Adaptive Compressive Sensing Approach

• Adaptive compressive sensing is a dynamic technique that adjusts measurement matrices based on signal feedback to optimize sparse signal recovery.
• The methodology employs probabilistic SOT-MRAM crossbar architectures and iterative update logic to reduce energy consumption and enhance reconstruction accuracy.
• Empirical results demonstrate up to 5 dB TNMSE improvement and highlight challenges like crossbar parasitics, driving innovations in adaptive sensing techniques.

Adaptive compressive sensing (ACS) refers to a class of compressive sensing algorithms and hardware frameworks that dynamically modify their measurement, sampling, or reconstruction strategy based on properties of the signal, the environment, or partial reconstruction feedback. By exploiting adaptivity, ACS seeks to improve accuracy, reduce energy and hardware costs, and provide robustness across heterogeneous application contexts, ranging from signal acquisition in fluctuating environments to model-based signal compression in constrained hardware.

1. Mathematical Formulation and Adaptive Measurement Design

Adaptive compressive sensing retains the canonical CS model for a $k$-sparse signal $x \in \mathbb{R}^n$,

$y = \Phi x + e,$

where $y \in \mathbb{R}^m$ are $m \ll n$ compressed measurements, $\Phi \in \mathbb{R}^{m \times n}$ is the measurement matrix, and $e$ models measurement noise. Classical CS assumes $\Phi$ is fixed and random (e.g., Gaussian, Bernoulli), with reconstruction via $$x̂ = \arg\min \lVert x \rVert_1 \text{ s.t. } \lVert y - \Phi x \rVert_2 \leq \epsilon$$.

In the adaptive variant, the design or selection of $\Phi$ evolves during sampling. For non-uniform or region-of-interest (RoI) scenarios, adaptivity is realized by a weighting vector $x \in \mathbb{R}^n$0 that biases measurement design, e.g.,

$x \in \mathbb{R}^n$1

with $x \in \mathbb{R}^n$2 updated in response to estimated per-coefficient errors, such as

$x \in \mathbb{R}^n$3

where $x \in \mathbb{R}^n$4 is a moving average and $x \in \mathbb{R}^n$5 tunes sensitivity. This dynamic allocation allows for intensified measurement in high-uncertainty regions (Salehi et al., 2019).

Update-frequency control is also employed: an update period $x \in \mathbb{R}^n$6 (often based on an energy budget $x \in \mathbb{R}^n$7 and threshold $x \in \mathbb{R}^n$8) regulates how often the measurement matrix is refreshed, balancing error reduction and write energy consumption.

2. Enabling Hardware Architectures

Practical adaptive CS requires in-memory processing capabilities and tunable, low-power hardware. Multibit Spin-Orbit Torque Magnetic Random Access Memory (SOT-MRAM) crossbar arrays instantiate $x \in \mathbb{R}^n$9 in analog conductance form (Salehi et al., 2019):

• Each SOT-MRAM multibit cell comprises four parallel magnetic tunnel junctions, yielding $y = \Phi x + e,$0 programmable resistance levels.
• Measurement rows are probabilistically programmed via trains of current pulses, physically realizing targeted Bernoulli or Gaussian statistics.
• An $y = \Phi x + e,$1 SOT-MRAM array encodes $y = \Phi x + e,$2; input voltages represent $y = \Phi x + e,$3 and output currents sum to $y = \Phi x + e,$4.
• Peripheral circuitry supports analog-to-digital conversion, row/column selection, sense amplification, and winner-take-all logic for iterative solvers.
• Power gating for unused array subregions and instant-on memory enable zero static leakage and context-aware operation.

SPICE sub-block modeling demonstrates efficient switching and readout, crucial for system-level performance evaluation.

