Deploying deep vision models on edge hardware requires understanding how architecture and training hyperparameters jointly determine accuracy and inference latency, yet these relationships remain poorly characterized in a systematic, data-driven manner. This paper presents a two-stage statistical framework providing interpretable, closed-form insights into both. In the first stage, we apply distance correlation (dCor) and the maximal information coefficient (MIC) across seven image-classification datasets, revealing that batch size and total layer count are the strongest universal accuracy predictors (mean dCor: 0.228 and 0.174), while learning rate achieves the highest MIC (0.226), reflecting a non-monotonic relationship with accuracy. In the second stage, PySR symbolic regression (representing, to our knowledge, the first application to cross-dataset vision model accuracy prediction) derives compact, interpretable formulas. Dataset-specific models achieve from 0.20 to 0.45; a universal model achieves a mean leave-one-dataset-out of 0.23, remaining strictly positive on all held-out datasets, whereas ordinary linear regression collapses to . We further derive device-specific inference latency formulas for CPU, GPU, and NPU, outperforming classical baselines by – in and confirming fundamental device heterogeneity. Together, these results offer interpretable surrogate models for screening deep vision architectures under accuracy and latency constraints in edge deployment.
On-device latency ultimately determines whether a vision model is deployable on smartphones, yet collecting real-device latency at scale is costly and fragile. We present AI on the Edge, an automated measurement pipeline and open dataset for large-scale delegate-resolved Android benchmarking. AI on the Edge converts a PyTorch model zoo to TensorFlow Lite in FP32 and full-integer INT8 precision, then benchmarks each model on physical phones with three forced execution paths: CPU (XNNPACK), GPU (TFLite GPU), and NPU (NNAPI), so every latency is attributable to a known backend. Using one workstation and five commodity phones, AI on the Edge measured 586 architectures and produced 5,826 model × device × precision records, comprising about 17 500 backend benchmarks and 351 600 timed inferences (586 × 5 × 3 × 2 × 20), in roughly 56 unattended device-hours (up to 201 models/h/phone). Campaigns include both end-to-end Stage A+Stage B runs and Stage B-only replay from prebuilt TFLite artifacts, with identical on-device benchmark settings. The pipeline is designed for unattended operation via USB-reconnection handling, thermal cool-down scheduling, and a resumable processing ledger. We release the full per-run dataset with documented schema, public model/code resources, and a reproducibility harness that regenerates headline statistics from raw records. AI on the Edge reduces the practical cost of real-device benchmarking and provides a reusable substrate for hardware-aware architecture research.