The history of Bear

In 2008 I was looking graphically at statistical dependencies in some stock market data in what I thought was a novel way. (It wasn’t.)

By late 2010 I had started to build data structures and algorithms to extract the statistically significant structure out of these sorts of dependencies, while discarding the statistical noise, in a way that was computationally tractable.

The day after my 45th birthday, in September 2011, a new television series premiered in Australia, Person of Interest, which was based on a machine learning system. It quickly became a favorite for my wife Sally and me. It prompted me to google machine learning systems, and I soon got the idea that my data structures and algorithms could be the basis of a simple machine learning system.

In early 2012, as I was developing this system, a former colleague told me about the Kaggle Data Science Competition website. I decided to use one of the competitions, the Heritage Health Prize, as test data. In May and early June I made some test submissions.

Then, on June 5, 2012, Kaggle announced a new competition: the Facebook Recruiting Competition. Intrigued, and despite not having authorization to work in the U.S., I switched my attention to this competition, making my first submission on June 8. Over the next two weeks I was amazed to find that my system was consistently providing submissions ranked number 2 on the Leaderboard, despite it being an after-hours rudimentary project running on just a dual-core laptop.

On June 24, another competitor “blew up” the competition, essentially restarting it from a new baseline. Due to this incident, on June 26 I announced my withdrawal and made a final submission, including all of my C# source code and documentation.

I thought that that was the end of the story.

After the competition finished on July 10 (my final June 26 submission had dropped to #24 by then), and after the Kaggle and Facebook staff had worked through all of the top-finishing entrants who had the right to work in the United States, I was emailed on July 27 by a Facebook recruiter, offering me the opportunity to interview for a Software Engineer position in a just-announced new London office. After explaining that the U.S. E-3 Visa meant that it was actually much easier for me to work in the United States than the United Kingdom, I was given a phone screen for a U.S. position on August 7, a full interview loop at Facebook’s new headquarters in Menlo Park, California, on August 29 (following a full loop the previous day at Google, who had approached me in the mean time), was offered and accepted the position of Software Engineer on September 5 (superior to the Google position offered just minutes later), and flew to San Francisco, California, on October 6 to start work at Facebook headquarters on October 8, 2012.

What a whirlwind!

Just after my birthday, and before flying out to start my new job at Facebook, a new character was introduced in the Season 2 premiere of Person of Interest: Bear. Our love of dogs together with our love of Person of Interest made our next decision an easy one: my machine learning system, which had opened up this incredible opportunity for us, would from that time onwards be codenamed Bear.

Ironically, but understandably, Facebook was actually interested in me—my software engineering skills and abilities—not Bear, which was consequently relegated to the backburner for more than a decade!

After that long hibernation, I finally released Bear here on this website, as a beta Version 0.1, as free and open source software under the MIT License, on March 21, 2023.

I updated Bear to Version 0.2 on April 9, 2023, to automatically swap memory out to storage. Version 0.3, released on April 28, 2023, enabled streaming data through a pre-trained model deployed to production.

These versions classified the MNIST database to 99% accuracy within minutes. I intended to release a full “Convolutional Bear” some time in 2023, but soon realized that my architecture for Bear was going down the wrong path.

In Version 0.4, “Bear Bones,” released on August 18, 2023, I strived to simplify Bear back towards its 2012 world-beating philosophy.

After an extended vacation back in Australia for the last few months of 2023, in early 2024 I borrowed liberally from the philosophy underlying random forests to completely overhaul Bear away from a “best model” decision-making approach towards a forest-of-models approach. Version 0.5, “Bear Forest,” released on July 7, 2024, under the MIT-0 License, is the result of this overhaul.