I recently got the chance to go to EmTech Digital 2019, an event put on by the MIT Technology Review to discuss the ethical and social implications of recent advancements in artificial intelligence in a variety of industries. I’ve since returned to school, but it was a really thrilling event, so this post is my attempt to summarize it all.
I was exposed to a wide range of knowledge about this enormous and wonderful field called AI, from transportation and autonomous vehicles to hacking AI, materials, and deepfakes, so I thought I’d post this as a way of synthesizing it all in a way that could be useful for others interested in the field as well. I’ll discuss a handful of the more noteworthy incidents as well as some concerns I had about the event’s substance.
The opening address was given by Microsoft’s Harry Shum, who among other things introduced me to Microsoft’s Chinese chatbot Xiaoice and some of the decisions that go into deciding how human a chatbot should be. The first of the two days was devoted to discussions about the evolving nature of AI in fields like human interaction. How, for instance, does one decide when a chatbot should cease conversing with a particular person? after an uninterrupted hour of conversation? 2 hours? The emotional component is another thing to consider; how does the chatbot handle difficulties like declarations of love or suicidal thoughts? The last query was more generic, but it had a social undertone and was truly concerning the current condition of chatbots and virtual assistants: why are they all based on women? This particular issue—the reinforced prejudice that guarantees that women are frequently perceived as having superior skills for providing services—was one we spoke about in my Natural Language Processing class last semester.
Another intriguing topic was GANs, or Generative Adversarial Networks, which Sam Gregory, the Program Director of Witness, discussed from an ethical and social standpoint. GANs, a type of neural network architecture developed by Ian Goodfellow, pit two neural networks against one another to simulate a distribution of data.
Anyway, what I found intriguing about this subject was the fact that GANs are increasingly being used to produce false media, or DeepFakes, and that, as usual, the societal implications of these new developments have the ability to tilt the scale in a world where there is no such thing as truth. They’re a really significant development since trust (or lack thereof) has the ability to influence how we see information dissemination even more. You can see what I mean from the image above and the Buzzfeed video below.
Sergey Levine, an assistant professor at UC Berkeley, introduced the method by which robots can be taught more about the world by using the specific example of reinforcement learning in the learning process of a non-mobile robot. Reinforcement learning is another AI technology that was discussed. In order to provide goals for the robots, he combined reinforcement learning with the usage of GANs, introducing the concept of cognition in the robot learning process. In a way, the robot may “imagine” a goal based on its current condition rather than having one officially programmed. The goal would then be worked towards via reinforcement learning, until the robot was able to approach its envisioned goal as accurately as feasible.
Another topic covered in my machine learning class this semester was reinforcement learning, which essentially entails instructing an agent in an environment to execute behaviors based on the possible rewards it sees for itself when it does so. RL is often best applied in games, but I think there are some intriguing connections between ML ideas like Reinforcement Learning and real life, and seeing applications like those produced by Sergey and his research team makes me wonder what more there may be.
Professor Fei Fei Li spoke on developing human-centered artificial intelligence (AI) after the session on real-time learning (RL), continuing a theme that was present throughout the program and also explored developing inclusive AI for everyone. This was supported further by speakers like Riedet Abede, who discussed the significance of eliminating bias from our data, enhancing diversity, and include the underrepresented in AI.
Being from Africa, I believe the concept of inclusion really appealed to me. Most technological advancements I’ve seen thus far have been implemented in settings that don’t truly reflect my reality, and the data has been obtained from and patterned after individuals who don’t look like me. Amazon’s face recognition software, Amazon Rekognition, which not only fails to reliably identify minorities but is also offered to law enforcement, is one frequently cited example of how poorly AI can manage racial variation.
A presentation by Dr. Solomon Assefa of IBM Research’s African division helped to dispel my concerns about AI in Africa by pointing out that there are, in fact, some emerging uses of AI on the continent, though information about these uses isn’t as widely disseminated as it ought to be. He discussed other initiatives utilizing AI in Africa, such as Hello Tractor’s usage of blockchain to rent out tractors.
One of the co-founders of Embrace Innovations, Rahul Panicker, gave a discussion that stood out during the event about the use of AI to approximate the weights of underweight/premature newborns using 3D technology. By using individual health workers with the assistance of a mobile application (on which the ML models are deployed), the correct estimates of these children can be found without placing them on a balance, and the appropriate interventions are recommended per weight range. This application was in the Indian context, and the issue it was solving made me realize the extent to which culture could be a force working against technology.
Really fascinating material. Dr. Dawn Song deserves special notice for her remarks on the need for more secure AI frameworks. She asked, “What if I could reverse engineer some insights and find the underlying data on which it was generated?”
