— Written by Tim Marzullo —
Over the past 80 years, we have exploited the invention of computers to make calculations much faster than our human brains can. From the ENIAC machines of post-WWII predicting ballistic trajectories to the contemporary Google Colab notebooks we now use that process our team’s electrophysiology data in the cloud, we have marveled at the speed of information processing in computational systems and have used it for economic, scientific, medical, and national defense benefits.
Throughout the ’80s, ’90s, and 2000s, we accepted that while yes, computers are faster than humans at calculations, we are animals that can move throughout the world. Moreover, we do so at low power. (How many times have you only had a coffee and an apple for breakfast, and managed to work for an entire day, albeit unhappily, on your body’s energy reserves?) But over the past decades, laptops have approached power requirements on the scale of the human brain (20 W). Robots such as the Boston Dynamics Atlas still have high power requirements though for their human size robots.
During the computer revolution, we have used electronics, either with vacuum tubes, transistors, and integrated circuits, to construct calculating machines. But could we build computational systems using only biological components, such as cultured neurons? The technical answer, is, obviously, of course, as human society is made up of biologically composed computational systems. But as the English expression goes, the devil is in the details. How would you build a biologically based computational system artificially (and not through biological reproduction as animals do), and how would you interface with it?
Moreover, will artificially constructed biological computational systems be important for anything? Or will the future continue to be defined by silicon? Neurons are fast (10-100 m/s) but not as fast as electronic circuits (half the speed of light). Gentle readers, it is the opinion of your correspondent, that yes, we will continue on the route of germanium (0.3 V) or silicon (0.6 V), but biological computing, using artificially constructed biological neural networks, will remain an interesting research curiosity for the near (and maybe even far) future.
This is why a recent neural network paper caught my attention. Is it possible to control a videogame with an artificially constructed biological network (neurons in a dish)?
Neural Network playing Pong with DishBrain
This group, led by Brett Kagan at the start-up (2018) Cortical Labs, based in Australia, seeks to use cultured neural networks as computational systems, and as a proof-of-concept, trained the neural network to play the iconic video game Pong. (Pong is considered the first commercially successful video arcade game, created by Atari by coder Allan Alcorn in 1972. The first video arcade game is considered Computer Space from 1971, developed by Atari co-founders Nolan Bushnell and Ted Dabney.) The team used a system called DishBrain (consisting of multi-electrode arrays embedded in a neuron culture dish) and ran tests on neural networks made up of two types of cells, a human neural cell line and a rodent neuron cell line.
There were two types of electrode arrays, which we can call the “reception array” and one which we can call the “action array.” The reception array described the position of the ball by electrically stimulating different neurons, and the action array was used to move the paddle by measuring the patterns of the spikes (electrical impulses) of the neurons growing on top of the action array. The action array moved the Pong “paddle” to the appropriate position in a demonstration of what the authors entitled “Synthetic Biological Intelligence” (SBI).
There is a term in AI that is called “embodiment” – an intelligence that interacts with the physical external world (your brain and body going for a walk and avoiding obstacles, staying balanced, etc., is an example of this), and the authors argue that their system approaches this through the control of a video game. The memory of how to “play the game” was not maintained over days. Every day a new experiment began, the network needed to be trained again, but the data appears to show their system working and adapting over every experiment.
Reading the paper, it is not clear how robust the system actually is. The work is a difficult read and quite long at 17 pages, and all the figures are highly schematized (always a suspicion – the more abstract the data, the more potentially dubious it is). The killer demo would have been the obvious one, make the system portable, and do public talks, television appearances, etc, with random people playing against the neural network. Or, perhaps, create an online version where the public could interact with the neural culture dish. My main criticism of biological neural network computational research is that live demos are never done, and the systems are always presented in papers as abstracted highly transformed data figures. The ultimate killer demo would have been showing a high school student playing this game against the neural network, the high school student understanding what is going on in the technology, and describing it to his/her fellow students. (Disclosure: The company I co-founded, Backyard Brains, focuses on exactly this, making abstract neurotechnology simple enough that high school students can use and understand it.)
The work is certainly ambitious, and I leave the paper with an optimistic feeling of “I want to believe.” But there needs to be more public demos before I become a full believer.