Explanation and Trust for Adaptive Systems Alyssa Glass (Stanford / SRI / Willow Garage) In collaboration with Deborah McGuinness (Stanford/RPI), Michael Wolverton (SRI), and Paulo Pinheiro da Silva (UTEP)
Dec 22, 2015
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Explanation and Trust for Adaptive Systems
Alyssa Glass(Stanford / SRI / Willow Garage)
In collaboration with Deborah McGuinness (Stanford/RPI), Michael Wolverton (SRI), and Paulo Pinheiro da Silva (UTEP)
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Outline
Adaptive Systems The User Perspective Trusting Adaptive Systems Discussion & Future Work
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Why is “trust” an issue? Systems are getting more complex
Hybrid and distributed processing Multiple learning components Multiple heterogeneous, distributed information
sources Highly variable reliability of information sources Less transparency of system computation and
reasoning Systems are taking more autonomous control
Guide/assist user actions Perform autonomous actions on behalf of user “reason, learn from experience, be told what to do,
explain what they are doing, reflect on their experience, and respond robustly to surprise” *
* DARPA PAL program: http://www.darpa.mil/ipto/programs/pal/
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
One Adaptive System: CALO Cognitive Assistant that Learns and
Organizes Personal office assistant, tasked with:
Noticing things in the cyber and physical environments Aggregating what it notices, thinks, and does Executing, adding/deleting, suspending/resuming
tasks Planning to achieve abstract objectives Anticipating things it may be called upon to do or
respond to Interacting with the user Adapting its behavior in response to past experience,
user guidance Contributed to by 22 different organizations
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Working with a Cognitive Assistant
CALO users need to Understand system behavior and responses Trust system reasoning and actions
To believe and act on recommendations from CALO, users need ways of exploring how and why the system acted, responded, recommended, and reasoned the way it did.
Additional wrinkle: CALO knowledge, behavior, and assumptions are constantly changing through several forms of machine learning.
A unified framework for explaining behavior and reasoning is essential for users to trust and adopt cognitive assistants.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Outline
Adaptive Systems The User Perspective Trusting Adaptive Systems Discussion & Future Work
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Interacting with Complex Systems
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Study Procedure 14 participants
12 men, 2 women Wide range of ages, education,
previous CALO experience Assigned tasks to accomplish
with CALO (many scripted) Told about trust study in
advance Structured interview format Identified 8 themes, in 3 major
categories
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Usability
Theme 1: Basic usability is important even in prototype-level systems.
“I can’t tell you how much I would love to have [the system]. But I also can’t tell you how much I can’t stand it.”
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Usability
Theme 2: Learning algorithms can give the impression that the user is being ignored.
“You specify something, and [the system] comes up with something completely different. And you’re like, it’s ignoring what I want!”
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
ExplanationsTheme 3: Users consistently want to ask context-sensitive questions, particularly when they are surprised by responses or failures.
What are you doing? Why did you do that? When will you be finished? What information sources did you use?
“If there had been an option to ask a question, I would have loved to ask a question.”
“I asked [‘Why?’] all the time, but I wasn’t getting answers!”
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Explanations
Theme 4: The granularity of feedback is important.
“I don’t just want an idiot light.”
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Trust
Theme 5: Users don’t trust opaque systems; they want transparency.
“The ability to check up on the system, ask it questions, get transparency to verify what it is doing, is the number one thing that would make me want to use it.”
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Trust
Theme 6: Access to information provenance can improve trust in both the information and the automated reasoning.
“[The system] needs a better way to have a meta-conversation.”
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Trust
Theme 7: Like in politics and the economy, gaining user trust relies on properly managing expectations.
“I was paralyzed with fear about what it would understand and what it would not.”
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Trust
Theme 8: Most users have a “trust but verify” attitude that makes system autonomy difficult without explainable verification.
“I trust [the system’s] accuracy, but not its judgment.”
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Outline
Adaptive Systems The User Perspective Trusting Adaptive Systems Discussion & Future Work
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
The Use-Ask-Understand-Update Cycle
Use Ask
UnderstandUpdate
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
What is an “explanation”…
<user>: Why are you doing <subtask>?<system>: I am trying to do <high-level-task>
and <subtask> is one subgoal in the process.
<user>: Why are you doing <high-level-task>?<user>: Why haven’t you completed
<subtask> yet?<user>: Why is <subtask> a subgoal of <high-
level-task>?<user>: When will you finish <subtask>?<user>: What sources did you use to do
<subtask>?
Initial request and answer strategy
Follow-up questions for mixed initiative dialogue
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
How Explanation Can Help
T1: Basic Usability T2: Being Ignored T3: Context-Sensitive Questions T4: Granularity of Feedback T5: Transparency T6: Provenance T7: Managing Expectations T8: Autonomy & Verification
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
How Explanation Can Help
T1: Basic Usability T2: Being Ignored T3: Context-Sensitive Questions T4: Granularity of Feedback T5: Transparency T6: Provenance T7: Managing Expectations T8: Autonomy & Verification
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
How Explanation Can Help
T1: Basic Usability T2: Being Ignored T3: Context-Sensitive Questions T4: Granularity of Feedback T5: Transparency T6: Provenance T7: Managing Expectations T8: Autonomy & Verification
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
The Integrated Cognitive Explanation Environment (ICEE)
Unified framework for explaining logical and task reasoning.
