CACTUS: Detecting and Resolving Conflicts in Objective Functions

Abstract

Machine learning (ML) models are constructed by expert ML practitioners using various coding languages, in which they tune and select models hyperparameters and learning algorithms for a given problem domain. They also carefully design an objective function or loss function (often with multiple objectives) that captures the desired output for a given ML task such as classification, regression, etc. In multi-objective optimization, conflicting objectives and constraints is a major area of concern. In such problems, several competing objectives are seen for which no single optimal solution is found that satisfies all desired objectives simultaneously. In the past VA systems have allowed users to interactively construct objective functions for a classifier. In this paper, we extend this line of work by prototyping a technique to visualize multi-objective objective functions either defined in a Jupyter notebook or defined using an interactive visual interface to help users to: (1) perceive and interpret complex mathematical terms in it and (2) detect and resolve conflicting objectives. Visualization of the objective function enlightens potentially conflicting objectives that obstructs selecting correct solution(s) for the desired ML task or goal. We also present an enumeration of potential conflicts in objective specification in multi-objective objective functions for classifier selection. Furthermore, we demonstrate our approach in a VA system that helps users in specifying meaningful objective functions to a classifier by detecting and resolving conflicting objectives and constraints. Through a within-subject quantitative and qualitative user study, we present results showing that our technique helps users interactively specify meaningful objective functions by resolving potential conflicts for a classification task.

References in corpus (16)

• Show, Attend and Tell: Neural Image Caption Generation with Visual Attention
• Axiomatic Attribution for Deep Networks
• Learning Important Features Through Propagating Activation Differences
• Interpretable classifiers using rules and Bayesian analysis: Building a better stroke prediction model
• SmoothGrad: removing noise by adding noise
• Sanity Checks for Saliency Maps
• Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)
• Multi-Task Learning as Multi-Objective Optimization
• Supersparse Linear Integer Models for Optimized Medical Scoring Systems
• SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability
• Do Convolutional Neural Networks Learn Class Hierarchy?
• Adaptively Learning the Crowd Kernel
• Deep Learning for Case-Based Reasoning through Prototypes: A Neural Network that Explains Its Predictions
• An Overview of Machine Teaching
• How to Read Many-Objective Solution Sets in Parallel Coordinates
• Interpretable Image Recognition with Hierarchical Prototypes