Source linked

Analog Kolmogorov-Arnold Networks Cut Area 55%, Power 50% for Wearables

arxiv.org@systems_wire3 hours ago·Systems Engineering·2 comments

A hardware-software co-optimized analog neural network achieves up to 55% area reduction and 50% power savings for complex function approximation on flexible electronics, with pruning actually improving accuracy.

kolmogorov arnold networksflexible electronicsanalog computinglow powerhardware software co optimizationiot sensors

Up to 55% less chip area and 50% less power for wearable sensors—that's what Analog Kolmogorov-Arnold Networks (AKANs) deliver, and the pruning technique actually makes them more accurate, not less.

Why Analog Function Approximation Matters for Wearables

Flexible electronics need on-sensor processing for biosignals, calibration curves, and log/power operations. Those functions are transcendental and multivariate—expensive to implement digitally. Analog approximation skips the energy-hungry analog-to-digital conversion, but flexible electronics impose brutal constraints on circuit density and power. The authors from an unnamed institution tackle this by co-optimizing both the hardware topology and the software training process, incorporating circuit-level error models into the loss function.

How AKANs Co-Optimize Hardware and Software

Kolmogorov-Arnold Networks naturally decompose multivariate functions into sums of univariate splines—a structure suited to analog implementation. The team prunes network weights at both the software level (during training) and the hardware level (by removing unused spline nodes). That pruning is where the real trick lies: it reduces area and power while simultaneously regularizing the spline parameters, preventing overfitting. Across multiple benchmarks the average area savings hit nearly 30%, and power drops by a similar margin.

Pruning That Improves Accuracy – A Surprising Result

You'd expect aggressive pruning to degrade accuracy. Instead, the regularization effect of forcing sparse spline representations cleans up the approximation. The paper reports that the pruning methodology can “improve approximation accuracy by regularizing spline parameters.” So you get smaller, cheaper circuits that also predict better. That’s a rare win-win in hardware acceleration.

AKANs give flexible electronics a practical path to on-sensor intelligence without the digital conversion overhead, making always-on biosignal processing a realistic near-term bet.

Source: Co-Optimization of Analog Kolmogorov-Arnold Networks for Low-Power Function Approximation in Flexible Electronics
Domain: arxiv.org

Read original source ->

External source stays available while the OJO article and comment thread stay local.

More in Systems Engineering

view topic
Random Forest Surrogate Predicts HPC Performance from Just 54 Benchmarks

A cold-start random forest trained on 54 LAMMPS+SPICA runs ranks hybrid MPI+OpenMP configurations with 4.0% relative error - but only within the same hardware regime.

97-99% Accuracy Retained While Slashing TPOT: Cascaded LLM Serving That Routes Smart

A two-stage cascaded framework keeps 97-99% of the strongest model's accuracy but directs only hard queries to expensive models-an interpretable hyperparameter controls the cost-accuracy knob.

DiStash Puts DRAM, SSD, and NVM Under One Transactional Roof

DiStash, a disaggregated transactional key-value store built on FoundationDB, runs a single atomic transaction across DRAM, SSD, HDD, and NVM pools-eliminating race conditions and tiering complexity.

Host-Driven SRv6 Flowlet Balancing Cuts Tail Latency 33% vs ECMP

Shifting flowlet detection from switches to hosts and using SRv6 for stateless steering reduces tail latency by 33% over ECMP and 15% over random flowlet balancing in testbed experiments.

Comments load interactively on the live page.