights of individuals regarding PII, such as the rights to forget and consent. On March 1, 2021, Changemakers held an educational webinar, allowing participants the opportunity to ask data privacy experts difficult questions directly.
Changemakers: Dara C (Group Leader). Alisher Y (Group Leader). Dhriti G. Anushka M. Brenda S. Peyton S. Imani W. Stephanie X.
Mentors: Clare Garvie, Senior Fellow, Georgetown Center on Technology & Policy; Emiliano Falcon-Morano, Policy Counsel, Technology for Liberty Program, ACLU of Massachusetts
AI & Education
Introduction to Artificial Intelligence for Beginners
Designed by AI4ALL Changemakers, the Introduction to Artificial Intelligence for Beginners curriculum breaks down the five fundamental branches of AI — computer vision, natural language processing, neural networks, supervised vs unsupervised learning, and ethics— and takes you step-by-step through the code necessary to build your own classification network, using labeled images to train a model.
The sigmoid function “activates” the model, compressing the contents of the output matrix into values ranging from 0–1.
Changemakers reached out to 220+ schools in various states including Massachusetts, Arizona, Alaska, California, New York, and Pennsylvania with Introduction to Artificial Intelligence for Beginners, a set of curriculum to teach AI in high school classrooms. The curriculum includes Neural Network Walk-Through, an interactive project walk-through website designed by Changemakers that takes you step-by-step through the creation of your own classification network. A classification network utilizes reinforcement learning, which is when labels for corresponding pictures in the dataset are used to train the model. Follow the website’s code in your own browser as Changemakers walk you through everything necessary to build a neural network model, including the selection of a dataset, how to use code to develop the neural network layers, and how to train the model using data. Learn how to begin building your own models and making your own predictions with the guidance of AI4ALL Changemakers.
Changemakers: Stella C (Group Leader). Nikhita P (Group Leader). Mehak G. Ivy Z. Jessica T. Jia Y L. Anya G. Olivia G.
Mentor: Sarah Judd, Curriculum Manager, AI4ALL Open Learning; Océane Boulais, Machine Learning Engineer, NOAA Fisheries & Northern Gulf Institute
JumpstartAI
JumpstartAI is a student-led initiative to expand AI access to all middle school students through interactive weekend workshops, covering AI fundamentals and ethical considerations.
High school-age Changemakers held JumpstartAI, a 3-day interactive AI camp for middle schoolers, from March 12th — March 14th, 2021. JumpstartAI exists to give students, regardless of zip code or income level, a fun introduction to AI. Leveraging an introductory middle school AI curriculum from MIT, Changemakers guided students in creating simple algorithms to categorize the ingredients necessary to make their favorite meals, introducing concepts like data selection and optimization. Changemakers then discussed how bias enters algorithms and guided students in identifying their own subconscious biases. The event concluded with a discussion on the ethical concerns with AI’s use today, successfully reaching 25 middle school students, and equipping them to identify bias and discuss AI’s impact in their communities. In the future, Changemakers plan to run more JumpstartAI sessions.
Changemakers: Maritza A (Group Leader). Shreya S (Group Leader). Raza A. Emma M B. Anusree C. Afifa F. James T. Claire S (Supplementary Support).