Digital Divides in Scene Recognition: Uncovering Socioeconomic Biases in Deep Learning Systems (2401.13097v1)

Abstract: Computer-based scene understanding has influenced fields ranging from urban planning to autonomous vehicle performance, yet little is known about how well these technologies work across social differences. We investigate the biases of deep convolutional neural networks (dCNNs) in scene classification, using nearly one million images from global and US sources, including user-submitted home photographs and Airbnb listings. We applied statistical models to quantify the impact of socioeconomic indicators such as family income, Human Development Index (HDI), and demographic factors from public data sources (CIA and US Census) on dCNN performance. Our analyses revealed significant socioeconomic bias, where pretrained dCNNs demonstrated lower classification accuracy, lower classification confidence, and a higher tendency to assign labels that could be offensive when applied to homes (e.g., "ruin", "slum"), especially in images from homes with lower socioeconomic status (SES). This trend is consistent across two datasets of international images and within the diverse economic and racial landscapes of the United States. This research contributes to understanding biases in computer vision, emphasizing the need for more inclusive and representative training datasets. By mitigating the bias in the computer vision pipelines, we can ensure fairer and more equitable outcomes for applied computer vision, including home valuation and smart home security systems. There is urgency in addressing these biases, which can significantly impact critical decisions in urban development and resource allocation. Our findings also motivate the development of AI systems that better understand and serve diverse communities, moving towards technology that equitably benefits all sectors of society.

Introduction

The efficacy of computer-based scene recognition is paramount in a diverse range of applications. However, examining whether these systems are equitable across social demographics has gained urgency. Researchers at Barnard College and Bates College have presented an in-depth analysis of deep convolutional neural networks (dCNNs) and their propensity to exhibit socioeconomic bias in scene classification tasks.

Dataset and Methodology

Focusing on nearly one million images from both global and US sources, the paper analyzed images from user-submitted home photographs and Airbnb listings. The multidimensional approach included statistical models to understand the influence of socioeconomic indicators such as family income, Human Development Index (HDI), and various demographic factors on dCNN performance. This allowed the identification of any correlative biases present in pretrained dCNNs, specifically regarding classification accuracy, confidence, and the potential assignment of offensive labels such as "slum" or "ruin" to images of homes.

Results and Analysis

Results revealed that pretrained dCNNs demonstrate lower classification accuracy and confidence, and a higher tendency to assign offensive labels, particularly in relation to homes from lower socioeconomic status (SES). This trend held true in both international comparisons and within the United States, suggesting a consistent bias correlating with economic and developmental factors. Furthermore, in a granular analysis using the Airbnb dataset, the researchers identified that classification entropy—reflecting uncertainty in classification—demonstrates a bias in computer vision systems, favoring images from more developed countries and areas with higher GDP per capita, literacy rate, and urbanization percentage.

Implications and Future Directions

These findings accentuate significant disparities in AI performance influenced by socioeconomic factors. The paper carries profound implications for fairness and equity in applied AI technologies, such as smart home devices and urban planning tools. By elucidating the biases in existing systems, the research points toward the necessity of constructing more inclusive training datasets. It proposes that rectifying the composition of training data is critical to prevent deep learning systems from echoing societal inequities. Additionally, the paper calls for further investigation into the development processes of AI systems, advocating for algorithmic diversity and a culture of conscious inclusivity in the field of AI. Moving forward, addressing these biases is critical to ensure technology equitably benefits all sectors of society.