Neuromorphic object localization using resistive memory, ultrasonic transducers | Hacker News Reader

Neuromorphic object localization using resistive memory, ultrasonic transducers

nature.com

47 pointsrntn9 comments

I'm dumb. What exactly does this mean - memory enhancement, memory augmentation, just better mimicry, something else?

Where are you seeing "memory enhancement" or "memory augmentation"?

This article is about analog computation hardware, which can potentially solve computer vision tasks with far less power than typical hardware with digital processing.

I saw stuff about synaptic gain.

Where? That's not in the article.

This is not talking about actual brains or synapses, if that's what you're thinking. It's just using the same words but these systems are only "inspired" by actual neurons, they don't interact with neurons and don't really behave like neurons except in high-level ways.

Like I said, I'm dumb.

I will take a stab at this... it is about object localization via sound instead of sight, using very little power. For comparison, let's say you have a sound sensing setup on a Raspberry Pi, known for its low power draw, attached to a pair of microphones, looking to alert the location of an object by similar sound triangulations. That processing could take maybe 4-6W of power (continuous running of the Pi and attached mics, as a generous estimate), and could be quite effective.

However the technique in this paper is _ultra_ low power. First off, they model the design off of a barn owl, and using "neuromorphic memristors" (sounds technobabble to me but I didn't understand that part). But in the Results part of the paper they claim they can sense movement sampling every 1/10th of a second using only 250 microwatts of power, orders of magnitude more efficient than a naive approach, with only 22 floating-point calcluations per sample. Sounds quite impressive, but I wonder what the actual applications of this will be, even though I'd love to be able to track mice around my house in real time grrrr.

"neuromorphic memristors" let you build neural-nets in hardware. Memristor act as a sort of variable resistor based on how much current has flowed through them. The weights in neural networks can be stored as the memory effect in each memristor.

A few years back HP was researching memristors to produce neural net processors, but I never heard of anything coming from it.

It's a pretty clever way to deploy a neural net algorithm using very low power. Maybe HP was looking at the wrong market.

It is not clear what the role of the RRAM is. I get that it applies weights, but where do the weights come from? Is this a system that you train with lots of known stimulus and location samples? And having trained one, you can read it out and program those weights into another?

Here's a TL;DR (I'm an amateur with interest in neuroscience - I have no creds and this is my personal, non-specific interpretation of the important parts - please feel free to correct me on anything I may have misinterpreted or how this explanation could be improved): There's a circuit that's fed, in an event-based manner, analog, ultrasonic data from sensors (data is filtered in a few ways first). The circuit has resistive memory which is basically non-volatile RAM. The RRAM modules can store weights, and block/allow inputs depending on its mode. The circuitry also involves LIF (Leaky integrate and fire) neurons which translate the amplitude of voltage into frequency (higher amplitude -> higher frequency @ same amplitude). Processed waves from the ultrasonic sensors are interpreted as degrees away and distance (with a relatively small error margin, on quite low energy compared to the same task on a microprocessor).

Reply on news.ycombinator.com