This Week in AI, March 1st, 2018
Simulated Robotics Environments from OpenAI, Facebook’s Text-to-Speech model, a new Technical Director for AI for Google Cloud, and more!
OpenAI Releases Simulated Robotics Environments
OpenAI is well known for their Gym environments commonly used for reinforcement learning research. This week, they released a set of new environments focusing on robotics. One environment has a robotic hand that you can train to grasp and manipulate objects. The other has a robotic arm that can push, pull, and pick up objects.
Along with that announcement, OpenAI also released a paper on a new reinforcement learning method called Hindsight Experience Replay. This enables agents to learn from mistakes by pretending final incorrect actions were right all along. Learning from mistakes is a powerful tool for humans, this could prove to be crucial for reinforcement learning.
Facebook Enters the Text-To-Speech Field
Along with Google and Baidu, Facebook has been hard at work on a neural text-to-speech model. Previously, these types of models required a lot of data to learn to generate new voices. The Facebook team built a model that can generalize to new voices using only a small amount of untranscribed audio. This means it can learn from “in-the-wild” recordings instead of needing controlled environments.
One Pixel Changes Everything
A team at Kyushu University in Japan found a way to fool computer vision neural networks with a single pixel. By changing the color of a single pixel in an image, they were able to force a network to misclassify images it would otherwise predict perfectly.
This is known as an adversarial attack. It’s been known for a while now that you can add imperceptible noise to an image and get pre-trained convolutional networks to predict whatever you want. The thing is, the images themselves look completely unchanged to humans. This can even be done with simple stickers which can effect computer vision systems in the real world (think automated cars).
Similar to the one pixel attack, researchers at Google Brain found a method to fool computers and humans.
To Structure or not to Structure
Yann LeCunn and Christopher Manning (famed AI professors at Facebook/NYU and Stanford respectively) had an interesting conversation about imposing structure on neural networks. Do we build in connections as we do in convolutional networks and capsule networks, or do we let networks discover appropriate configurations themselves? Human brains are highly structured, so perhaps we should attempt to emulate them? On the other hand, at some point in the distant past, the animal brain was unstructured, and over a billion years of evolution, it developed structure.
From Alexa to Google: Musical Chairs in the AI world
Ashwin Ram, the lead for AI on Amazon Alexa, has joined Google as the technical director of AI for Google Cloud. Amazon, Microsoft, and Google are in fierce competition to capture the cloud market specifically with respect to AI. These companies, and plenty more with deep pockets, are fighting over AI talent, sometimes paying a million dollars for experienced practitioners.
~
Stay tuned for new updates as we continue to review all that’s new in the world of AI! And if you’re interested in mastering these transformational skills, and building a rewarding career in this amazing space, consider one of our Nanodegree programs:
