Episode 178 – L33T Learning Siri


Just say no to enabling autonomous bots & drones with the ability to apply lethal force.  Following on last week’s episode 177, Michael & Michael start off with a continuation of the discussion on the Pandora’s Box of armed robots.

Moving on to a happier topic, while still staying on the machine learning concept, the pair talks about how Siri and other voice interactive systems have been improved with deep learning.  Not just recorded scripts, but actual intelligence, and more natural interaction is apparent when you listen to how far things have come since the earliest days of Siri — listen to the examples from the Apple Machine Learning Journal article, and we think you’ll agree.

Understanding that you’re talking with a bot vs a human is important to help avoid the uncanny valley where the interaction becomes weird vs expected — perhaps this is why Michael R preferred the Australian accent reminiscent of J.A.R.V.I.S, and how in Star Trek and other science fiction, interaction with the computer was prefaced with “computer!” before issuing a request.

Feedback loops getting faster & faster, which is really evident in machine learning examples applied to games.  When the machine — the program — is experimenting with the game to learn, and figure out what is rewarded, what is not rewarded, and determine a heuristic model for playing the game, which for a human is for fun and for the machine learning example is to maximize the outcome.  The example from the O’Reilly article was particularly intriguing to Michael M who wondered why the AI playing the game did not scoop up each and every present while playing the game.  Michael R made the interesting point that if the machine learning algorithms assume that there’s an infinite number of presents ahead, it makes more sense to quickly move forward and grab the presents because there are more ahead of you than behind.  The screenshot above is from a machine learning model playing Super Mario Kart, where it is easy to see how machine learning quickly gets better and better with successive plays of the game.

Closing out this episode, Michael and Michael include David Ma’s videos of food imagined in the style of  famous directors, a funny song about waffles and a discussion on Zork.  Hope you enjoy these tasty treats!

Selected links 

last week’s show, E177 — What could possibly go wrong? — http://gamesatwork.biz/2017/08/31/episode-177-what-could-possibly-go-wrong/

An open letter to the United Nations Convention on Certain Conventional Weapons — https://futureoflife.org/autonomous-weapons-open-letter-2017

We don’t have long to act — https://qz.com/1058280/we-do-not-have-long-to-act-teslas-tsla-elon-musk-and-others-warn-the-un-about-autonomous-weapons

Six Colors article:  Deep Learning Improves Siri’s voice — https://sixcolors.com/link/2017/08/deep-learning-improves-siris-voice/

Apple’s Machine Learning Journal:  Deep Learning for Siri’s Voice: On-device Deep Mixture Density Networks for Hybrid Unit Selection Synthesis — https://machinelearning.apple.com/2017/08/06/siri-voices.html

Uncanny Valley — https://en.wikipedia.org/wiki/Uncanny_valley

J.A.R.V.I.S. — https://en.wikipedia.org/wiki/Edwin_Jarvis#J.A.R.V.I.S.

O’Reilly article:  Bringing gaming to life with AI and deep learning — https://www.oreilly.com/ideas/bringing-gaming-to-life-with-ai-and-deep-learning

YouTube:  Super MarI/O Kart Commentary — https://www.youtube.com/watch?v=S9Y_I9vY8Qw

Food by Favorite Directors (Sort Of) — https://www.subtraction.com/2017/08/14/food-by-famous-directors-sort-of/

YouTube:  What if Alfonso Cuaron made pancakes?  — https://www.youtube.com/watch?v=6_hpJHNt4IE

Do You Like Waffles?  By Parry Gripp — https://itunes.apple.com/us/album/do-you-like-waffles/id468568946?i=468568983

MIT Technology Review — The Enduring Legacy of Zork — https://www.technologyreview.com/s/608670/the-enduring-legacy-of-zork/

MIT Technology Review article:  Siri for Business — https://www.technologyreview.com/s/600990/siri-for-business/