Radical Reads

The Future is Federated

By Daniel Mulet, Investor


Credit: Analytics India Magazine

Federated Learning (FL) is a powerful machine learning technique that has been quietly gaining adoption in settings where preserving privacy is a top priority.

The core concept behind FL is that an AI model can be trained using data from multiple owners without sacrificing privacy. For example, a FL-powered application running on a smartphone will not use personal data to tune its model. Rather, it will share anonymized data to a centralized model for updating. At scale, across hundreds of millions of phones, such a system would maintain the privacy of individual users while still serving them high performance AI models.

Facing regulatory pressure in multiple jurisdictions, Google has been working on applying FL to their advertising business to improve its privacy measures. Earlier this week the search giant announced a proposal to “eliminate third-party cookies by replacing them with viable privacy-first alternatives,” through the creation of Federated Learning of Cohorts (FLoC). Instead of having third party cookies sending your browsing activity to centralized databases, local models running in your browser will place you in a “cohort” of several thousand people based on your browsing behaviour. The data on these cohorts would then replace cookies, and be sent via an API to parties interested in serving the cluster relevant ads, while preserving the privacy of each individual.

This decentralization of personal data opens up a new paradigm for AI as compute moves from the cloud and data centers, to the increasingly powerful AI chips in our devices. We believe privacy-preserving technologies like FL will be widely adopted within healthcare, finance and other regulated industries where preserving individual privacy is essential.

AI News This Week

  • Unlimited computer fractals can help train AI to see  (MIT Technology Review)

    Most image-recognition systems are trained using large databases that contain millions of photos of everyday objects. Now researchers in Japan have shown that AIs can start learning to recognize everyday objects by being trained on computer-generated fractals instead. Using an endless supply of synthetic images rather than photos scraped from the internet avoids problems with existing hand-crafted data sets that are hard to produce and may contain bias.

  • The cloud infrastructure market hit $128B in 2020  (TechCrunch)

    Digital transformation trends continue to accelerate. The cloud infrastructure market in 2020 grew to $129B for the year according to data from Synergy Research Group, up from $97B in 2019. Amazon, Microsoft and Google hold the top three spots of the growing market. Gartner estimates that less than 4% of IT spend is on cloud infrastructure, leaving room for more growth in the coming years.

  • As Robots Fill the Workplace, They Must Learn to Get Along  (Wired)

    Warehouses, factories and hospitals are deploying more robots often made by different companies that can lead to communication problems. Changi General Hospital in Singapore has about 50 robots from eight manufacturers. To alleviate communication problems, Changi is using open-source software to let robots from different manufacturers talk to each other and negotiate safe passage. Mobile robots are increasingly found in factories, warehouses, hospitals and stores, ferrying goods, inspecting shelves or cleaning floors and will need to look for a similar open source platform for interoperability.

  • Pngme, a financial data platform, closes $3M seed to accelerate growth in Sub-Saharan Africa  (TechCrunch)

    There has been a proliferation of fintech services ranging from wallets, to savings, to loans, creating a lot of fragmented data in the process. This makes it difficult for financial institutions to understand and provide insights from users’ data. Pngme, a unified financial data API platform, is looking to solve this problem. In many cases, the aggregation of financial and non-financial data will allow unbanked and underserved customers to have a maiden credit score, opening up more opportunities to access financial products. The Africa-focused, but US-based, company announced its seed round of $3 million this week led by Radical Ventures (Fund 1), Raptor Group, Lateral Capital and EchoVC.

  • AI Emerges as Crucial Tool for Groups Seeking Justice for Syria War Crimes  (Wall Street Journal)

    As the United Nations, Europeans authorities and human-rights groups build war-crimes cases, they have turned to artificial intelligence. AI and machine learning is playing an integral role in sorting through the huge trove of evidence to bring warm criminals to justice in Syria and can serve as a model for investigations into other modern-day conflicts. The technology is aimed at helping process, organize and analyze the data and reduce the time human investigators spend sifting through terabytes of traumatic videos and images.

Radical Reads is edited by Leah Morris (Senior Director, Velocity Program, Radical Ventures).