Abstract

Embedded systems were once built from clear, algorithmic components where each computation was interpretable and debuggable. Modern embedded AI increasingly relies on black-box neural networks: accurate, but hard to explain, debug, or control. We propose Algorithm-Informed Neural Networks (AINNs), where neural architectures are derived from classical algorithmic blueprints and decomposed into neural blocks that mirror logic blocks with explicit invariants. This structure improves data efficiency, interpretability, and enables program-like debugging at the block level.

AINN Framework

Classical algorithms (e.g., knapsack, HMM, DTW+KNN) are decomposed into logic blocks and mapped to neural blocks with invariants and monitors.

Example 1 – Fall Detection (HMM Blueprint)

For wearable fall detection, AINN is constructed from a Hidden Markov Model (HMM) blueprint with separate fall and non-fall chains. Neural blocks learn the observation mapping and emissions, while the HMM forward recursion remains a structured, interpretable layer.

AINN constructed from a two-chain HMM (fall vs. non-fall) with neural observation and emission blocks.

Accuracy vs. number of training samples on UMAFall.

Block-level encoders separate normal vs. anomalous executions in a 2D latent space.

On the UMAFall dataset, AINN achieves 12.6–21.78% higher accuracy than BRNN/MLP baselines while using only 6.58% of the training data and running in milliseconds on a Google Pixel 6a.

Example 2 – Keyword Spotting (DTW + KNN Blueprint)

For keyword spotting, AINN mirrors a DTW+KNN pipeline: a Cost Estimator approximates frame-wise distances, a Dynamic Programmer learns soft DTW recursion, and a Score Aggregator implements differentiable KNN-style voting over reference utterances.

AINN blocks for DTW-based sequence alignment and KNN-based decision rules.

Accuracy vs. training samples on Google Speech Commands.

Relationship between learned DP updates and ground-truth minima.

Block-level anomaly clusters from Cost and Aggregation encoders.

On the Google Speech Commands dataset, AINN achieves up to 7.75% higher accuracy using half the training data and a 46× smaller model, while preserving explicit alignment structure and enabling anomaly detection inside the dynamic programming block.

BibTeX