Liquid Neural Networks Do More With Less

[Ramin Hasani] and colleague [Mathias Lechner] have been working with a new type of Artificial Neural Network called Liquid Neural Networks, and presented some of the exciting results at a recent TEDxMIT.

Liquid neural networks are inspired by biological neurons to implement algorithms that remain adaptable even after training. [Hasani] demonstrates a machine vision system that steers a car to perform lane keeping with the use of a liquid neural network. The system performs quite well using only 19 neurons, which is profoundly fewer than the typically large model intelligence systems we’ve come to expect. Furthermore, an attention map helps us visualize that the system seems to attend to particular aspects of the visual field quite similar to a human driver’s behavior.

 

Mathias Lechner and Ramin Hasani
[Mathias Lechner] and [Ramin Hasani]
The typical scaling law of neural networks suggests that accuracy is improved with larger models, which is to say, more neurons. Liquid neural networks may break this law to show that scale is not the whole story. A smaller model can be computed more efficiently. Also, a compact model can improve accountability since decision activity is more readily located within the network. Surprisingly though, liquid neural network performance can also improve generalization, robustness, and fairness.

A liquid neural network can implement synaptic weights using nonlinear probabilities instead of simple scalar values. The synaptic connections and response times can adapt based on sensory inputs to more flexibly react to perturbations in the natural environment.

We should probably expect to see the operational gap between biological neural networks and artificial neural networks continue to close and blur. We’ve previously presented on wetware examples of building neural networks with actual neurons and ever advancing brain-computer interfaces.

Continue reading “Liquid Neural Networks Do More With Less”

High Voltage Power Supply From USB

Those who work in different spaces may have different definitions of the term “high voltage”. For someone working on the GPIO pins of a Raspberry Pi it might be as little as 5 volts, someone working on a Tesla coil might consider that to be around 20 kV, and an electrical line worker might not reference something as HV until 115 kV. What we could perhaps all agree on, though, is that getting 300 volts out of a USB power supply is certainly a “high voltage” we wouldn’t normally expect to see in that kind of context, but [Aylo6061] needed just such a power supply and was eventually able to create one.

In this case, the high voltages will eventually be used for electrophoresis or electrowetting. But before getting there, [Aylo6061] has built one of the safest looking circuits we’ve seen in recent memory. Every high voltage part is hidden behind double insulation, and there is complete isolation between the high and low voltage sides thanks to a flyback converter. This has the benefit of a floating ground which reduces the risk of accidental shock. This does cause some challenges though, as voltage sensing on the high side is difficult while maintaining isolation, so some clever tricks were implemented to maintain the correct target output voltage.

The control circuitry is based around an RP2040 chip and is impressive in its own right, with USB isolation for the data lines as well. Additionally the project code can be found at its GitHub page. Thanks to a part shortage, [Aylo6061] dedicated an entire core of the microprocessor to decoding digital data from the high voltage sensor circuitry. For something with a little less refinement, less safety, and a much higher voltage output, though, take a look at this power supply which tops its output voltage around 30 kV.

Hackaday Links Column Banner

Hackaday Links: April 30, 2023

Cloudy with a chance of concrete? The “success” of last week’s brief but eventful Starship launch has apparently raised some regulatory eyebrows, with the Federal Aviation Administration launching an investigation into the destruction wrought by the mighty rocket. And it’s not just the hapless Dodge Caravan that they’re concerned with — although we found some fantastic POV footage that shows the kill shot as well as close-ups of the results — but also the damage rained down upon residents around the Boca Chica launch complex. Tons of concrete and rebar were excavated by the 33 Raptor engines during the launch and sent in all directions, reportedly landing up to 6 miles (10 kilometers) from the pad. What’s worse, a lot of debris ended up on beaches that are home to endangered species, which has the Sierra Club also taking an interest. The FAA has apparently nixed any launches from the Texas facility until they complete their investigation.

Continue reading “Hackaday Links: April 30, 2023”

A Microneedle Vaccine Patch Printer For Thermostable MRNA Vaccines

What if you could get vaccinated with the ease of putting on an adhesive bandage? This is the promise of microneedle patches (MNP), which are essentially what they sound like. These would also have uses in diagnostics that might one day obliviate the need for drawing blood. The one major issue with MNPs is their manufacturing, which has been a laborious and highly manual process. In a recent paper in Nature Biotechnology researchers detail the construction and testing of a MNP printer, or microneedle vaccine printer (MVP) that can print dissolving polymers containing stabilized mRNA vaccine.

These mRNA strands are as usual encapsulated in a liquid nanoparticle container, which is mixed with the soluble and biocompatible polymer. This mixture is then added to a mold and dried, after which it retains the microneedle structure of the mold. On tests involving pig skin, the MNPs were capable of penetrating the skin and delivering the vaccine contained in the needles. Produced patches were shown to be shelf-stable for at least six months, which would make these ideal for vaccine distribution in areas where refrigeration and similar are problematic.

