Probabilistic Sensing: Intelligence in Data Sampling

TL;DR

The paper addresses the challenge of excessive data generation by sensors by introducing probabilistic sensing with p-neurons that decide sampling in real time based on analog feature extraction. The approach enables reflex-like, autonomous activation of data acquisition, combining spintronic p-bits and FPGA-based implementations to control sampling probability through the p-neuron and a clock signal $V_{Sync}$. It demonstrates lossless probabilistic data acquisition on active seismic data, achieving $NMSE = 0.41\,6\%$ in the $0$–$200\,6\mathrm{Hz}$ range and a $93\,6\%$ reduction in samples and ADC active time, with microsecond response times ($\approx 2.8\,6\mathrm{\mu s}$) that break the sub-sampling-rate limit. This work highlights a path to energy-efficient, real-time intelligent sensing applicable to geophysical surveys and beyond.

Abstract

Extending the intelligence of sensors to the data-acquisition process - deciding whether to sample or not - can result in transformative energy-efficiency gains. However, making such a decision in a deterministic manner involves risk of losing information. Here we present a sensing paradigm that enables making such a decision in a probabilistic manner. The paradigm takes inspiration from the autonomous nervous system and employs a probabilistic neuron (p-neuron) driven by an analog feature extraction circuit. The response time of the system is on the order of microseconds, over-coming the sub-sampling-rate response time limit and enabling real-time intelligent autonomous activation of data-sampling. Validation experiments on active seismic survey data demonstrate lossless probabilistic data acquisition, with a normalized mean squared error of 0.41%, and 93% saving in the active operation time of the system and the number of generated samples.

Figures (4)

• Figure 1: Probabilistic sensing using p-neurons. (a) The neuro-inspiration behind probabilistic sensing. (b) Conventional event-based sensing, where event features are extracted from the sensor signal and then these features are encoded into spikes or spike trains. (c) Probabilistic sensing, where the extracted event features from the sensor signal are used to activate a probabilistic neuron (p-neuron), which in turn activates data-acquisition, capturing the actual sensor signal and retaining the original data, not just a few features of it. The inset shows a spintronic p-neuron implemented using a probabilistic bit (p-bit).
• Figure 2: Overall probabilistic sensing system design, including the (a) feature extraction, (b) activation, and (c) data acquisition units. The p-neuron is in the activation unit, shown here with two different implementations; digital (top) and spintronic (bottom). The entropy source in the digital implementation is a Linear Feedback Shift Register (LFSR) and in the spintronic implementation is a stochastic MTJ. Magnetization dynamics of the sMTJ and its impact on probabilistic sampling. (d) A stochastic MTJ (sMTJ) with a low-barrier (LB) free layer magnet. Magnetization dynamics of $m_z$ for (e) a circular in-plane magnetic anisotropy (IMA) magnet, and (f)-(g) an isotropic magnet. Sensor signal sampling with the sMTJ retention time (h) longer than the sampling time and (i) comparable to the sampling time. A seismic geophone senosor signal and the corresponding event-detection signal obtain from the activation unit ($V_{Clk}$), where in (j) the average random sampling rate (X) is smaller than its value in (k).
• Figure 3: Experimental implementation of the probabilistic sensing system, (a) using a field programmable gate array (FPGA) and discrete electronic components. (b) The design of the digital p-neuron implemented using the FPGA. Experimental results showing the average random sampling rate plotted against (c) the input signal to the p-neuron ($V_{IN}$) and (d) the slope of the original sensor signal.
• Figure 4: Survey-level validation of the probabilistic sensing system using active seismic geophone sensor data. (a) schematic illustration on an active seismic survey. (b) Survey-level results based on both implementations; digital FPGA-based (experiment) and spintronic sMTJ-based (simulation). Comparison of reconstructed seismic signals using linear interpolation after sampling using a regular ADC (green) and a probabilistic ADC (blue), (c) in the time domain and (e) in the frequency domain (frequency range 0 - 200 Hz). (d) Comparisson of the normalized mean square error (NMSE) between data acquired using an R-ADC and a P-ADC, for both the FPGA-based experiments and the sMTJ-based simulations. (f) Number of generated samples using the P-ADC in percentage relative to regular continuous sampling at a 2 kHz sampling frequency using an R-ADC.