When Bigger is Worse: A Practitioner's Guide to Model Selection Under Data Scarcity

Pixel-wise Earth-observation (EO) foundation models are now achieving state-of-the-art performance via generated spatial embeddings. However, how these models scale and how best to spend a pretraining budget remain poorly understood. We present the largest controlled scaling study for EO to date: 395 training runs on 1,024 GH200 superchips within a fixed pixel-wise Barlow Twins family, each evaluated on 15 downstream tasks. We find that pretraining loss barely predicts downstream performance (|Pearson r| < 0.2), so selecting models by loss wastes a large share of the compute. We also find that, as the training budget grows, the encoder and the data should grow together while the projector stays fixed, which gives a simple rule for allocating compute. Using this rule, we train a family of pixel-wise models (0.5B and 1B, with a 2B model in training) and distill them into compact students for embeddings-as-data deployment. The 21-million-parameter distilled TESSERA v2-1B-M in aggregate outperforms all open and proprietary models tested, some of which are orders of magnitude larger. These students produce Matryoshka representations that are inexpensive to serve: a 16-dimensional prefix keeps 92% of the full 128-dimensional performance at 1/8 of the storage. Upon completion of training we plan to release v2 global embeddings covering 2017-2025. Together, these results give a concrete, empirically grounded recipe for scaling pixel-wise EO foundation models: train large encoders, select by downstream performance, and distil into flexible student models. All code will be released at https://github.com/ucam-eo/tessera.

Empirical scaling laws relate loss to parameters $N$, training tokens $D$ and compute $C$. This technical note explains three mechanisms behind their exponents. Local approximation bounds function error at rate $N_e^{-γs/d}$, with effective parameters $N_e$, smoothness $s$, reuse exponent $γ$ and dimension $d$; the loss function determines the risk bound. Independent feature coverage gives $α_D=(β-1)/ζ$ from predictive-weight decay $β$ and occurrence rarity $ζ$. Separable-loss optimization determines training allocations and deployment crossovers. Nonlinear scoring can produce apparent emergent abilities, while grokking concerns delayed generalization and verification constrains test-time scaling. For regular molecular configurations, the $\mathrm{SE}(3)$ quotient gives $d_{\mathrm{int}}=d-6$ and relative approximation-exponent gain $6/d_{\mathrm{int}}$. Optimized dense-to-equivariant excess-loss ratios scale as $C^{p_{\mathrm{eq}}-p_{\mathrm{dense}}}$, where $p_j$ are compute-loss exponents and the prefactor depends on loss amplitudes and FLOP costs. Domain examples also cover visual resolution, conformational coverage and correlated-sample precision.

While much effort has focused on developing and benchmarking high-performance neural networks, less attention has been given to how dataset properties, known to practitioners, can guide efficient model selection. Neural models are typically evaluated on datasets with thousands of classes, yet many real-world applications involve fewer than ten. To address this understudied but common setting, we develop a measure of classification difficulty based on data-side properties and show how it enables more efficient model selection for few-class datasets, where traditional approaches are less effective. We term this phenomenon "few-class distinctiveness". Our metric allows comparison of models and datasets 6 to 29$\times$ faster than repeated training and testing. Leveraging this insight, we extend scaled model families below the smallest published models, achieving greater efficiency at similar accuracy, for example models up to 42% smaller than YOLOv5-nano for a mobile robot task. Targeting resource-constrained applications, we demonstrate few-class model selection across mobile robot, drone, and IoT scenarios, highlighting practical gains in efficiency without sacrificing performance.