Grant: Discovery Grants Program - Individual
University of Alberta
$41,000.00
Improving the Explainability and Robustness of Deep Learning via Computational Intelligence
EF 2023-24
Résumé
This proposal requests funding for an ongoing program of research in computational intelligence, with a focus on fuzzy logic and neural networks. Over the past six years, my group has worked on the theory of complex fuzzy sets and logic; their operationalization via machine learning; comparing those systems and competing intelligent approaches for monitoring complex physical systems; mitigating sample selection bias; and designing deep fuzzy systems for improved explainability. This work has led to three completed PhD dissertations, five MSc theses, and 22 journal publications submitted or published in the last six years. In the current proposal, we build upon this foundation to investigate two critically-important challenges in deep learning: 1) mitigating brittleness in deep neural nets, and 2) improving the accuracy of explainable deep models. Both are topics of intense current interest, as they bear directly on the issue of trust in AI, a technology that is expected to be a wellspring of economic growth. If, however, users do not trust AIs - because they are too brittle, or because high-performing AIs are incomprehensible - all that promise may be lost, as humanity refuses to use AI systems. In (1), based on recent investigations and earlier work, we propose that a combination of sample selection bias and small disjuncts within the training set support are the common cause of brittleness and adversarial examples. To test our proposal, we will use decision-tree models to duplicate decision boundaries induced from in-distribution and out-of-distribution examples. Disagreements between the two are expected to indicate brittle regions of input space; we can then examine those regions in more detail for small disjuncts. To mitigate brittleness, we will then design robust hybrids of deep learning and fuzzy logic. In (2) we follow up on recent findings that explainability could improve the accuracy of deep learning, an apparent contradiction of the well-known accuracy/interpretability tradeoff. We interpret this as regularizing deep learning with the reflective process underlying explanation generation. We propose multiple approaches for designing such regularized networks, and a screening strategy to select which ones are worth evaluating in user studies. This strategy relies on results from an industrial condition-monitoring project where we observed brittleness firsthand; our test is if the explanations can duplicate the findings of a root-cause analysis currently underway. The proposed research is anticipated to offer new insights into these two phenomena, and evaluate AI designs that improve these dimensions of trustworthiness. 2 PhD and 2 MSc students will be trained in designing AIs, while students on other other funded projects focus on AI applications in intelligent systems.
Programme
Discovery Grants Program - Individual
Bénéficiaire
Emplacement
Alberta, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data
Tel que publié
Application id
RGPIN-2023-04455
Name nom
Dick, Scott
Department département
Electrical and Computer Engineering - Electrical and Computer Engineering
Organization id
9
Institution établissement
University of Alberta
Province en
Alberta
Province fr
Alberta
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
41000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
802
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Information systems and technology
Area of application fr
Systèmes d'information et technologie de l'information
Research subject code
9999
Research subject group en
Not available
Research subject group fr
Not available
Research subject en
Non disponible
Research subject fr
Non disponible
Application title
Improving the Explainability and Robustness of Deep Learning via Computational Intelligence
Keyword
Computational Intelligence; Deep Learning; Expalinable Artificial Intelligence; Fuzzy Logic; Machine Learning
Application summary
This proposal requests funding for an ongoing program of research in computational intelligence, with a focus on fuzzy logic and neural networks. Over the past six years, my group has worked on the theory of complex fuzzy sets and logic; their operationalization via machine learning; comparing those systems and competing intelligent approaches for monitoring complex physical systems; mitigating sample selection bias; and designing deep fuzzy systems for improved explainability. This work has led to three completed PhD dissertations, five MSc theses, and 22 journal publications submitted or published in the last six years. In the current proposal, we build upon this foundation to investigate two critically-important challenges in deep learning: 1) mitigating brittleness in deep neural nets, and 2) improving the accuracy of explainable deep models. Both are topics of intense current interest, as they bear directly on the issue of trust in AI, a technology that is expected to be a wellspring of economic growth. If, however, users do not trust AIs - because they are too brittle, or because high-performing AIs are incomprehensible - all that promise may be lost, as humanity refuses to use AI systems. In (1), based on recent investigations and earlier work, we propose that a combination of sample selection bias and small disjuncts within the training set support are the common cause of brittleness and adversarial examples. To test our proposal, we will use decision-tree models to duplicate decision boundaries induced from in-distribution and out-of-distribution examples. Disagreements between the two are expected to indicate brittle regions of input space; we can then examine those regions in more detail for small disjuncts. To mitigate brittleness, we will then design robust hybrids of deep learning and fuzzy logic. In (2) we follow up on recent findings that explainability could improve the accuracy of deep learning, an apparent contradiction of the well-known accuracy/interpretability tradeoff. We interpret this as regularizing deep learning with the reflective process underlying explanation generation. We propose multiple approaches for designing such regularized networks, and a screening strategy to select which ones are worth evaluating in user studies. This strategy relies on results from an industrial condition-monitoring project where we observed brittleness firsthand; our test is if the explanations can duplicate the findings of a root-cause analysis currently underway. The proposed research is anticipated to offer new insights into these two phenomena, and evaluate AI designs that improve these dimensions of trustworthiness. 2 PhD and 2 MSc students will be trained in designing AIs, while students on other other funded projects focus on AI applications in intelligent systems.