Progress in Artificial Intelligence Requires Diversity

Source: Deep Learning on Medium

By Mahir Jethanandani

Recall some of the most infuriating problems you have dealt with: massive programs that inexplicably failed at test time, furniture pieces that could not be assembled, or cooking recipes that turned out disastrously. The most infuriating and complex problems required diverse and sometimes strange solutions: printed schematics galore, dives between the couch cushions for nuts and bolts, or adding rice vinegar to a recipe. This is the process of diverse problem-solving: devising interdisciplinary solutions from all sorts of backgrounds through cycles of hypothesizing, testing, and evaluating. So, in a field like artificial intelligence which is meant to simulate human problem solving, shouldn’t having an intellectually diverse workforce capable of diverse problem-solving be the standard?

Recent Turing Award co-winner Geoffrey Hinton claims “I have always been convinced that the only way to get artificial intelligence to work is to do the computation in a way similar to the human brain.” Intellectual diversity is achieved not only by mixing minds from an amalgam of academic interests but from every walk of life. Racial, gender, sexual orientation, nationality, and educational diversity is just the beginning of how each one of us thinks differently from one another. Many fundamental skills in artificial intelligence hint at the importance of novel problem solving, including feature engineering.

In the process of feature engineering, or selecting data features for a problem, different interpretations of data are necessary. Here is an example: The classic classification problem of spam/not spam emails challenged computer scientists to reverse engineer how spam emails were written, and how they slipped past complex spam filters. A young computer scientist well-versed in slang can fully understand its meaning perfectly, aiding the team in comprehending the lexical structure of spam emails. Likewise, when devising neural network architectures, an engineer with a background in neuroscience can best piece together network layers that simulate how humans would examine an image. Unique approaches can unlock logical parallels to problems so crucial in developing and advancing fundamental skills in artificial intelligence.

I stepped into the world of artificial intelligence through my university’s flagship artificial intelligence course, CS188, at the University of California, Berkeley, taught by Anca Dragan and Sergey Levine. The course is a survey of the field of AI, touching on path-finding problems to neural networks. I wrestled through six projects that semester and along the way strengthened my knack for tackling classical machine learning problems. Being a survey course, CS188 also expanded the number of subfields and focuses I could specialize in, leading me to select deep learning.

The full immersion of my time towards the subject illuminated the reigning experts in the field, including Professor Pieter Abbeel. His courses, including Deep Unsupervised Learning and his iteration of CS188 received glowing recommendations, made him a well-recognized name in Sutardja Dai Hall, the on-campus center of innovation and entrepreneurship. I shadowed his Deep Reinforcement Learning Bootcamp in August 2017, and remained alert for any of his future Bootcamp offerings. When the Full Stack Deep Learning Bootcamp was announced, I quickly signed up and eagerly awaited admittance. Needless to say, I was floored when the acceptance came in — with a scholarship from Turnitin!

Artificial intelligence needs to encourage mental diversity and breadth in developing applications. Crucial steps in the development pipeline — like model selection, feature engineering, and adversary protection — require clever maneuvers to best explain how a model should function. “Street smarts” — we might call the opposite of “book smarts” — enables teams to discover clever solutions to the spam/not spam problem. In the same way, diversity in thinking leads to a proliferation of neocognitron: the neural network model for pattern recognition, inspired by the human visual cortex.

Imagine how two project groups could predict housing prices given information about the estate and its surrounding areas. Quantitative minds could hone in on the physical dimensions of the estate, while unorthodox individuals might classify the style of the house (Victorian, cottage, colonial), a classification more likely to be thought of by an individual with an architecture minor in their studies. Sadly, there is currently little room to flex these creative muscles when interviewing for artificial intelligence job positions, which seems shortsighted and likely to limit attracting diverse thought in tackling these technical problems.

To bring diversity in thought and of individuals, we need a cycle of development, growth, and support for underrepresented individuals. Role models need to return to their roots and encourage those even slightly interested individuals to develop a fundamental understanding of artificial intelligence. Many such role models have begun this cycle: Daphne Koller, Suchi Saria, and Anna Penn developed PhysiScore, which predicted premature babies’ likelihood of health defects based on genetic history and postnatal behavior for underprivileged environments. Anca Dragan and AI4All run a free artificial intelligence camp for high-potential high school students in the Bay Area. These role models, and many others are a blueprint for developing young talent, by helping them navigate through higher education and into employment.

There are ever-widening cracks to fall through at each difficult concept and systemic failure in public education. We need to reassure and uplift underrepresented individuals so they will not lose heart in pursuing an education in artificial intelligence. Teaching the prerequisites to core artificial intelligence concepts can begin in elementary school, where the smallest decision to ditch school can snowball into chronic absenteeism. Mentorship is the single most important factor in creating diversity — the physical presence of successful role models is the most motivating factor in fostering intellectual diversity and breath.

Return home and teach a mini-lesson on classifying sneakers as authentic or fake using computer vision. Refer promising students to potential scholarship programs and internships, and build diversity-focused groups at your place of work. Develop, grow, and advise our next generation’s diversity so the artificial intelligence field is a level one.

Source: Mahir Jethanandani

About the Author
Mahir Jethanandani was a recipient of the 2019 Turnitin Full Stack Deep Learning Scholarship.The program brought promising young programmers of diverse backgrounds to the Full Stack Deep Learning Bootcamp. Attendees explored new ways to bring artificial intelligence applications to fruition, rubbed elbows with leading researchers and programmers, and helped envision the future of deep learning.