GECO: Advanced Algorithms & Frameworks

• GECO is a collection of advanced algorithms and models that leverage constraint optimization, geometric modeling, and generative techniques to address challenges in machine learning and scientific computing.
• Its methods—including a constrained VAE training framework, geometry-aware dense feature extraction, and scalable graph modeling—combine mathematical rigor with demonstrable empirical gains.
• By enhancing model interpretability, efficiency, and robustness, GECO offers actionable insights and improved performance in domains like computer vision, robotics, and galaxy evolution.

GECO

GECO encompasses a diverse set of algorithms and models across machine learning, computer vision, graph modeling, control, numerical analysis, scientific discovery, and physical sciences. The acronym and associated terminology encapsulate several distinct but technically advanced frameworks. This entry surveys the principal approaches known as GECO and closely related methods, focusing on their mathematical principles, algorithmic structures, application domains, theoretical guarantees, and empirical results.

1. Generalized ELBO with Constrained Optimization (GECO) for VAEs

The GECO algorithm ("Generalized ELBO with Constrained Optimization") defines variational autoencoder (VAE) training as a constrained optimization problem with explicit reconstruction-type constraints in the data space, instead of the usual approach of tuning a penalty factor for rate-distortion trade-offs. The generic GECO VAE objective is:

$$\min_{\phi, \theta} \mathbb{E}_x \left[ \mathrm{KL}(q_\phi(z|x) \parallel \pi(z)) \right] \;\; \text{subject to } \mathbb{E}_{x,z\sim q_\phi(z|x)} [C(x, g_\theta(z))] \le 0$$

with

$$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$

This enforcement is achieved using Lagrange multipliers λ. At each iteration,

• Gradients of the combined Lagrangian (KL penalty plus the weighted constraint) are used to update θ and φ,
• λ is updated using a dual ascent strategy:

$$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$

where $\bar{C}^t$ is an exponential moving average of the current batch constraint.

The key features are: directly interpretable constraints in pixel MSE or other metrics, self-regularizing Lagrange multipliers with stop-gradient correction for numerical stability, and training dynamics characterized by initial aggressive reduction of constraint violation followed by KL minimization under the active constraint.

GECO delivers tight adherence to reconstruction thresholds, lower average and marginal KL relative to standard ELBO or β-annealing, and produces well-structured latent spaces as evidenced across multiple datasets and architectures (Rezende et al., 2018). The GECO optimizer is foundational to more recent approaches for dynamic latent bottleneck selection in VAEs, where it is combined with differentiable $L_0$-regularization and ARM gradient estimators to enable on-the-fly latent dimensionality pruning under a fixed reconstruction error tolerance and without hyperparameter sweeps (Boom et al., 2020).

2. GECO and GeCo in Geometric and Computer Vision Contexts

2.1 GECO for Geometric Correspondence

GECO ("Geometrically Consistent Embeddings with Lightspeed Inference") provides fast and geometry-aware dense feature extraction for visual correspondence. Given the limitations of classical backbone features in distinguishing symmetric or geometrically ambiguous parts, GECO integrates a lightweight LoRA (Low-Rank Adaptation) adapter into a frozen DINOv2-B transformer backbone and applies an unbalanced, entropy-regularized optimal transport loss between patch-wise features of image pairs:

$$\widehat{\mathbf{P}} = \arg\max_{\mathbf{P}\ge0} \langle \mathbf{P}, \mathbf{C}\rangle + \lambda H(\mathbf{P}) - \alpha \mathrm{KL}(\mathbf{P}\mathbf{1} \| \mathbf{a}) - \beta \mathrm{KL}(\mathbf{P}^T \mathbf{1} \| \mathbf{b})$$

with explicit "dustbin" indexing for occlusions. Networks are supervised using sparse keypoint correspondence and negative-pair constraints, enabling gradient signal beyond annotated points and robustness to occlusions/disocclusions.

GECO achieves PCK improvements of 6.0%–6.2% across benchmarks (PFPascal, APK, CUB), runs at 30 fps (~40 ms), and introduces the geometry-aware PGCK metric to better distinguish symmetric feature confusion (Hartwig et al., 1 Aug 2025).

