Ep. 44: Forget Polls, Here's What Street View, and AI, Can Tell You About How People Will Vote
Dec 20, 2017
auto_awesome
Timnit Gebru, a postdoctoral researcher at Microsoft and a Stanford AI Lab PhD graduate, dives into the intriguing intersection of AI and voting behaviors. She discusses how Google Street View data can predict demographic trends by analyzing vehicle types in neighborhoods. With a focus on the challenges of fine-grained image recognition and the ethics of AI, she emphasizes the need for fairness and accountability in algorithms. Gebru also shares fascinating insights from her research, underscoring the biases that can influence societal outcomes.
Using deep learning on Google Street View images enables insights into neighborhood demographics that traditional polling methods cannot provide.
The correlation between vehicle types and voting behavior highlights the potential of visual data to predict political leanings based on socio-economic status.
Deep dives
Innovative Use of AI for Demographic Insights
Using deep learning in conjunction with Google Street View images provides valuable insights into the demographic makeup of neighborhoods across the United States. By analyzing 15 million images from 200 populous cities, researchers were able to classify over 2,600 distinct car types to infer the socio-economic characteristics of surrounding areas. The methodology involved correlating data on these cars with ground truth information from census data and voting records, allowing the team to draw meaningful conclusions. This innovative approach differs from traditional polling by relying on visual data rather than direct voter feedback, potentially offering a more nuanced view of community demographics.
Challenges in Fine-Grained Image Recognition
The project required a sophisticated level of fine-grained image recognition to distinguish between visually similar car models, which posed a significant challenge. Assembling a dataset that accurately labeled cars for training the model was one of the most arduous tasks, involving whittling down an initial list of 15,000 car types to a manageable 2,600 classes based on recognizable visual characteristics. The classification process not only required the detection of the car type but also the use of convolutional neural networks to improve accuracy. Despite achieving a 33% accuracy rate with complex classifications, researchers emphasized the difficulties inherent in this nuanced segmentation of data.
Linking Automobiles to Voting Patterns
The research unveiled intriguing correlations between vehicle types and voting behavior in U.S. elections, suggesting that the presence of certain cars could indicate the political leanings of a neighborhood. For instance, sedans were strongly associated with Democratic precincts, while pickup trucks were correlated with Republican ones, reflecting broader trends in vehicle ownership linked to socio-economic status. By aggregating data features, such as average miles per gallon and car density, researchers were able to conduct analyses that predicted voting patterns with notable accuracy. However, they cautioned against over-claiming the implications, stressing that these findings pertain to aggregate data rather than individual behavior.
Election polling is an inexact science. If you've been paying attention to American politics at all over the past year or two, you don't need us to tell you that. But what if instead of asking voters their opinions on the candidates or the issues you took a different approach, one that involves artificial intelligence... and cars. Joining us for this edition of the AI podcast is Timnit Gebru, a post-doctoral researcher at Microsoft Research in New York and a newly minted PhD from the Stanford Artificial Intelligence Laboratory. Timnit is co-author of a paper titled "Using Deep learning and Street View to Estimate the Demographic Makeup of Neighborhoods Across the United States."
Get the Snipd podcast app
Unlock the knowledge in podcasts with the podcast player of the future.
AI-powered podcast player
Listen to all your favourite podcasts with AI-powered features
Discover highlights
Listen to the best highlights from the podcasts you love and dive into the full episode
Save any moment
Hear something you like? Tap your headphones to save it with AI-generated key takeaways
Share & Export
Send highlights to Twitter, WhatsApp or export them to Notion, Readwise & more
AI-powered podcast player
Listen to all your favourite podcasts with AI-powered features
Discover highlights
Listen to the best highlights from the podcasts you love and dive into the full episode