On Monday, Oct. 29, Vassar students converged in Rockefeller Hall 300 to attend a presentation on artificial intelligence (AI) in health. Organized by Xiaoqing Xu ’19 and Vanessa Achoy ’19, alongside University of Oregon Professor of Computer and Information Science Dejing Dou, the lecture informed attendees about the current use and potentials of AI in healthcare.
To start off the event, Xu briefly introduced listeners to the AI Challenger Global contest, a non-commercial, open-source platform of datasets and programming competitions for designers of artificial intelligence. Hoping to encourage AI talents to bring valuable contributions to the world through research and practical application, this international competition provides rich, high-quality data resources. In 2017, AI Challenger housed 8,892 teams from 65 countries, marking it as the most widely scaled Chinese dataset platform and non-commercial competition.
Inspired by this contest, Xu founded Vassar AI Challenger, hoping to provide a similar experience and opportunity for campus AI enthusiasts. Xu explained, “This summer I went to an AI bootcamp organized by Sinovation Ventures, a venture capital firm that also serves as an incubator for startups using AI technology. Since last year, they started a non-commercial platform for data science contests and open-source datasets, called AI Challenger. As part of their promotion for this platform, they are encouraging lecture events like this across the globe to help create opportunities for students to get in touch with the advances in the field of AI.”
Xu is now starting a small interest group, organizing weekly workshops on Saturdays in Sanders Physics or New England in which participants test out AI applications and discuss AI-related topics, ranging from algorithms to applications to ethics.
After the introduction, Dou gave a live presentation demonstrating AI’s evolution and potential. As he explained, “It kind of went back to 1940s, with McCulloch and Pitts’ Boolean circuit model of [the] brain, and since the 2000s, data mining has been a big part of AI. And in 2010, deep learning dominated AI.”
In particular, data mining is characterized by its information extraction technology. The information from data mining can be used for various tasks, and the technology uses ontology, the study of concepts in a specific domain that show the relationship between them.
Explaining the impact of semantic data mining with respect to healthcare, Dou noted, “This is important. We have to be careful with each patient, because of the different medical records, specific vocabulary domains, issues surrounding policies and then drugs. [A medical record] becomes zeros and ones. This technology will…[help] to annotate the terminologies [to better explain each patient’s particular case]. With typical data mining, we can do this.”
Dou next engaged attendees in a discussion on deep learning, a powerful, multilayered neural network of non-linear information processing for learning representation based on data. This data-analytic technology, according to Dou, can process a vast amount of information, ranging from image classification to face detection to medical diagnoses generation.
In the field of healthcare particularly, deep learning has allowed for semantic mining in activity, social and health data. Drawing on examples such as the alarming obesity rate in the United States, Dou further championed an unlimited access to information.
In an ensuing question-and-answer period, one student inquired about evaluating the use of ontology in constructing intelligent models. Dou responded, “For this special project…what we’re trying to find is the ontology that focuses on the specific dimensions serving our project.”
Xu added, “In the health industry, there are a lot of mature data structures. Now, you have a system that you can collect data from, via hospitals, or sometimes [it is] building new things. This technology is not only analyzing data, but are also building data-collecting structures.”
Continuing the discussion with a presentation on her summer research in the Murray Lab at the Yale School of Medicine, Achoy, a neuroscience major, applied computational modeling techniques to gain insights into the nature of psychiatric disorders. The algorithms, along with three datasets, helped her research team visualize differences between the control and schizophrenic groups, allowing them to generate meaningful samples.
Achoy then carefully described the operation of three encoder techniques: vanilla autoencoders, variational autoencoders (VAE) and conditional variational autoencoders (CVAE). She explained, “[The vanilla autoencoder is fed] an image into this neural network, [which travels] through the network, and then [it] output[s] an image as close to the original as possible.”
The VAEs, which Achoy employed during her summer research at Yale, enabled her to create and store blueprints of input data in the system and then produce real and genuine data.
Discussing existing shortcomings in representing data using latent space, Achoy noted that the storage systems’ two-dimensionality and data distribution can hinder accurate data representation. In the future, Achoy plans to test the same set of data on CVAE, which can understand the numbers it receives contrary to VAE’s capabilities, and further investigate the impact of two-dimensionality on representing data.
When asked about the impact data amplification may bear on efficacy in healthcare, Achoy explained in an emailed statement: “Creating new data for research will help so much! It’s hard to locate and have a statistically reliable number of human subjects in trials. If we can artificially create data, it would improve the statistical power of healthcare research as well as lessen financial and time constraints.”
Yet, she remained realistic: “Unfortunately, there is again the issue of making sure that any artificially created data is unique, reliable and true enough to be trusted as part of a dataset.”
All in all, the lecture served to educate students about how AI-related technologies could impact the health industry, raised possibilities for future research and encouraged students interested in AI to explore real-world applications.
