Seven Open Problems in Edge Intelligence
This is not a survey. It's a working document — a list of problems we encounter in shipping real systems that we don't yet have good answers for.
1. Continual Learning Without Catastrophic Forgetting at the Edge
Cloud models retrain on new data constantly. Edge models cannot afford full retraining cycles. Continual learning methods (EWC, PackNet, etc.) help but break down when the task distribution shifts rapidly. We don't have a production-grade solution for this.
2. Uncertainty Quantification Under 10ms
Knowing what a model doesn't know is as important as knowing what it knows. Bayesian methods are too slow. MC Dropout is fast but poorly calibrated on distribution shift. Conformal prediction is promising but has issues with non-exchangeable data.
3. Secure Aggregation in Heterogeneous Meshes
Federated learning assumes all nodes contribute honest gradients. In adversarial mesh deployments, this assumption fails. Existing Byzantine-robust aggregation methods add 40–200% overhead.
4. Efficient Attention for Long Sensor Sequences
Already discussed in our earlier note. We have a partial answer. It's not complete.
5. The Calibration-Accuracy Tradeoff at Low Precision
Models quantized to INT4 are typically overconfident. The relationship between precision and calibration is poorly understood.
6. Cross-Modal Representation Alignment Without Paired Data
Fusing gaze, gesture, and voice requires a shared embedding space. Collecting paired training data is expensive. Self-supervised alignment across modalities is an open research problem.
7. Power-Aware Neural Architecture Search
NAS has largely optimized for accuracy × latency. Power consumption on real hardware (with memory, IO, and cooling effects) is much harder to model. Our power models are off by 15–40%.
---
If you are working on any of these, we would like to talk.