Nadja Rhodes is enamoured with artificial intelligence. A Seattle-based Microsoft software developer self-taught on AI techniques such as deep learning, Rhodes had applied to a number of tech company sponsored AI residency initiatives, but to no avail. And so she was thrilled to be accepted by OpenAI Scholars.
Launched in March, OpenAI Scholars is a diversity-oriented initiative that provides mentorships and stipends to enable individuals new to machine learning to study full-time for three months and complete an open-source project. Last week OpenAI held a Scholars Demo Day at its San Francisco office.
Rhodes’ project is an automated music commentary generator — a bot that can write music reviews after receiving musical attributes of a Spotify song. “People always have a super descriptive and creative way of writing music commentary. Plus, I always want to do something unique,” says Rhodes.
Six of the eight scholars are female and the group represents a wide range of ethnicities and backgrounds, from a Navajo painter to a Russian former ballerina. There are no white males. The diversity push underlines OpenAI’s commitment to promoting underrepresented groups in machine learning and the institute’s long-term mission to advance digital intelligence in ways likely to benefit humanity as a whole.
Public concern has emerged regarding bias in machine learning, particularly in tasks related to race. A study by MIT and Microsoft researchers released earlier this year found that facial recognition tech from Microsoft, IBM, and Chinese AI unicorn Megavii have an average accuracy rate of only 65–80 percent when recognizing dark-skinned female subjects — in sharp contrast to their 99.6 percent accuracy when identifying light-skinned male subjects.
“We are in a diversity crisis for AI,” NIPS 2017 Black in AI workshop organizer Timnit Gebru told MIT Tech Review earlier this year. At this spring’s Google Developer Conference I/O, Stanford University Professor Fei-Fei Li said “I believe in the future of AI changing the world. The question is: who is changing AI? It is really important to bring diverse groups of students and future leaders into the development of AI.”
Last year Dr. Li launched AI4ALL, a non-profit organization that aims to nurture the next generation of AI technologists with representation from a broad range of ethnic backgrounds.
OpenAI Project Manager and Scholar Program lead Larissa Schiavo suggests that people of color are more likely to become engineers than spend five years in a PhD program. “Traditionally, machine learning has been dominated by people who come from a more traditional academic background.” The OpenAI scholar program aims to fast-track people of colour’s entry into machine learning.
Each scholar is assigned an OpenAI researcher as a mentor. They work about 40 hours a week, first acquiring basic machine learning skills, then developing a research project; and write weekly blog posts updating the community on their progress.
Rhodes says she’s learned a lot. For her final project, she collected a set of training data comprising some 20,000 music reviews, which were split into 104,500 sentences. The model is built on a sequence-to-sequence conditional variational autoencoder (seq2seq CVAE) model with an additional latent constraint generative adversarial network (LC-GAN) model to generate embeddings of text conditioned on a thought vector z and attributes of the referenced music a. Rhodes’ trained model “deephypebot” regularly tweets “weird but interesting comments” on popular Hype Machine songs.
Rhodes is on her way to becoming a machine learning engineer.
Founded in 2015 by Elon Musk and Sam Altman, OpenAI is a non-profit AI research institute. Open AI plans to repeat its Scholars program in 2019. Click here to read more on this year’s final projects.
Journalist: Tony Peng | Editor: Michael Sarazen