In the wake of recent acts of extreme brutality, injustice and mass protests, we are examining our role in perpetuating systems of inequality. We are responsible for our impact as a technology company, as news readers, and most importantly, as machine learning algorithm developers for Leo, your AI research assistant.
Artificial intelligence and machine learning are powerful tools that allow Leo to read thousands of articles published every day and prioritize a better selection based on the topics, organizations and trends that interest you. However, if not intentionally designed, these tools run the risk of strengthening harmful cultural prejudices.
The bias creeps into machine learning algorithms via incomplete or unbalanced training data. Without realizing it, we lose or over-represent some variables and the algorithm learns the wrong information, often with dangerous outcomes.
In Leo’s case, we risk introducing bias when we teach him broad topics like “leadership”. Leo learns about these topics by finding common themes in series of articles edited by the Feedly team. For the “leadership” topic, Leo could choose themes such as strong management skills and building a supportive team culture. However, if more articles about male leaders are published or added to the training set than about females, Leo may also learn that being male is a quality of leadership. Tracking the topics Leo learns is an essential part of topic modeling that helps us avoid reinforcing our own prejudices or those of the author or editor of the article.
It is up to us as developers to be deliberate and transparent how we take prejudices into account in our training process. With this in mind, we are excited to share what we are working on to reduce bias at the most crucial stage: training data.
Break down the silos
Collaboration between people from different backgrounds helps us explain our blind spots. However, to make this collaboration possible, we need an accessible tool. The new topic modeler is that tool, designed so that anyone in the Feedly community can help curate a dataset to train Leo on topics they’re passionate about.
The Topic Modeler leverages the Feedly UI we know and love to allow multiple users to search for articles for the training set and review Leo’s learning progress. Our goal is to connect with experts in a variety of fields to create solid topics that represent the entire community, not just the engineering team.
Test yourself: the theme of diversity
Recently, two members of the Feedly team with no machine learning experience and interested in diversity issues road-tested the new tool to redesign our diversity topic. The result is a rich and nuanced topic: instead of focusing only on the keyword “diversity,” Leo will search for thousands of related keywords, including representation, inclusion, prejudice, discrimination, equal rights and intersectionality. Now you can train Leos to track diversity and inclusion progress in your industry and find essential information on how to build and maintain inclusive work cultures and hiring practices.
Leo learns all the time
Topic modeling isn’t the only way to collaborate. Any Feedly user can help Leo learn. When Leo is wrong, you can use the “Less like this” down arrow button to let him know that an article he has prioritized is not about a particular topic.
Leo will also seek your feedback occasionally via a message at the top of an article. If you see “This article is about? [topic]?, “Let him know! Your feedback is incorporated into Leo’s training set to fill in any gaps we have missed and strengthen his understanding.
Join the movement
In addition to the in-app feedback, please feel free to contact us by email or join the Feedly Community Slack channel, especially if you have a topic that Leo needs to explore. This is the tip of the iceberg when it comes to addressing and dismantling systemic bias. We take our role as content brokers seriously and know we are indebted to those who have struggled so long to bring these issues to our attention. Leo is listening and learning.