Modeling of moral decisions with deep learning

Abstract

One example of an artificial intelligence ethical dilemma is the autonomous vehicle situation presented by Massachusetts Institute of Technology researchers in the Moral Machine Experiment. To solve such dilemmas, the MIT researchers used a classic statistical method known as the hierarchical Bayesian (HB) model. This paper builds upon previous work for modeling moral decision making, applies a deep learning method to learn human ethics in this context, and compares it to the HB approach. These methods were tested to predict moral decisions of simulated populations of Moral Machine participants. Overall, test results indicate that deep neural networks can be effective in learning the group morality of a population through observation, and outperform the Bayesian model in the cases of model mismatches.

Keywords: Artificial intelligence; Bayesian method; Deep learning; Moral machine experiment.

Fig. 1

Moral Machine Example: Example screenshot of a scenario from a Moral Machine scenario [8]. The participant must decide the more ethical course of action for the self-driving vehicle: swerving would kill all vehicle passengers (right), while maintaining the course would result in the death of all pedestrians (left) [8]

Fig. 2

Feature Transform: Binary transformation matrix A used in Kim, et al., which converts a set of character traits θ into quantifiable features [9]

Fig. 3

Moral Principle Mean: Mean value for all underlying distributions of the moral principle vector w

Fig. 4

Moral Principle Covariance: Covariance matrix for all underlying distributions of the moral principle vector w

Fig. 5

Distribution Transform: Histograms of marginal distributions of the ’Human’ value for the normal dataset (left) and one of the transformed datasets (right)

Fig. 6

Neural Network Architecture: Network architecture for the DL moral decision model. Batchnorm signifies a batch normalization layer

Fig. 7

Generated Distribution Mean: Mean values of the principle vector w used to synthesize more general data

Fig. 8

Generated Distribution Covariance: Covariance matrix of the principle vector w used to synthesize more general data

Fig. 9

DL Accuracy: DL model predictive accuracy with different sample sizes and underlying distributions of w. Various distributions for w are denoted by the approximate average of absolute shape factors k ($\left|\overline{k}\right|=0$ denotes the Gaussian distribution)

Fig. 10

DL Accuracy (Deterministic Decision Process): DL model predictive accuracy with different sample sizes and underlying distributions of w. Various distributions for w are denoted by the approximate average of absolute shape factors k ($\left|\overline{k}\right|=0$ denotes the Gaussian distribution) In this test, decisions were simulated with the deterministic version of Eq. 2

Fig. 11

Model Performances: Comparison of model performances (DL trained with 2,000 participants involved) over datasets with various underlying distributions of w, denoted by the approximate average of absolute shape factors k ($\left|\overline{k}\right|=0$ denotes the Gaussian distribution). ‘GT’ denotes the predictive model in which the exact w for each participant is known

Fig. 12

Model Performances (Deterministic Decision Process): Comparison of the model performances (DL trained with 2,000 participants involved) over datasets with various underlying distributions of w, denoted by the approximate average of absolute shape factors k ($\left|\overline{k}\right|=0$ denotes the Gaussian distribution). In this test, decisions were simulated with the deterministic version of Eq. 2

Fig. 13

DL Model Accuracy (Generated): DL model predictive accuracy with different sample sizes and underlying distributions of w, applied to the generalized, abstract data. Various distributions for w are denoted by the approximate average of absolute shape factors k ($\left|\overline{k}\right|=0$ denotes the Gaussian distribution)

Fig. 14

DL Model Accuracy (Generated; Deterministic Process): DL model predictive accuracy with different sample sizes and underlying distributions of w, applied to the generalized, abstract data. Various distributions for w are denoted by the approximate average of absolute shape factors k ($\left|\overline{k}\right|=0$ denotes the Gaussian distribution) In this test, decisions were simulated with deterministic version of Eq. 2

Fig. 15

Model Performances (Generated): Comparison of model performances (DL trained with 2,000 participants involved) over datasets with various underlying distributions of w, denoted by the approximate average of absolute shape factors k ($\left|\overline{k}\right|=0$ denotes the Gaussian distribution), as applied to the generalized, abstract data. ‘GT’ is the predictive model in which the exact w for each test participant is known

Fig. 16

Model Performances (Generated; Deterministic Process): Comparison of model performances (DL trained with 2,000 participants involved) over datasets with various underlying distributions of w, denoted by the approximate average of absolute shape factors k ($\left|\overline{k}\right|=0$ denotes the Gaussian distribution), as applied to the generalized, abstract data. In this test, decisions were simulated with the deterministic version of Eq. 2