Bruce Blumberg is Assistant Professor in the MIT Program of Media Arts and Sciences, Head of the Synthetic Characters Group and Asahi Broadcasting Corporation Career Development Professor of Media Arts and Sciences. Using animals such as dogs as his model, his research goal is to build animated characters that possess the same type of every day common-sense, empathy, and the ability to learn that one finds in animals such as dogs. His group has had a number of prominent installations such as SWAMPED!, void*: A Cast of Characters, sheep|dog: Trial By Eire, and this summer Alpha Wolf, which will be part of the Emerging Technology venue of Siggraph 2001 in Los Angeles. Prior to the Media Lab, he held positions with Apple computer and with NeXT Inc., where he was the first employee hired by the founders. Blumberg holds a B.A. in economics from Amherst (1977), an S.M. in management (1981) from MIT Sloan School of Management and a Ph.D. in Media Arts & Sciences (1996), from the MIT Media Laboratory.

Blumberg's response:
I got the idea in 1994 to focus on dogs. It was a combination of a suggestion of my professor at the time to pick dogs as a good animal to model and my love of dogs which I have had since I can remember.

The very first dog that we did would only run on a $1 Million dollar machine. Duncan runs on $1000 PCs today. The cost of building Duncan really is in the people who worked on the system. Duncan was the collective work of several graduate students, artists and myself.

Yes, one of my students has made a virtual wolf pack and I have always wanted to make a virtual beaver pond. But we will continue to focus on dogs for the foreseeable future.

Blumberg's response:
I hope that one day you will be able to play with Duncan or one of his descendants. One of my dreams, in fact, is to one day create virtual worlds inhabited by virtual animals like lions, chimps or wolves that kids could interact with in ways that they might not be able to in the real world. Wouldn't it be cool if you could become part of a lion pride for a day and see what life might be like as a lion cub?

In the meantime, I hope that one day you can get a real dog.

Blumberg's response:
Dogs are an ideal model for studying what I call "everyday commonsense", that is the commonsense, social savvy, and ability to learn that animals such as dogs possess and which gets them through their day. If you think about it, dogs do a remarkable job of fitting into human society and do so in a way that is very rewarding to us humans. It would be great if the technological devices that we interact with every day were as capable in this regard as a dog. This is not to say that the ideal computer interface is a dog, but rather that if we could imbue the interface with some of the commonsense, adaptability and social savvy found in creatures such as dogs we might be able to create a more engaging and inherently more rewarding interaction.

Perhaps you saw pictures of the search and rescue dogs and their handlers working at the WTC site. The dogs were performing 2 functions in a sense. On the one hand, their job was to locate potential survivors, and they could do this better than any machine or person for that matter . On the other hand, an equally important part of their job was to be there to provide emotional support to their handlers and the other rescue personnel. That is, to remind the workers what was good about life in the midst of the horror.

In developing technology we should keep search and rescue dogs and their dual functions as a model of what to strive for.

As for your final question, ultimately, I would like to do much of the same kind of training that I do with him when I am at home. As we proceed with this research I would also like to find ways that we can enrich the lives of dogs who are at home while their human companions are away. It would be great if you could do a 5-10 play period every couple of hours.

10.24.01 Glenn Murdoch asks:
Hello Bruce,

If it's not too much trouble, I would like to know what data structures you use to store Duncan of Innisfree's knowledge.

I too am researching the nature of intelligence using software, but I am using a "bottom up" approach rather than your "top down" one. I have created a virtual world of very primitive creatures and am trying to get them to evolve their own intelligence via natural selection. Their intelligence has not yet reached that of insects, but the storage space is already quite heavy since the world has a capacity for over five thousand creatures. My current software model is already showing limitations, particularly since creatures can only inherit knowledge and cannot learn new ideas based on interaction with their world.

I would be very interested in how you get Duncan to identify the important aspects of any learning experience and how you store this knowledge without requiring gigabytes of disc space. Duncan seems quite intelligent so I would like to know how he decides what course of action is most appropriate under the current circumstances.

Blumberg's response:
The best thing to do would be to look at our web page for papers. Here is a link:

http://www.media.mit.edu/characters/resources.html

One of the cornerstones of our work has been to look for "scaffolding", that is those structural elements which if incorporated into the system make it far easier for the creature to respond intelligently or learn what it ought to learn. For example, Duncan has an explicit representation of time and rate, and this makes the task of learning far easier. While we do not take an evolutionary approach, the idea is still very applicable. We also make assumptions about how the world works and this combined with a hierarchical search strategy reduces the amount of memory required. Plus as you will see from reading our papers, there is a lot of built in structure in other ways as well. Anyway, take a look at our papers. It sounds like you are doing interesting work.