Very low power data acquisition is possible using an Analog Neural Network to analyse the Rate of Innovation of a signal.

SiliconInterevntion did not invent the concept of “Rate of Innovation”,  it is known in the signal processing world. You may start your exploration of the idea with “Sampling Signals with Finite Rate of Innovation, Martin Vetterli et al, IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 6, JUNE 2002. DOI: 10.1109/TSP.2002.1003065. The concept is well defined, but perhaps not so well known.

Examples explaining Finite Rate of Innovation

AM Radio

Some of the team at SiliconIntervention, myself included, grew up in the UK listening to BBC Radio 1. It was an AM RF carrier at 1214 kHz and about 10kHz audio bandwidth. If I were to describe by phone how to set up a replica Radio 1 system in your lab it would begin with “tune your oscillator to 1214Khz and set the amplitude to 10mV”.

Having done that, if our conversation got interrupted (our phone call got dropped) your oscillator would continue to output 1214Khz 10mV until you called me back. It could be an hour later before we got in touch again, but your 1214Khz is still running.

The Rate of Innovation of the signal at this point is zero – I don’t need to say anymore, the 1214khz is running producing a steady signal. But that’s no use – the amplitude must vary for you to hear The Rolling Stones or whatever.

So, here come the instructions: “adjust the amplitude to 12mV, now to 14mV, now to 9mV, now to 5mV …”  These are a series of instructions for you to keep adjusting the amplitude, I am constructing them such that amplitude waveform is the audio of a Rolling Stones song.

At what rate do I need to send you those instructions to change the amplitude?

20k instructions per second is sufficient – since we said we have an audio bandwidth of 10kHz. The Rate of Innovation of your oscillator is now 20kHz. Thus, a 1214Khz oscillator is running in your lab and making an AM radio, but the rate of updates – the rate of data flowing to you – is not 2.418Mhz but the Rate of Innovation of the signal, 20khz, not 2.4Mhz. We need only communicate at the rate of innovation.

Note the surprising thing: if you had an oscilloscope set to 50nS per division connected to your lab oscillator you would see the radio signal as it was transmitted – at high speed. You can reconstruct the high bandwidth signal from the low Rate of Innovation data that you are being sent. Transmitting at the rate of innovation does not lose any data, the high frequency signal can be exactly reproduced.

While this example is understandable the inclination is to conclude that it is working because it’s an audio signal impressed upon a high frequency carrier. But the principle applies to all signals that have a Finite Rate of Innovation, and the concept works in situations much more sophisticated than the AM Radio case,  so let’s move on to another example.

Example in an ECG instrument

Consider an ECG machine that picks up the signals from the beating heart. The doctor wants to see the details of the signal, normally labelled PQRST, the characteristic signal from the beat of the heart. The analog neural network (ANN) at the point of data acquisition determines the Rate of Innovation of the signal and sends sufficient data to the application processor to reproduce the PQRST waveform. This assessment of the rate of innovation is continuous – it is not a one-time event.

When the ECG changes, goes into fibrillation for example, the ANN instantly adapts such that the reconstruction in the application processor is precise and this fibrillation is visible.

ECG instruments may sample data in an ADC at 1Khz, every milli-second. But the rate of innovation of medical data is much lower, perhaps a few hertz at most.

The ANN’s ability to constantly track and communicate at the Rate of Innovation has reduced communication bandwidth by a factor of 500. This vastly reduce power. As in the AM radio, the application processor can reproduce the exact series of samples at 1kHz – the process is lossless.