3. Algorithmic Workflow and Adaptive Update Logic

Adaptive crossbar-based approaches such as ACMCA follow a multi-stage workflow (Salehi et al., 2019):

1. Initialize $y = \Phi x + e,$5 via a deterministic reset.
2. At each iteration $y = \Phi x + e,$6, optionally re-program rows of $y = \Phi x + e,$7 by probabilistic switching according to current $y = \Phi x + e,$8.
3. Acquire measurement $y = \Phi x + e,$9.
4. Reconstruct $y \in \mathbb{R}^m$0 using basis pursuit or Bayesian solvers.
5. Evaluate per-coefficient error $y \in \mathbb{R}^m$1 and update $y \in \mathbb{R}^m$2.
6. Adjust the global update-frequency variable $y \in \mathbb{R}^m$3.
7. Iterate, with frequency of $y \in \mathbb{R}^m$4 reconfiguration governed by the reconstructed error and energy budget.

This tight loop between measurement, feedback-driven allocation, and hardware programming achieves fine-grained adaptation at runtime.

4. Performance Metrics and Empirical Results

Key quantitative metrics include:

• Time-Averaged Normalized Mean Squared Error (TNMSE):

$y \in \mathbb{R}^m$5

• Energy and Area:
  • Area reduction up to $y \in \mathbb{R}^m$6 in $y \in \mathbb{R}^m$7 nm processes vs. prior random-number generator (TRNG) + CS-generator CMOS designs.
  • Crossbar vector-matrix multiplication (VMM) energy per $y \in \mathbb{R}^m$8 array: SMC crossbar $y \in \mathbb{R}^m$9 nJ vs. CMOS $m \ll n$0 nJ ($m \ll n$1 improvement).
  • Matrix generation/write energy: $m \ll n$2 nJ, with infrequent updates making this negligible in the total budget.
• Signal Reconstruction:
  • Region-of-interest (RoI) TNMSE improved by $m \ll n$3 dB (Bernoulli $m \ll n$4) to $m \ll n$5 dB (Gaussian $m \ll n$6) relative to uniform CS.
  • Average full-signal TNMSE improved by $m \ll n$7 dB.
  • Comparable or better energy efficiency and area than prior adaptive/non-adaptive hardware.
• Trade-offs:
  • Adjustable trade-off between reconstruction error and write energy via $m \ll n$8.
  • Crossbar non-idealities (IR drop, device mismatch) can degrade reconstruction but are actively mitigated by dynamic adaptive updates.

Empirical validation uses MATLAB algorithm simulations with hardware-in-the-loop SPICE modeling, substantiating both algorithmic improvements and hardware resource savings (Salehi et al., 2019).

5. Comparative Analysis and Context

Relative to uniform CS hardware (e.g., flash ADCs, CMOS-based pseudo-random generators), adaptive crossbar-based approaches:

• Yield $m \ll n$9 dB lower TNMSE and greater hardware/energy efficiency.
• Exploit the intrinsic stochasticity of SMCs, obviating the need for dedicated random-number circuits.
• Integrate adaptivity, which is largely absent or implemented with higher cost in prior non-adaptive spintronic CS proposals.
• Offer rapid, zero-leakage power gating and in-memory vector-matrix multiplication not feasible in conventional mixed-signal/CMOS-only designs.

However, crossbar-based ACS faces scalability issues with large array parasitics, motivating future dynamic subarray selection and optimized materials to eliminate the need for external field switching.

6. Limitations and Future Directions

Outstanding challenges and development opportunities include:

• Mitigating line-to-line parasitics and voltage drop in large crossbars, possibly via dynamic sub-array selection or architectural innovations (Salehi et al., 2019).
• Extending update-frequency control using multi-objective optimization (e.g., via genetic algorithms or convex programming).
• Pursuing field-free SOT-MRAM switching materials to simplify integration.
• Integration of on-chip sparse recovery/iterative-solver accelerators to enable a fully in-hardware adaptive compressed sensing pipeline.
• Exploring hardware-aware algorithmic co-design for further improvements in energy, latency, and reconstruction quality.

In sum, the Adaptive Compressive Sensing Approach, as exemplified by the ACMCA framework, synthesizes feedback-driven adaptive measurement design, device-level probabilistic switching in SOT-MRAM crossbars, and high-efficiency in-memory computation. This yields robust, resource-minimal sparse signal recovery for rapidly-varying, constraint-driven sensing environments (Salehi et al., 2019).