Depthwise-separable CNN variants and MobileNet-style blocks are common because they reduce multiply cost while preserving useful representational power. For some signal tasks, small residual networks remain competitive when memory allows.

Vision trigger tasks and keyword spotting often have different receptive-field needs and preprocessing assumptions. Treat architecture choice as task-specific rather than forcing a single model family everywhere.

Maintain a small architecture ladder with known budget characteristics and reuse it across projects. Standardized ladders reduce decision noise and speed experimentation.

When false alarms are expensive, prioritize precision and robust negative separation; when misses are costly, optimize recall with calibrated thresholds. Architecture and loss choices should reflect this business tradeoff directly.

Adding layers can improve offline accuracy but may violate latency and energy limits. Keep a strict model-size ladder and promote only when gains justify operational cost.

Benchmark-driven architecture selection can ignore deployment constraints and produce models that cannot pass firmware budgets. Always cross-check architecture choices against real device limits.

Models that cannot observe enough context underperform on events that unfold over time or space. Increasing receptive field must be balanced against compute and memory growth.

Robust front-end features and carefully tuned normalization can reduce the need for larger models. This is often the most efficient way to improve field stability under noisy inputs.

Evaluate architectural variants with shared edge-case suites and threshold stress tests. This reveals stability differences that aggregate metrics hide.

Use input shapes, quantization paths, and preprocessing that mirror deployment settings during training. Late-stage mismatch between training and runtime assumptions is a common source of regressions.

Track calibration behavior, confidence distribution, and class-wise confusion during training. These signals help identify models that look accurate but are unstable after thresholding on-device.

Require architecture promotion decisions to include both quality deltas and operational cost deltas. This prevents accidental complexity creep.

Start with a small architecture shortlist that already fits raw memory and latency limits, then run shared evaluation sets across all candidates. This narrows options quickly without over-investing in one design too early.

Apply compression and threshold tuning only to finalists, then validate under device-level load tests and environmental variance. Promote the smallest model that meets quality and reliability gates.

Publish architecture rationale documents with task assumptions and failure boundaries for future model maintainers. Review it at each release checkpoint so assumptions remain current.

Symptoms include unstable threshold behavior, sensitivity to minor preprocessing drift, and unacceptable tail latency on target hardware. These indicate architecture-task mismatch even when top-line metrics look acceptable.

When misfit is detected, prefer architecture simplification or feature redesign before adding training complexity. Simpler models with robust inputs are often more reliable on constrained devices.

Validate memory fit, latency envelope, calibration stability, and hard-negative behavior for every candidate. Require evidence from on-device tests rather than host-only benchmarks.

Freeze only when one candidate is clearly best across quality and operations gates with measurable margin. Clear freeze criteria reduce churn during compression and integration stages.