# Interpretable deep-learning models to help achieve the Sustainable Development Goals

**Authors:** Ricardo Vinuesa, Beril Sirmacek  
**arXiv:** [2108.10744](https://arxiv.org/abs/2108.10744)  
**Date:** 2021-08-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.10744v1/

## Key points

- Interpretable AI models play a crucial role in developing ethical AI systems and data-driven solutions aligned with the Sustainable Development Goals (SDGs).
- Extracting truly interpretable models from deep-learning methods, such as through symbolic models obtained via inductive biases, ensures sustainable advancement of AI technology.
- Transparency and interpretability are needed in applications like facial recognition, loan assignments, recruitment decisions, and healthcare to understand decisions made by black-box AI models.
- Tracking poverty using satellite images and convolutional neural networks (CNNs) can accurately predict economic consumption per capita; adding interpretability enhances understanding of influencing factors and improves poverty-tracking efforts aligned with SDG 1.
- Methodologies like those proposed by Cranmer et al., which add interpretability to AI models based on symbolic equations, help professionals and policymakers gain insights into model reasoning for enhancing improvements aligned with SDGs.
- Prioritizing interpretable AI models for societal benefit and environmental sustainability influences AI researchers and policymakers towards fostering transparency, accountability, and ethical decision-making within AI systems for a more sustainable future.

## Summary

In our discussion on interpretable artificial intelligence (AI) models, we emphasize their crucial role in the development of ethical AI systems and data-driven solutions aligned with the Sustainable Development Goals (SDGs). By extracting truly interpretable models from deep-learning methods, such as through symbolic models obtained via inductive biases, we can ensure a sustainable advancement of AI technology. We provide examples where black-box AI models make decisions with significant implications for individuals, highlighting the need for transparency and interpretability in applications like facial recognition, loan assignments, recruitment decisions, and healthcare. One specific example showcases how tracking poverty using satellite images and convolutional neural networks (CNNs) can accurately predict average economic consumption per capita by identifying features like night-light intensity and distance to urban areas. Adding interpretability to this model could enhance understanding of the factors influencing outcomes and improve poverty-tracking efforts aligned with SDG 1. Through the methodology proposed by Cranmer et al., which adds interpretability to AI models based on symbolic equations, professionals and policymakers can gain insights into model reasoning to enhance model improvements and design actions in line with SDGs. We aim to influence AI researchers and policymakers towards prioritizing interpretable AI models for societal benefit and environmental sustainability. Achieving interpretable models not only aligns with Trustworthy-AI assessment guidelines but also enhances real-world applications that serve the greater good. Overall, our focus on interpretable AI models underscores their importance in fostering transparency, accountability, and ethical decision-making within AI systems for a more sustainable future.