Oliver Selfridge is a P.IM.P. pt. 1

Pretty damn IMportant Person in the world of A.I. !

I've been meaning to share this for a few days, but it'll probably take a few posts. So I'll just start with this: On Thursday, October 25th, I had the privilege of seeing Oliver Selfridge give a presentation to the Symbolic Systems posse at Stanford and it was incredible. I didn't know what to expect, but I walked in and there was John McCarthy, Nils Nillson, and Ed Feigenbaum sitting in the audience. It was a small affair, probably around 40 people, and half of them were O.G. veterans of A.I. There were probably other greats there too that I just didn't know or catch. Half the room was gray hairs and the other half were undergrads. It was kinda like something out of Dark Crystal, all the Mystics gathered and dropped some serious science while us Skeksis waited to merge, but anyway, onto the good stuff....

His lecture was titled "Let's Improve Machine Learning" and I took notes voraciously, so I'll just share some highlights. Also, the experience has given me 3 threads of thought I find myself meditating on, that's why I'll be breakin the posts up:

Part 1) His presentation itself and what it means
Part 2) Hearing the history of A.I. shared first person by himself and the attendees
Part 3) How come so many people ride FB's jock for the valuation, or attend the web2.0 summit but sleep on the greats who actually deliver? In fact, we still use and teach what these people delivered 40+ years later while I'm pretty confident we'll forget some of these dot-com frontrepeneurs after a short time.

Part 1: The Presentation

"I want software to give a shit."

That direct quote sums it up in a nutshell. He himself quoted Norbert Weiner who proposed "a machine can learn" , and Ed Berkeley who said, "a machine, therefore, can think." From this he went on to question how much does software think today? He also asked, well, if we want software to think, how might we do that?

To answer those questions, he focused on giving software a Purpose, and what that entails. His notion of Purpose: "Having Purpose means having a goal, and trying to achieve it."

To have software give a shit, he Proposed "Purpose Loops". Adaptive, learning loops, with an emphasis that software should remember what it has learned. He pointed out that today, most apps reach a goal, but forget how they got there and don't build off of it.

He also stated that Math, while great, should not be an impediment to getting software to care. His background is Math, and he didn't mean to offend any Mathematicians, he just thinks that sometimes scientists get lost in the equations and at that point stop getting things done in regards to advancing learning applications forward.

From there he stated some of the insights and challenges with Purposes, and he often used people and children as a metaphor for what the software could/might do:

Purpose Insights:

1)You can have many at once.
2)There are different kinds.
3)You remember them, and learning is not isolated.
(you build off experiences; sequences of met goals.)
4)You collaborate: sharing what you've learned;
we learn from our experiences and other's;
we don't reach goals or learn in a big-bang fashion, it builds.
5)If you don't reach the goal the first time, you try something else.
6)Learning is self-improvement; what the learner perceives to be important.

Purpose Challenges:

1)You can have a hierarchy of Purpose:

"Win the war!" - The Commander
"Take the city!" - Major
"Take the hill!" - Corporal
"Take out enemy tanks!" - Sgt.
"Stay Alive!" - Private

2)You can have many Purposes at once.
3) The world changes: values, cost, environments, you change...

From this he pointed out, that many times we work to deliver the very best software, but given this, he quotes Minsky: "The best is the enemy of the good". So again, the message was just get out there and make something happen!

4)Sharing; you can't always share and when you can there's no standard way. Language helps, but in some situations, you need examples, instructions, teaching. (He used Tennis as an example.) He also pointed out that many times Sharers have something in common.

Now, here's where it got kind of Matrixy; he proposed apps that help other apps in the way an adult helps a child. When people asked technical questions during the course of this event, he'd often respond: "How would you (insert noun/verb from question here) with a child?" But he made complete sense. It was Brilliant!

Some Challenges with the Challenges:

He provided some more examples where Learning and Sharing are just tough:

1) Learning to raise an eyebrow; or raise each individually.
2) Sometimes you can't teach how to do something, but how to learn how to do something.
3) Counting; how did man learn to count and why, it must be important.
4) How to get to the grocery? Requires maps/models; we often make the map in our head and remember landmarks.
5) Spoonerisms: I'm riding a well-boiled icicle
6) SkyDiving: Trying and failing may not work.
7) Learning a new skill
8) Solving quadratic equations (there are many ways: factoring, graphing, formulas, etc.)

What To Do:

He started to wrap up with proposals for what we can work on to help give software a purpose.
1) A full descriptive taxonomy of Purposes.
2) A language for expressing Purposes.
3) Flexible adaptive modules for our Purpose Loops and Learning.
4) A language for remember what we've learned
5) Hierarchies of Purposes
6) Kinds of Change

A particularly useful problem: Recognition of English Text, there's no ideal way.
(Note: NLP is something I'm keenly interested in and have been focused on for quite some time, so I was right there with him loving it by the time he got to this point.)

In Conclusion:

He concluded with this parting thought:
You can't program a kid, you can only open a door for a kid and show them what's out there, then they care. Do the same for software.

Final Thoughts:

Ok, some of those points above may be concise, but I hope you got something out of it. It was a great presentation, and some of McCarthy's quotes during the presentation were priceless:

When Oliver spoke on learning, McCarthy stated with authority in response:
"You can only learn what your system will permit."

When Oliver spoke on the advancement of machine learning and A.I., McCarthy pointed out:
"Complete success has not been achieved."
which caused many to laugh.

And finally, when people were questioning Oliver's points on scientists getting lost in Math, and not advancing machine learning as quickly as may be possible,
McCarthy blurted out: "It's not the fault of Mathematics!"

So marinate on that for awhile, I'll followup with some more A.I. goodness later.