SITE INDEX l General Info l Schedule l Abstracts l Call For Papers l
Please email oblio@us.ibm.com with participation requests.
Humans need to trust that intelligent systems are behaving correctly, and one way to achieve such trust is to enable people to understand the inputs, outputs, and algorithms used as well as any new knowledge acquired through learning. As the use of machine learning increases in critical operations it is being applied increasingly in domains where the learning system's inputs and outputs must be understood, or even modified, by human operators.
For instance, e-mail classification systems may need to gain the user's trust by explaining their predictions in a language the user can understand. Intelligent office assistants learn from a user's preferences and behavior, but in order to be useful, the user must trust that agent will make the same decisions the human would under the same conditions. Machine learning has also been widely used to support credit approval decisions; yet banks are becoming increasingly responsible for explaining the reasons behind a denial of credit. Autonomic systems are beginning to employ machine learning to support common administrative policies; yet system administrators are reluctant to trust automated technology they do not understand.
In this workshop we explore issues of human comprehensibility as it relates to machine learning.
Topics include the following:
·
Human-assisted learning
·
Knowledge acquisition for learning
·
Establishing and maintaining trust of users
·
Human understanding or modification of learning algorithms
·
Comprehensibility of the input or bias for learning
·
Comprehensibility of the induced model
·
Learning and explanation generation
·
Exploration and exploitation trade offs in the context of human
use
July 9th
|
Time |
Speaker |
Title |
|
|
Opening Remarks |
|
|
|
Mike Pazzani |
Comprehensible Knowledge Discovery in Databases |
|
|
Pat Langley |
Computational Discovery of Communicable Scientific Models |
|
|
Pedro Domingos |
Markov Logic: A Language For Human-Comprehensible Machine
Learning |
|
|
BREAK |
|
|
|
Rob Holte |
Two Fielded Machine Learning Applications Requiring Comprehensibility |
|
|
Guided Discussion |
|
|
|
LUNCH |
|
|
|
Noboru Matsuda |
Applying Programming by Demonstration in an Intelligent Authoring Tool for Cognitive Tutors |
|
|
Andrea Thomaz |
Real-Time Interactive Reinforcement Learning for Robots |
|
|
Rich Caruana |
Machine Learning and Medicine: When Models Collide |
|
|
Simone C Stumpf |
Predicting user tasks: I know what you’re doing! |
|
|
Flavian Vasile |
TRIPPER: Rule learning using taxonomies |
|
|
John D Burge |
Comprehensibility of Generative vs. Class Discriminative Dynamic Bayesian Multinets |
|
|
Closing |
|
Learning human-comprehensible models requires easy incorporation of user knowledge into them, easy explanation of model output, and a division of labor between human and machine that makes the best of each one's strengths. Markov logic is a language that meets these desiderata via a simple combination of logic and probability: users provide knowledge in the form of first-order sentences, the sentences are refined and weights for them are learned from data, and the result is interpreted as a set of templates for features of Markov networks, with explanations provided by inference over these. In this talk I will review Markov logic, the algorithms for learning and inference we have developed for it, some of its applications to date, and some of the challenges that remain. (Joint work with Stanley Kok, Matt Richardson and Parag Singla.)
Computing Science Department,
This talk describes two fielded applications of machine learning, Proteome Analyst and SAGA-ML, for which comprehensibility is an essential requirement.
Proteome Analyst is a web-based bioinformatics tool used by hundreds of molecular biologists worldwide to analyze new proteins. Several of its components are classifiers built by machine learning. The biologists require explanations of Proteome Analyst's classifications for two reasons: (1) to help gain the trust of the biologists in the tool, and (2) neither the classifiers nor the training data from which they are built are error-free and it is important to be able to trace classifications back to their source in order to uncover errors.
SAGA-ML is a system for the semi-automated analysis of the behaviour of game software. It has been applied to the 1999 version of Electronic Arts's FIFA soccer title and is now ready for use in the development of future versions. The purpose of machine learning in SAGA-ML is to summarize large logs of the game's behaviour so that the game designer can easily identify anomalous behaviours. This is only possible if the output of the machine learning is presented to the designer in an intuitive visual form.
