Google’s DeepMind project is responsible for headline-grabbing wins for artificial intelligence (AI) over human capabilities. For example, DeepMind technology was used to build AlphaGo, the software program responsible for beating 9th dan Go player Lee Sedol 4-1 in March 2016 (strangely, Lee won his sole game thanks to an extraordinarily rare move known as “the Divine Hand”) in a best of five match.
In 2017, AlphaGo beat Ke Jie, who at the time held the number one rank in Go for two years in a three-game series. Later that year, AlphaGo finally lost ... to AlphaGo Zero, another DeepMind AI creation. AlphaGo Zero destroyed AlphaGo with self-taught strategies after three days of teaching itself. It won with less processing power as well.
Why Use Neuroscience to Study AI?
As a species, we now have the unprecedented ability to build tools that work better than humans without fully understanding how they work better.
In this vein of inquiry, researchers at Harvard University and DeepMind collaborated to create a virtual AI rat to carry out complex tasks in a digital environment. The research, described in a paper titled, “Deep neuroethology of a virtual rodent,” details their goal of creating a method to study artificial intelligence in much the same way that lab researchers study rodents.
The virtual rat’s decision-making process was created from neural networks with differing levels of realistic biological qualities to test how well it managed various complicated tasks. In other words, the digital biology of the virtual rodent was designed from the physical biology of real rats.
Using measurements and movements of rodent bones, joints and muscles gives the virtual rat limitations in movement that are similar to those of a real rat. The virtual rat was coded with vision and proprioception, giving it a digital awareness and a virtual proprioception that act as the virtual rat’s awareness and feedback systems.
Bottom Line
As you can see from the video, the virtual rat’s movements don’t resemble those of an actual rat. But that is beside the point.
The main goal of the collaboration between DeepMind and Harvard researchers is to use established neuroscientific methods to study AI.
More specifically, the main goal is to study how AI encapsulated in an agent (in this case, a virtual rat created from the biological data of a real rat) makes deep-learning network-based decisions.