Monish Ramadoss, a computer science & engineering major with a keen interest in artificial intelligence, recalls hearing about the then-new AI@UCI Club in spring 2017. That first meeting, held in a windowless basement room of the Donald Bren School of Information & Computer Sciences, attracted a dozen or so like-minded undergraduates who wanted to discuss machine learning, computer vision and other aspects of AI.
Ramadoss kept attending the meetings for the new friendships he made as well as the stimulating subjects covered.
“The idea of the club was to create an open forum for everybody to communicate their ideas about machine learning and other different topics – an area where people could come together and develop new projects or even just connect,” says Ramadoss, now a senior and one of 15 leaders of the AI@UCI Club.
The student-run organization currently has a mailing list of around 2,000 members.
During the academic year, 100 or so students gather twice a month for free workshops in which they get hands-on experience designing such things as chatbots – think Alexa and other services that simulate conversations with humans by leveraging AI and natural language processing – and fake online dating profiles, to illustrate the concept of generative adversarial networks (see definitions below). Once a quarter, guests speak at AI@UCI Club meetings about real-world applications of artificial intelligence.
Anthony Luu, a computer science major who, along with Ramadoss, is a group leader (they call themselves mentors), is interested in an AI-related career in the medical or cybersecurity field. At one club seminar, he talked about an app he’s developing for UCI Dining Services: “UCI is focused on zero waste. I’m working on an app that allows users to take a picture of their trash, and then the app will tell them what bin to throw it in, along with instructions on how to properly dispose of it.”
Club President Iman “Amy” Elsayed, a fifth-year computer science & engineering major, works closely with the group’s academic adviser, Alex Ihler, a professor of computer science who teaches many of UCI’s undergraduate machine learning classes.
“AI encompasses so many different fields,” Elsayed says, “but at the end of the day, it’s all math. And I really enjoy math.”
Says Ihler: “The students approached me. The undergraduates wanted to have a club so they could get together and learn about different projects. A few years ago, AI clubs at universities were somewhat unusual. One reason is that machine learning in the past has not been very accessible to undergraduates. Now it’s starting to become more accessible and more relevant.”
Shivan Vipani, a junior majoring in computer science, joined the AI@UCI Club when he was a freshman. “I was interested in the workshops, and I wanted a little head start on being introduced to the world of AI,” Vipani says. “I thought it’d be a cool way to learn and get my hands dirty.”
To Andrew Laird, another club member and computer science major, AI “feels like magic.” He adds: “It’s super impressive what people have done with AI on the internet. When you get down to it, it’s all math and such, but it’s nice to be the magician.”
With AI terminology a veritable alphabet soup of head-scratchers for the uninitiated, UCI Magazine asked mentors of the AI@UCI Club to explain in their own words what some of the key phrases mean.
“‘Artificial intelligence’ is a broad term that means for a machine to demonstrate intelligence similar to humans or animals. The field encompasses several subfields such as computer vision, machine learning and natural language processing. While AI in the media often depicts a doomsday scenario, such as in ‘The Terminator,’ machines don’t have their own conscience and really only excel at what the programmer tells them to do. I like to describe programs as really smart 5-year-olds. They do exactly what you tell them to do.”
(object recognition and visual understanding)
“Face ID on iPhones is a popular application of computer vision, which allows technology to make sense of images,
such as recognizing objects. When you open your iPhone, computer vision allows the phone to ‘see’ you and unlock
the phone once it recognizes you.”
(genetic algorithms, genetic programming)
“Think of this as a computerized version of Darwinism. ‘Evolutionary computation’ refers to a machine learning method of optimization and learning inspired by genetics and evolution. Starting with a large group of possible solutions, we take the characteristics of the best-performing ones and produce a new set to eventually end up with the ‘fittest.’ It’s
much like finding the best recipe for homemade cake: You try multiple different methods and find which aspects of each attempt produce the best-tasting cake. On each subsequent attempt, you only use methods that you’ve found create the best flavor. Over time, you end up with the perfect cake, making all your friends crown you the Cake Master.”
(speech recognition and production)
“Virtual assistant technologies like Alexa, Siri or Google Assistant are great examples of speech processing AIs. When a user says ‘Hey Siri,’ the iPhone employs speech recognition to understand what the user said. Then the AI’s response is converted back to sound using speech synthesis, giving a more natural way to interact with your phone.”
Natural Language Processing
“Natural language processing is the intersection between computer science and linguistics, dealing with how computers process and analyze human language. When your phone guesses what you’ll type next or Google answers a question for you, that’s NLP hard at work to understand your sentence structure and the inherent meaning behind it. NLP is how systems like chatbots and Google Translate are able to run. As the volume of text information increases over time, NLP will play a key role in making sense of it all.”
“Currently, AI is a ‘black box’ where we don’t always understand how an AI program makes a decision or comes to a conclusion. Explainable AI tries to extract and communicate why the AI program makes its decision in a way that humans can understand.” – Amy Elsayed
(scheduling, game playing)
“Reinforcement learning is the process of learning by interacting with an environment. Games are excellent examples of reinforcement learning. In 2016, Google DeepMind’s AI, AlphaGo, beat the human world champion of Go – an ancient Japanese game significantly more complex than chess. However, RL techniques are being applied to more than just games. Researchers are using RL to control robotic systems, optimize business strategies and even predict protein folding, which helps biologists fighting diseases, including COVID-19.”
“Machine learning is a subfield of AI that uses statistical methods to make AI perform better with experience, or examples. We can ‘teach’ a machine to recognize puppies by showing it many images of puppies and non-puppies (an example of supervised learning). The more examples of puppies we feed the machine to learn on, the better it will be at recognizing a puppy in a new image.” – Amy Elsayed
“Most of the machine learning that you hear people talk about is supervised machine learning. Supervised learning uses labeled data and maps it between some observable information, or features (x) and output (y). This mapping allows the algorithm to take in something it’s never seen before and apply the mapping it learned from the data to
produce a prediction of the ‘correct’ output.”
“Data mining is pretty great, despite its bad rap in popular media. (Remember how Target got slammed for being able to deduce, based on her purchases, that a teenage shopper was pregnant before her father knew it?) Instead of mining the Earth, data mining operates on large raw datasets, and instead of pickaxes, it uses tools like machine learning and statistics. In both, the goal is to chip away the surface and uncover hidden value – in this case, patterns, trends and information. Such information can then be used for tasks ranging from providing better music recommendations to more accurately detecting cancer in X-rays.”
Generative Adversarial Networks
“Think of the classic ‘cops and robbers’ scenario. In games, AI agents can use self-play to improve their skills. Generative adversarial networks are a similar idea used for supervised learning. We build two neural networks: a ‘generator’ to make fake outputs and a ‘discriminator’ to evaluate how real they are. Together they make a ‘cops and robbers’ relationship in which the generator tries to fool the discriminator by creating more realistic outputs and the discriminator searches for ways to tell the difference. An example of this application is image upscaling, where a generator is trained to take a low-resolution image and upscale it to a higher resolution; the discriminator makes sure that the results look realistic.”
Recurrent Neural Networks
“Recurrent neural networks are robust learning models that have found their application in various complex systems. Unmanned vehicles and robots defying gravity opened avenues in research in science and technology. Similarly, businesses benefit a lot from these methodologies in helping them analyze their data and everyday activities better. RNN serves as the state-of-the-art approach for determining future states of objects and data based on objects that happened earlier. Google Translate, Google Finance and Amazon’s Alexa chatbot all use such intuitive tools at the heart of their software.”