For more information:
Computational Learning Laboratory; Center for the Study of Language and Information
Most research on computational knowledge discovery has focused on descriptive models that only summarize data and utilized formalisms developed in AI or statistics. In contrast, scientists typically aim to develop explanatory models that make contact with background knowledge and use established scientific notations. In this talk, I present an approach to computational discovery that encodes scientific models as sets of processes that incorporate differential equations, simulates these models' behavior over time, incorporates background knowledge to constrain model construction, and induces the models from time-series data. I illustrate this framework on data and models from a number of scientific domains. In closing, I describe our progress toward an interactive software environment for the construction, evaluation, and revision of such communicable scientific models. This talk describes joint work with Nima Asgharbeygi, Will Bridewell, Andrew Pohorille, Jeff Shrager, and Ljupco Todorovski.
National Science Foundation (NSF)
Knowledge discovery in databases is a field whose goal is to turn data into knowledge. For example, by analyzing a database of credit card customers we can determine what types of customers are most likely to be profitable for the company. By "mining" databases of medical records, new cost-effective procedures for screening for diseases may be uncovered. I review advances in the field over the past two decades of research in statistics, neural networks and artificial intelligence that have identified a variety of approaches that produce accurate descriptive or predictive models. However, I show that experts are unwilling to accept the results of these techniques when they don't make sense, are difficult to understand, or violate prior understanding. I discuss factors that make learned knowledge acceptable to experts and discuss modifications to rule learning, linear regression and text classification algorithms that make the learned models more comprehensible.
http://cs.cornell.edu/~caruana
First I'll present a case-study where the best model on an
important
problem in pneumonia risk
prediction turns out to be an opaque neural
net containing 100,000+
weights. This model has *much* better
AUC than
the standard logistic regression
model, but the doctors went to clinical
trials with logistic regression
anyway. I'll explain why.
Then I'll describe a method called Case-Based Explanation we
devised to
generate explanations for opaque
models like neural nets. We developed
this method specifically because we
needed some way of making doctors
more comfortable with predictions
made by complex models such as neural
nets, SVMs,
and bagged/boosted decision trees. Case-Based Explanation
uses a complex learned model as a
distance function for k-nearest
neighbor so that kNN can generate explanations for each prediction the
complex model makes.
I'll wrap up by describing "Leo Breiman's
Law" which states that the
product of model error and model
opacity must be greater than Leo
Breiman's Constant. I'll discuss the evidence for this law, it's
implications for machine learning,
and what my research group is doing
to try to violate it.
Humans need to trust that intelligent systems are behaving correctly, and one way to achieve such trust is to enable people to understand the inputs, outputs, and algorithms used as well as any new knowledge acquired through learning. As the use of machine learning increases in critical operations it is being applied increasingly in domains where the learning system's inputs and outputs must be understood, or even modified, by human operators.
For instance, e-mail classification systems may need to gain the user's trust by explaining their predictions in a language the user can understand. Intelligent office assistants learn from a user's preferences and behavior, but in order to be useful, the user must trust that agent will make the same decisions the human would under the same conditions. Machine learning has also been widely used to support credit approval decisions; yet banks are becoming increasingly responsible for explaining the reasons behind a denial of credit. Autonomic systems are beginning to employ machine learning to support common administrative policies; yet system administrators are reluctant to trust automated technology they do not understand.
In this workshop we explore issues of human comprehensibility as it relates to machine learning.
Topics include the following:
§
Human-assisted learning
§
Knowledge acquisition for learning
§
Establishing and maintaining trust of users
§
Human understanding or modification of learning algorithms
§
Comprehensibility of the input or bias for learning
§
Comprehensibility of the induced model
§
Learning and explanation generation
§
Exploration and exploitation trade offs in the context of human
use
The workshop will feature talks by invited speakers as well as presentations of submitted work. The goal of the workshop is to facilitate discussion between participants. Thus, the schedule will include time for an extended group lunch session as well as a poster session in the afternoon where the participants will be able to share their ideas and open problems related to human-comprehensible machine learning.
We welcome submissions describing either relevant work or proposals for discussion topics that will be of interest to the workshop. Submissions are accepted in PDF format only, using the AAAI formatting guidelines at www. aaai.org/Workshops/. Submissions must be no longer than eight pages in length, including references and figures. Please e-mail submissions to oblio@us.ibm.com.
Dan Oblinger (Primary contact)
IBM T.J. Watson Research and
Telephone: 914-784-7531
Fax: 914-784-7455
E-mail: oblio@us.ibm.com
Mathias Bauer, German Research Center for
AI (DFKI) (bauer@dfki.de);
Yolanda Gil, USC Information Sciences Institute (gil@isi.edu);
Tessa Lau, IBM T.J. Watson Research (tessalau@us.ibm.com)