A per-camera vision pipeline with quality gates ran a full fleet training cycle for about seven cents of compute
AspireAI community · AWS Machine Learning Blog (Kiowa Jackson)
In a production inspection deployment, one pooled model across many cameras consistently underperforms individual cameras enough that each camera ends up needing its own fine-tuned model. This AWS post walks that pattern end to end with the SageMaker Python SDK v3, replacing the deprecated SageMaker Edge Manager with SageMaker Neo. A dual-branch pipeline trains an anomaly detector and a defect classifier per camera, exports each to ONNX, evaluates the exported artifact against held-out data, and blocks underperforming models at a quality gate before compiling with Neo and registering the result.
The orchestration layer is what makes fleet scale workable. Re-training requests land on an SQS FIFO queue, a Lambda function starts a Step Functions execution per request, and each execution acquires a slot through a DynamoDB semaphore, so a burst drains at a steady rate instead of launching everything at once. The default admits at most five concurrent pipeline executions, which keeps training inside standard account quotas. A full pipeline run on synthetic data costs about $0.07 in compute, and the complete sample ships in the aws-samples GitHub repository.
The limitation is scope: this is a vendor reference architecture rather than an independent evaluation, and the $0.07 figure is a synthetic-data run, not a production cost. The author also notes that the per-camera pattern exists only because a single pooled model underperforms in the first place.
https://aws.amazon.com/blogs/machine-learning/fleet-wide-computer-vision-lifecycle-at-the-edge-with-sagemaker-sdk-v3
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.