cs162fek@sdcc18.ucsd.EDU (William Scarvie) (11/24/88)
We are part of a group working in a Cognitive Engineering course at
UC San Diego on the design of a tool which would allow the designers
of neural networks to view their network while it is training in a
clear and easily understandable manner. The following questionairre
is designed to help us pole the user populace and determine what
features user need or would like to see in such a tool. If you
would be so kind as to take a few minutes to reply to us we would
appreciate it greatly.
Please keep in mind that we are posing these questions from a
position of somewhat limited fluency in the field. If we have made
assumptions about neural nets which are invalid please feel free to
comment on this as well. Any input will be of great help to us.
Thank you very much,
William Scarvie
Colleen Sebald
Don Woodward
1) What form does your exposure to neural networks take? (do you
use them as a tool for problem solving, do you design and implement
your own, etc.)
2) When you are designing a network to solve a problem, what
factors are typically known to you when you begin?
3) What tools or displays are available to you which help you find
and solve problems with the new network?
4) Are there tools or displays which currently don't exist that you
would like to have available for this purpose?
5) With regard to the pictorial or greaphic representation of a
network, what schema do you prefer (matrix, node-and-edge drawing,
etc.) and why?
6) What kinds of things do you need or would you like to know about
any given network in order to analyze it's performance?
7) What kinds of features of the network would you like to be able
to observe while the network is in training (i.e., a representation
of the weight space, the patterns of activation of the hidden nodes,
etc.)?
8) In many representations a gradient of color is used to represent
a scale of weights (i.e., color ranging from red to blue used to
represent a scale from -1 to 1). Do you find this kind of
representation easy to use? Are there others which you prefer?
Thank you for your cooperation. Please e-mail your responses to:
William Scarvie
cs162fek@icse4.ucsd.edu