Applicable to multiple task execution systems.
Leverage existing InferenceWeb work for generating formal justifications.
Underlying task reasoning useful beyond explanation.
Provide sample implementation of end-to-end system.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Explanation ExampleSample question type: task motivation
Why are you doing <subtask>?
Strategy: reveal task hierarchyI am trying to do <high-level-task> and <subtask> is one subgoal
in the process.
Alternate strategies: Provide task abstraction Expose preconditions Expose termination conditions Reveal meta-information about task dependencies Explain provenance related to task preconditions or other
knowledge
Possible follow-up suggestions: Request additional detail Request clarification of of the given explanation Request an alternate strategy to the original query
McGuinness, D.L.; Glass, A.; Wolverton, M.; Pinheiro da Silva, P. A Categorization of Explanation Questions for Task Processing Systems. AAAI Workshop on Explanation-Aware Computing (ExaCt-2007), Vancouver, Canada, 2007.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Sidetrack:An InferenceWeb Primer
Trust
Explanation
Presentation
Abstraction
InferenceMeta-Language
InferenceRule
Specs
ProvenanceMeta-data
InformationManipulation
Data
Interaction
Understanding
Proof Markup Language
Framework for Framework for explainingexplaining reasoning and execution tasks by reasoning and execution tasks by abstracting, storing, exchanging, combining, annotating, filtering, abstracting, storing, exchanging, combining, annotating, filtering,
comparing, and rendering justifications from varied cognitive comparing, and rendering justifications from varied cognitive reasoners.reasoners.
1. Registry and service support for knowledge provenance.
2. Language for encoding hybrid, distributed proof fragments (both formal and informal).
3. Declarative inference rule representation for checking proofs.
4. Multiple strategies for proof abstraction, presentation, and interaction.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Sidetrack continued:Representations in PML
Proof Markup Language (PML) is a proof interlingua Used to represent justification of information
manipulation steps done by theorem provers, extractors, other reasoners
Main components concern inference representation and provenance issues
Specification written in OWL
iw:hasConclusion:
SupportsTopLevelGoal
iw:NodeSet
iw:isConsequenceOf
iw:InferenceStep
iw:hasLanguage:(Supports GA BL)
KIF
iw:hasRule:iw:hasSourceUsage:iw:hasEngine: SPARK
http://foo.com/Example.owl#Laptop
TailorComment
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Sample Interface Linked to ICEE
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Learning by Instruction
Relatively straight-forward to explain
Store instruction, resulting modification
Strategies present instruction and related meta-information
Demonstrated in CALO with Tailor task learning system
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Learning by Demonstration Generalizes user’s demonstration to learn a procedure One data point --> generalization will sometimes be
wrong Specifically, it will occasionally over generalize
Generalize the wrong variables, or too many variables Produce too general a procedure because of a coarse-
grained type hierarchy Explain the relevant aspects of the generalization
process To help the user identify and correct over
generalizations To help the user understand and trust the learned
procedures Working with LAPDOG task learning system in CALO
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Support-Vector Machines Augment SVM to gather additional meta-information about the SVM itself:
Support vectors identified by SVM Support vectors nearest to the query point Margin to the query point Average margin over all data points Non-support vectors nearest to the query point Kernel transformation used, if any
Represent SVM learning and meta-information as justification in PML, using added SVM rules
Design abstraction strategies for presenting justification to user as a similarity-based explanation
Demonstrated in CALO with PLIANT preference learning system PLIANT uses user-elicited preferences and past choices to learn user scheduling
preferences Inconsistent user preferences, over-constrained schedules, and necessity of
exploring the preference space result in user confusion about why a schedule is being presented.
Lack of user understanding of PLIANT’s updates creates confusion, mistrust, and the appearance that preferences are being ignored.
Provide justifications of schedule suggestions, without requiring user to understand SVM learning.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Future Work
Using conflicts to drive the learn-explain cycle
Using explanations to identify high-reward learning opportunities
Support more advanced dialogues and interfaces
User study using ICEE explanations
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Resources Explanation questions & strategies:
McGuinness, D.L., Glass, A., Wolverton, M., and Pinheiro da Silva, P. A Categorization of Explanation Questions for Task Processing Systems. AAAI 2007 Workshop on Explanation-Aware Computing (ExaCt-2007), Vancouver, Canada, 2007.
CALO trust study: Glass, A., McGuinness, D.L., and Wolverton, M. Toward Establishing Trust in
Adaptive Agents. Technical Report, KSL-07-04, Knowledge Systems, Artificial Intelligence Laboratory, Stanford University, 2007.
Explanation interfaces: McGuinness, D.L., Ding, L., Glass, A., Chang, C., Zeng, H., and Furtado, V. Explanation
Interfaces for the Semantic Web: Issues and Models. 3rd International Semantic Web User Interaction Workshop (SWUI’06), Athens, Georgia, 2006.
Overview of ICEE: McGuinness, D.L., Glass, A., Wolverton, M., and Pinheiro da Silva, P. Explaining Task
Processing in Cognitive Assistants that Learn. Proceedings of the 20th International FLAIRS Conference (FLAIRS-20), Key West, Florida, 2007.
Video demonstration of ICEE: http://iw.stanford.edu/2006/10/ICEE.640.mov