Fwd: RI Ph.D. Thesis Proposal: Nicholas Gisolfi

[Artur Dubrawski]

Team,

Please consider attending a thesis proposal presentation by Nick (this
Tuesday, ie. tomorrow, 10am, see zoom link below).
This talk looks to be really fun to hear.

Cheers
Artur

---------- Forwarded message ---------
From: Suzanne Lyons Muth <lyonsmuth at cmu.edu>
Date: Mon, Jan 11, 2021 at 9:40 AM
Subject: RI Ph.D. Thesis Proposal: Nicholas Gisolfi
To: ri-people at lists.andrew.cmu.edu <ri-people at lists.andrew.cmu.edu>


Date:  19 January 2021
Time:  10:00 a.m. (ET)
Place: *Virtual Presentation*
https://cmu.zoom.us/j/96434017510?pwd=amJXYlFTUDAvMmp6NkprUUtxR0xoUT09
Type:  Ph.D. Thesis Proposal
Who:  Nicholas Gisolfi
Title:  Verification and Accreditation of Artificial Intelligence

Abstract:
This work involves formally verifying a trained model's adherence to
important design specifications for the purpose of model accreditation.
Accreditation of a trained model requires enumeration of the explicit
operational conditions under which the model is certified to meet all
necessary specifications.  By verifying model adherence to specifications
set by developers, we increase the trustworthiness of the model along the
dimensions that are most relevant to earning a developer's trust.  We argue
that this gives developers a tool to quantitatively define and formally
answer the fundamental question of ‘should we trust this model’.

We intend to demonstrate utility of the framework with the verification of
a repertoire of design specifications on voting tree ensemble models for
classification.  Verification is posed as an instance of the Boolean
Satisfiability (SAT) problem.  Our novel SAT encoding yields disruptive
speed gains over related tree ensemble verification techniques, which
facilitates the verification of harder specifications on models of larger
scale than reported in literature.  It is possible to extend our framework
to other model classes and design specifications that are beyond our
current scope.

We show it is possible to use our framework to provide explanations for
model behavior.  We introduce a quantitative definition for
interpretability; a provably correct interpretation of a model is a causal
specification relating inputs to outputs that the model never violates.  We
demonstrate how we can mine data for a specific type of causal structure;
2D, axis-aligned range rules that offer candidate specifications for
verification.  These bounding box rules represent intelligible,
intermediate-level concepts that, pending verification, offer provably
correct summaries of model behavior.  In addition to explanations for
input-output behavior of a model, we demonstrate that it is possible to
provide explanations that diagnose discrepancies between what is observed
and what was prescribed; if the model violates a required specification, we
want to understand why.  We provide explanations by summarizing proofs of
unsatisfiability that accompany the certificates.  A minimized
unsatisfiable set comprises the literals and clauses implicated in a
logical contradiction.  These provide the basis for provably correct
explanations that are counterfactual and contrastive in nature.

We propose a few directions to further demonstrate and expand the utility
of our accreditation framework.  For a supervised radiation risk assessment
application, where safety is paramount, we propose verifying a suite of new
safety criteria that will increase domain-expert trust in the model's
predictions.  For selected clinical decision support applications, we
propose to diagnose the sources of systematic disagreements between
opinions of experts who annotate training data, and to assess the model's
adherence to common-sense-based expectations of its behavior.  We expect
these new capabilities to enhance adoption and effectiveness of artificial
intelligence applications to guide clinical decisions.

Thesis Committee Members:
Artur Dubrawski, Chair
Stephen Smith
Reid Simmons
Madalina Fiterau, University of Massachusetts Amherst

A copy of the thesis proposal document is available at:
https://cmu.box.com/s/gq9lumhvw7r0cnug3t9jnzrrb0jfcg5j
_______________________________________________
ri-people mailing list
ri-people at lists.andrew.cmu.edu
https://lists.andrew.cmu.edu/mailman/listinfo/ri-people
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/autonlab-users/attachments/20210118/b86ec843/attachment.html>