Real-Life Electronic Neurons

All the kids down at Stanford are talking about neural nets. Whether this is due to the actual utility of neural nets or because all those kids were born after AI’s last death in the mid-80s is anyone’s guess, but there is one significant drawback to this tiny subset of machine intelligence: it’s a complete abstraction. Nothing called a ‘neural net’ is actually like a nervous system, there are no dendrites or axions and you can’t learn how to do logic by connecting neurons together.

NeruroBytes is not a strange platform for neural nets. It’s physical neurons, rendered in PCBs and Molex connectors. Now, finally, it’s a Kickstarter project, and one of the more exciting educational electronic projects we’ve ever seen.

Regular Hackaday readers should be very familiar with NeuroBytes. It began as a project for the Hackaday Prize all the way back in 2015. There, it was recognized as a finalist for the Best Product, Since then, the team behind NeuroBytes have received an NHS grant, they’re certified Open Source Hardware through OSHWA, and there are now enough NeuroBytes to recreate the connectome of a flatworm. It’s doubtful the team actually has enough patience to recreate the brain of even the simplest organism, but is already an impressive feat.

The highlights of the NeuroBytes Kickstarter include seven different types of neurons for different sensory systems, kits to test the patellar reflex, and what is probably most interesting to the Hackaday crowd, a Braitenberg Vehicle chassis, meant to test the ideas set forth in Valentino Braitenberg’s book, Vehicles: Experiments in Synthetic Psychology. If that book doesn’t sound familiar, BEAM robots probably do; that’s where the idea for BEAM robots came from.

It’s been a long, long journey for [Zach] and the other creators of NeuroBytes to get to this point. It’s great that this project is now finally in the wild, and we can’t wait to see what comes of it. Hopefully a full flatworm connectome.

25 thoughts on “Real-Life Electronic Neurons

  1. Interesting. I read about NeuroBytes on HaD and was wondering since if we think like the brain more as a computer with components or regions that are for specialized (and can even be adapted to other) methods/functions with some hacking or modification… then we can observe reliably memory, memory associations with algorithms and then get into more complex memory + memory + memory repeating maybe like in fractal pattern operations. The neuron reminds me of a fractal branching pattern like a tree or one of the many types of neuronal cells.

    So, in regards to the complex memory associations… how are those patterns that our 5 well known sensory systems (vision, hearing, smell, taste, touch) that can be associated with forces (mechanical pressure, RF, etc.) in different coefficients of effect interact with the world at our boundaries and environment to input and process in our internal systems for output later?

    Like with using controls systems, maybe we can explain in more thorough detail also. Especially if we are thinking lean sigma to standardize the methods. Well, there may be linear mathematical associations where we can interpolate or extrapolate… there may be non-linear associations where we require more advanced algorithms to process the data memories to find the association equation.

    Genetic Algorithms? Neural Networks? Partial Least Squares? Hierarchical Cluster Analysis?

    What are the more complex algorithms to associate the memories and output the valid predicted repeatable observation in an accurate, precise, linear (or non-linear) robust way for the defined range and limits of detection?

    Maybe the exact brain process isn’t goin to be demonstrated to the masses? Or will the process in detail be decoded in the AC and DC ways and means? Like in the book, The Body Electric by Dr. Robert O Becker et.al…. there is a predictable system that performs the controls work like a Process Logic Controller does with large industrial machines.

      1. yup, pretty much. we do what we can to keep the modules as close to $10 ea. as possible, but we need to include accessories and packaging and distributor margin (someday). And profit. We are a for-profit company and need to pay salaries.

        for what it’s worth, our cost driver is connectors — the JST GH (not Molex as suggested above..) are tough to get below $0.17 ea. in multi-reel quantity, and their width ends up defining the PCB size in many cases (with a corresponding cost increase there, too). But the GH has survived many tests where others failed — resistance (over 10x the mfr recommended spec), usability by 10-year-olds, etc.

        I would also suggest that our modules actually aren’t that expensive compared to other modular toys on the market. Our advanced NID kit has 23 distinct boards with 32-bit ARM processors and costs $299. Most hardware companies don’t give you two dozen tiny computers in a box.

        … and finally.. if they’re too expensive for you, please grab our KiCad files and go make your own. the STM32L0 on each module lacks a ground pad, so you can easily hand-solder a bunch of NeuroBytes yourself. I’ll even loan you my crimper if you pick it up in Minneapolis…

        1. You’re based in Minneapolis? Have you ever looked into the Twin Cities Robotics Group. They meet third Thursdays at the Hack Factory in Minneapolis, and sometimes do additional meets outside the Hack Factory. Maybe you can get some word out that way? Can’t hurt.

  2. Interesting toys, but the implementation has it’s limits, still useful for when you are working with people that can’t pick up concepts in a more compressed and abstract way and need play as a form of motivational reward for participating and staying focused, not the average HAD crowd. :-) As for the H part of HAD I guess you could make one of those for a couple of dollars using like a a PIC18F24K40 because it has A/D and a DAC ports? Personally I prefer to play with simulators such as logisim where the hierarchical nature of neuronal circuits is more easily managed, particularly from the scalability side of things. Then once the fundamental ideas have been understood and it is time for a project running in the real world with sensors and actuators you can fit a lot of virtual neurons on a RPi module. As for understanding how the brain works that is like saying you can learn about coastal ecosystems from a few grains of sand.

    1. I agree. This should be viewed as nothing more than an educational toy used to introduce people to the concept of neurons. Anyone who wants to get above a high school level understanding will be using other approaches (such as a PC based simulator) as that is more feasible, cost-effective, and flexible.

      1. yup, that’s absolutely true. NBs aren’t terribly sophisticated; they’re really designed to teach the basics to middle- and high-schools students. concepts like integration, thresholds, decentralized processing, etc. if you want to explore artificial NNs and so forth a computer simulation is the way to go.

Leave a Reply

Please be kind and respectful to help make the comments section excellent. (Comment Policy)

This site uses Akismet to reduce spam. Learn how your comment data is processed.