RSA Wrapup: Song, Darling, Thrun

The closing session of the RSA conference was a sort of chat-show hosted by Hugh Thompson, who is the RSA Program Chair. He brought up three people, two of who have featured in Breakfast Bytes posts before.

The proceedings started with a session where DJ Shiftee, who had been spinning for us while we waited for the show to start, was challenged to pass the Turing test more convincingly than an anonymous Alexa/Cortana type voice assistant. He succeeded.

Then Hugh brought up each of the three people individually.

Dawn Song of UC Berkeley

First up was Dawn Song, who is a professor in the EE/CS department at UC Berkeley. Adversarial machine learning is where one AI system is trained against another, such as training driving algorithms against models of roads. "But what happens when you have an active adversary who is tampering with the training set?" Hugh asked in his introduction. "But luckily we have one of the world's leading experts."

She is an expert on changing images slightly so that humans cannot tell the difference but it can fool the computer vision system. If you want more background on this, then see my post Fooling Neural Networks. The stop signs in that post are part of her work. What I hadn't realized when I wrote that earlier post is that the stop sign with a few bits of tape on it are not a static image—they fool a vision system as the car approaches the stop sign and continue to fool it as they get close.

I tried to find the video she showed, since I assume it is on YouTube somewhere, but Dr Song's name, Dawn Song, is also the name of several songs, and they seem to overwhelm the search algorithm.

What she and her team had done was developed a mathematical method to figure out what perturbation should be injected to make the attack successful. It is complicated in a physical setting because the vision distance and angle are changing and the attack should be successful through the transition.

As Dawn left, Hugh exhorted us all toe "Read her papers…they will really terrify you."

Kate Darling of MIT Media Lab

Kate has the wonderful title of Mistress of Machines. Human-Robot Interaction, Robot Ethics, IP Theory and Policy. She showed up with an animatronic baby dinosaur. She says a baby dinosaur is great since it is easier for a person to suspend their disbelief when interacting, compared to a cat or dog, where we continually compare it to a real one. She pointed out that the dinosaur was making little cooing sounds.

She pointed out:

We are entering an area of computer robot interaction. We’re primed by science fiction and culture to want to treat them as agents, and assign intent. People are suckers for these sorts of designs, when people interact and engage with the technology.

Her work is on how to get the design right. "Clippy was super-annoying." However, she thinks we need to think about representations of AI systems in ways that are somewhat human but look a bit different.

My pet peeve we are constantly comparing AI to humans. That’s not where the true promise lies. People don’t like it since their performance is not up to snuff. Compare it to domestication of animals and think of AI that way, where they enhance our skills rather than replacing humans.

She mused about whether systems have rights and whether there will be ethical issues. "Is it always OK to take the batteries out?" Kate believes that empathy will get better the more the robots improve. If you pick up the baby dinosaur, it cries and looks uncomfortable. She said that they did a workshop where people tortured and killed dinosaur robots and it was very upsetting, even though everyone knew it was just a machine.

Sebastian Thrun of Udacity

Finally, it was Sebastian Thrum, billed as the founder of Udacity. But he is also famous for leading the team that won the second DARPA challenge driving through the Mojave desert. For more on that, see my post Ten Years Ago Self-Driving Cars Couldn't Go Ten Miles. Or for more of his interests, see my 2014 post Sebastian Thrun: Self-driving cars, MOOCs, Google Glass and More. 

“You have had a pretty amazing career," Hugh opened with. "It’s like taking a normal career and running a compression algorithm on it.”

His first question was "How will widespread use of AI change society?"

Sebastian replied:

I’m incredibly excited. Imagine 300 years ago that the steam engine was invented. You would have worked in the field as a farmer. Your worth was tied up in your strength, so a person could supply food for 400 people, instead of 4. So we moved into offices and do the same thing over and over again. Not just taxi drivers, but lawyers, medical doctors. Stanford train on skin cancers. They get $450,000 a year but our program was as good.

Hugh was worried for all those professions. It sounds as if this is going to free up humanity from the mundane things to free us up to do something else. "So what are those something elses?”

Sebastian thinks that we can unleash creativity. When all we did was farming, childbirth and warfare, we weren't creative. Then we moved to reading and science and released enormous creativity. "I believe every human is creative, just get a bottle of wine and thin of crazy ideas."

When Sebastian moved on to what is important, he had an interesting perspective:

My teacher says kids need to be good at math, and I say ‘no you can hire someone who’s good at math’. Curiosity. This sets Silicon Valley apart. Grit is really important. Don’t give up. Then I’d say people skills. Social and emotional skill. When Udacity works with employers, it is important to work as a team not as individuals. I love innovation and thinking here. I know a lady in Saudi Arabia that suddenly got a job programming in Australia from Udacity courses.

It's a Wrap

As Sebastian left the stage, Hugh wrapped up the RSA conference and brought us back to security.

I want people to think about how things can fail. I think with AI we can unlock the time we can spend with each other. This is a very human topic, where we can be the guardians of the future.

The RSA Conference next year (in San Francisco) is March 4-8th 2019. There are other RSA conferences in other locations too.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.