Interview: Andrew Olney

Last week, I wrote a piece for Motherboard about an android version of the science fiction writer Philip K. Dick. The story of the android is truly surreal, stranger than even Dick's flipped-out fiction, and I recommend you pop over to Motherboard and mainline it for yourselves. For the piece, I interviewed the lead programmer on the first version of the PKD Android, Dr. Andrew Olney. Aside from bringing science fiction legends back from the dead, Olney is an Assistant Professor in the Department of Psychology at the University of Memphis and Associate Director of the same university's Institute for Intelligent Systems. He makes intelligent tutoring systems and once hacked a Big Mouth Billy Bass.

I found his interview too interesting not to share in full.

Universe: This project lives in a world between computer science, sculpture, and literature. In terms of the response you've received, has it been understood primarily as an art piece or a feat of engineering?

Olney: I was attracted to the aspects of the project most opposite from what I normally do. I loved the art aspect, the resonance with PKD's written work. To me it was a kinetic interactive sculpture.

Universe: What makes this project fascinating is that PKD was obsessed, himself, with ideas of personhood in an age of androids. I can't help but think of Jack Bohlen, in Martian Time-Slip, servicing the simulacra in his son's school and having schizoid episodes where he believes that every person is secretly a mechanism. This is presented as being patently horrifying; based on your understanding of his work, do you think Dick would have liked the android or been terrified by it?

Olney: I can only speculate, of course. I imagine he would have appreciated the reification on a certain level, say on the nuts and bolts level. To the extent that it fell short of the androids he wrote about, I think he would have been relieved. After all it is not the basic idea of an android that is the most troubling, but rather the idea that an android can be more real, or more human, than a human. Since we no doubt missed the mark there, I doubt the robot itself would have terrified him. But as evidence that a future he wrote about was perhaps more possible than he imagined. Maybe that would make him a bit uneasy.

Universe: Can you tell me a little about the AI?

Olney: The basic AI is somewhat analogue to my reading of Dixie Flatline from Neuromancer. Imagine a giant recorder that pairs every stimulus you receive with every response you make. I had a rough approximation of that via PKD's interviews, letters, etc. With the rest of his writings (books, fiction, etc) I had enough material to make a giant semantic space, basically a vector space that maps words to vector representations. This let me create a giant associative memory of conversational stimulus/response pairs, mediated by the meaning of what was being said to the robot. That was most of the interesting stuff going on with the AI. In addition, I hooked up a chatbot for the goofy dialogue ("What is your name") and used a ranking procedure to decide what response to select. This is a bit of a simplification, but there were about 20 responses generated for every spoken input, and only one was selected. The rest of the robot was sensors (face recognition) and effectors (facial expression, lip sync, audio output) chained together using an expert system in order to drive certain behaviors, like turning the head to make eye contact. I also experimented right before the head was lost on dynamically generating emotional expressions based on the emotional valence of the robot's speech, which actually worked pretty well. We've since used that in other systems in my lab.

Universe: What is your end goal, in developing an AI like this? Are you interested in developing real machine intelligence or is it more about making useful systems to assist people in everyday tasks, like learning?

Olney: That's a great question. My position is that real-world contexts are an endless source of important research questions. AI is way past the point where toy systems can generate interesting questions. It's only by placing your system in the world that you can see where the holes in your theory are. So it's a very symbiotic relationship between science and engineering. On one level, you are establishing the computational specifications for a functional (though not necessarily human) intelligence. What you learn from AI can be taken back to understanding human cognition or animal cognition because it helps shape your questions-it helps you ask better questions about human and animal intelligence. On another level, you are creating intelligent machines that carry out useful tasks in the real world. Something that we do a lot of in my lab is create intelligent tutoring systems (ITS) that tutor students on various topics. This is societally relevant because there are not enough human tutors to meet the need and tutoring is expensive. So creating ITS as a supplement to classroom instruction has a positive social impact and it is a source of important research questions about natural language understanding, natural language generation, motivation, and learning.

Universe: What's your favorite Philip K. Dick novel or story?

Olney: Scanner Darkly is by far my favorite out of what I've read, which is only half a dozen or so. I've read maybe 20 or so short stories, but I suppose my favorite is the Golden Man, which I love for its non-cognitive-evolutionary storyline. Another I like is the Pre-Persons, not because I agree with the position on abortion, but because the story is hilarious.