Using MNPs for delivering vaccines has previously been researched for e.g. delivering rotavirus and poliovirus vaccine, and a 2021 study in Nature Biomedical Engineering looked at the viability of using MNPs to rapidly sample protein biomarkers in interstitial fluid, which could make diagnostics for certain biomarkers as uncomplicated as putting on the patch, removing it and examining it, removing the need for drawing blood or sampling large amounts of interstitial fluid for external analysis.

If the concept of the MVP and similar MNP printers can be commercialized, it might make it possible to strongly shorten the supply chain for vaccines in less developed regions, while also enabling diagnostics that are very costly and cumbersome today.

Embed Hardware Into 3D Prints, But Not In The Way You’re Thinking

[Christopher Helmke] is doing fantastic work in DIY systems for handling small hardware like fasteners, and that includes robotic placement of hardware into 3D prints. Usually this means dropping nuts into parts in mid-print so that the hardware is captive, but that’s not really the story here.

The really inventive part we want to highlight is the concept of reducing packaging and labor. Instead of including a zip-lock bag of a few bolts, how about embedding the bolts into a void in the 3D print, covered with a little snip-out retainer? Skip ahead to 1:54 in the video to see exactly what we mean. It’s a pretty compelling concept that we hope sparks a few ideas in others.

As clever as that concept is, the rest of the video is also worth a watch because [Christopher] shows off a DIY system that sits on top of his 3D printer and takes care of robotically placing the hardware in mid-print. He talks all about the challenges of such a system. It’s not perfect (yet), but seeing it in action is very cool.

We’ve recently seen a lot of fascinating stuff when it comes to [Christopher Helmke]’s automated handling of fasteners and similar hardware. His system makes rapid and accurate dispensing of bolts look easy, and his work on using compressed air to zip pieces around seems effective.

Continue reading “Embed Hardware Into 3D Prints, But Not In The Way You’re Thinking”

Add A Little Quindar To Your Comms For That Apollo-Era Sound

If there’s one thing that ties together all the media coming out of the Apollo era, it’s probably the iconic Quindar tones. These quarter-second beeps served as control tones for the globe-spanning communications network needed to talk to the Apollo astronauts, and any attempt to recreate the Apollo-era sound would be glaringly wrong without them. And that’s why [CuriousMarc] whipped up this Quindar tone system.

The video below starts with a detailed treatment of what Quindar tones are and why they were used, a topic we’ve covered ourselves in the past. To recap, Quindar tones are a form of in-band signaling, with a 2,525-Hz pure sine wave intro tone that signaled the transmitters connected to Mission Control in Houston over leased telephone lines to key up. The 2,475-Hz outro tone turned off the transmitters and connected the line to the receivers.

To recreate the sound quality of the original circuitry, and to keep in the retro vibe, [Marc]’s Quindar homage avoided digital circuitry as much as possible, opting instead to generate the two tones with an XR-2206 function generator chip. The chip can rapidly switch back and forth between two frequencies, making it perfect for FSK applications or, in this case, reproducing the two slightly different tones. [Marc] added a dual mono-stable multi-vibrator to pulse the tone, giving the 250-ms pulse, and an audio gate, which uses a MOSFET to switch the tone into an audio stream. All this got soldered up to a piece of perf board and stuffed in the base of a cheap intercom microphone, which while not period accurate still has a cool retro look — and now, a retro sound, too.

Hats off to [CuriousMarc] and his merry band for probing the mysteries of Apollo-era comms and keeping the accomplishments of all those engineers alive. The methods they used are still relevant after all these years, and there seems to be no end to what we can learn from them.

Continue reading “Add A Little Quindar To Your Comms For That Apollo-Era Sound”

When Your Smart Light Switches Stop Working, Build Your Own

If you want smart light switches in your house, you can buy from any one of hundreds of manufacturers. [Brian Boyle] had kitted out his home with TP Link devices, but after a few years of use, he found they all suddenly failed within a few months of each other. Decrying the state of things, he set about building his own instead.

[Brian]’s switches use the ESP32 for its handy in-built WiFi hardware. His aim was to produce smart switches that would fit neatly into standard “Decor” style switch boxes. The design uses two PCBs. One is charged with handling the mains power side of things. It carries an SPDT relay for switching AC power, and a DC power supply to run the ESP32 itself. The controller board holds the microcontroller, a Neopixel as a status indicator, and a pair of buttons — one for switching the lights on and off, the other for resetting to default settings. The physical housing is 3D printed, and looks great with the glowing status indicator in the middle of the switch.

[Brian]’s switches are triggerable via MQTT, a web interface, and the physical button onboard the device itself. Having built the devices on his own, he’ll be well-placed to troubleshoot any usability or reliability issues that crop up in the future. That’s a lot more than we can say about most smart devices on the market!