- Former Amazon Web Services and Google engineer Bindu Reddy wants her startup Abacus.AI (formerly RealityEngines.AI) to make artificial intelligence more accessible and less biased.
- To work on the goal of de-biasing AI, her team created an open source library with three algorithms meant to help de-bias existing AI models.
- The problem with biased AI often stems from the fact that the data used to train models is incomplete or tainted by human biases, but Abacus can adjust systems without needing new data or to retrain them "from scratch."
- This new tool is launching as the startup raises $13 million in Series A funding led by Index Ventures and adds two new board members: Mike Volpi from Index and Google founding board member Ram Shriram.
- With its new funding, Abacus.AI plans to continue researching AI issues and turning its findings into actual products for customers.
- Click here for more BI Prime stories.
Bindu Reddy wants to solve what she calls the "hardest problems" in artificial intelligence with her startup, Abacus.AI.
Reddy, a former Amazon Web Services and Google engineer, started the company (formerly called RealityEngines.AI) to make AI more accessible for companies with small teams and limited data sets, and to make AI less biased.
Bias in AI is often linked to the dataset originally used to train an AI model: For example, if a facial recognition system is trained using a dataset that has more pictures of white people's faces than Black people's faces, it will result in a model that's not able to detect Black people's faces as well, or that makes inaccurate predictions based on a small sample size. Similarly, if a voice assistant is trained with mostly male voices, it will be less capable of understanding female voices.
Reddy's team has created an open source library with three algorithms meant to help de-bias AI, that can be used after-the-fact on an already-trained AI system.
This new product comes as Abacus.AI announces $13 million in Series A funding led by Index Ventures. The company also added Index partner Mike Volpi, to its board of directors, as well as Google founding board member Ram Shriram, who participated in the funding round. The round tripled the company's valuation from its seed round in 2019, Reddy said, but declined to share specific metrics. PitchBook says the company's valuation after its seed round was $25.25 million.
Algorithms to help make AI less biased
Reddy said she hopes that Abacus' new open source tools for de-biasing solve the root of the issue, which is that the data many are using to create AI models is biased in the first place.
"The biggest issue is really the data itself is biased," Reddy told Business Insider, "So what we're saying is, 'Look, even if you retrain the model, you may not get it to be unbiased, because a lot of existing data that people get — because it reflects a social bias — is going to be biased.' By using these techniques, we can actually un-bias it."
The system works by adjusting weights in an existing model based on a "fairness constraint," and can de-bias based on attributes like race, gender, and age (read the team's paper on its methods here). The idea is that an AI model that's already been trained can be "fixed" through Abacus' tools, without the need to completely retrain it with new data.
Bias in AI has been hot topic recently, as companies like Amazon, Microsoft and IBM have suspended the use of their facial recognition technology by police forces, amid a reckoning over systemic racism in the US following the deaths of George Floyd, Breonna Taylor and other Black people at the hands of police.
Despite the recent attention, Reddy said bias in AI is not a new issue, although Abacus' approach is novel. Typically, a company that wanted to de-bias its AI would completely retrain the algorithms with new data. But Reddy argues that that doesn't in fact solve the problem because, realistically, all data sets are going to be inherently biased because of human and societal biases. Her system makes more sense, she said, because it doesn't require new data or re-training "from scratch."
Ultimately, she believes that humans need to continue to come up with new ways to check their algorithmic systems for biases. She thinks AI can be monitored in the same ways that servers are, and her company wants to provide the tools to do that. "We could also have monitoring tools, which monitor bias regularly," she said.
Turning its research into products
In addition to the un-biasing algorithms, the company has been creating a library of AI templates for specific use cases including financial planning, sales and marketing, and fraud and security — things that are common uses for AI in the workplace. (That's how it's solving its mission to make AI more accessible for smaller teams.)
Right now the startup has 1,500 people testing its templates, including 35 customers, like 1-800-Flowers, Flex, DailyLook and Prodege, some of whom have started paying the company as well.
The idea is to create a business model like fintech company Stripe but for AI. Customers should be able to add AI features like recommendations and personalized search to mobile apps and websites. Right now it has 16 templates available and plans to use a freemium, pay as you go business model for customers.
The new funding will help continue with research and product development and grow its 22 person team.
"The bias thing is absolutely novel and new, and we invented it," Reddy said.
They team will continue to fund that kind of research, she added: "The other thing is obviously packaging all of this research into easy to use products, which customers can use."
Got a tip? Contact this reporter via email at pzaveri@businessinsider.com or Signal at 925-364-4258. (PR pitches by email only, please.) You can also contact Business Insider securely via SecureDrop.
Join the conversation about this story »
NOW WATCH: Here's what it's like to travel during the coronavirus outbreak