2.2 Solar-GECO: Geometric-Aware Co-Attention for Perovskite Solar Cell PCE Prediction

Solar-GECO models power conversion efficiency (PCE) of perovskite solar cells by integrating geometric graph neural networks (CGCNN) encoding the 3D atomic structure of absorber crystals with LLM embeddings (MaterialsBERT) that process the chemical composition of transport layers. Fusion is implemented using a stack of multi-head self- and cross-attention (co-attention) blocks, followed by masked average pooling and a probabilistic regression head:

$$\mathcal{L} = \frac{1}{2B} \sum_{i=1}^{B} [\log \sigma_i^2 + (y_i - \mu_i)^2/\sigma_i^2]$$

Solar-GECO reduces MAE versus prior semantic GNNs (from 3.0668 to 2.936, p<0.05) and achieves R² = 0.418, Spearman’s ρ = 0.619, with well-aligned uncertainty calibration as PICP≈0.95 (Li et al., 24 Nov 2025).

2.3 GECO for Real-Time Geometric Correspondence in Robotics and 3D Modeling

GECO enables rapid geometric correspondence in vision-to-vision and vision-to-3D scenarios, including image-to-3D pipelines (GECO: Generative Image-to-3D within a SECOnd). GECO distills multi-view diffusion teacher models (Zero123Plus) into a single-step U-Net generator for rapid inference, and jointly finetunes generator plus 3D reconstruction modules for view-consistent generation. This pipeline delivers PSNR = 19.62, SSIM = 0.821, LPIPS = 0.159 on benchmarks, with <0.35 s runtime per scene (Wang et al., 2024).

3. GECO for Graph Modeling and Explanations

3.1 GECO: Graph-Enhanced Contextual Operator

GECO layers for scalable graph learning combine:

• A Local Context Block (LCB): neighbor aggregation and concatenation,

$H^{*(l)} = [\,H^{(l-1)},\,A\,H^{(l-1)}]$

• A Global Context Block (GCB): learnable FFT-based global convolutions across node indices with recurrence order $K$, capturing efficient “surrogate attention” in $O(N\log N+M)$ time,

delivering state-of-the-art performance and scalability to graphs with millions of nodes, e.g., obtaining accuracy of 73.10% (ogbn-arxiv), 55.55% (Flickr), and 96.65% (Reddit) with up to 169x speedup over attention-based methods (Sancak et al., 2024).

3.2 GECo for GNN Explanation

GECo ("Graph Explainability by COmmunities") is an instance-level perturbation-based explanation method that maps GNN decisions to meaningful graph communities. By performing modularity-based community detection, feeding each induced community into the trained GNN, and scoring their class confidence, GECo constructs explanation masks as unions of those communities exceeding a mean-confidence threshold. It is evaluated via Fidelity⁺ (necessity), Fidelity⁻ (sufficiency), characterization (harmonic mean), and Graph Explanation Accuracy (GEA):

$$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$0

GECo achieves Fid⁺ = 0.79–0.93, charact = 0.75–0.95, GEA = 0.31–0.56 on synthetic and real-world molecular datasets with <10 s/graph compute (Calderaro et al., 2024).

4. GECO and GeCo in Control, Robotics, and Anomaly Detection

4.1 GeCo-SRT: Geometric Continual Sim-to-Real Transfer in Robotics

GeCo-SRT introduces a geometry-aware mixture-of-experts residual module for sim-to-real continual transfer. Local geometric descriptors (planarity, linearity, saliency) derived from point clouds are assigned to specialized experts through soft gating, ensuring domain- and task-invariant representation. Geo-PER replay prioritizes batches that activate under-utilized experts, actively preventing catastrophic forgetting. Experiments across multiple manipulation tasks yield a 52% performance improvement over strong baselines and enable adaptation with ~1/6 data (Yu et al., 24 Feb 2026).

4.2 Generative Control as Optimization (GeCO)

In adaptive robotic control, GeCO reframes action-synthesis under flow-matching as stationary velocity field optimization:

$$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$1

where $$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$2 is a learned field whose zeros coincide with expert actions (stable attractors). At test time, actions are generated via iterative optimization until $$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$3. The field-norm provides a native OOD (out-of-distribution) safety score. Empirical results include a 91.9% success rate with 75% less compute versus time-conditioned flow, improved VLA-model transfer (success 18%→28%), and AUROC=0.93 for OOD detection (Zhang et al., 18 Mar 2026).

4.3 GeCo for Anomalous Sound Detection

GeCo leverages Predictive AutoEncoders (PAE) with transformers for generative frame-level prediction, augmenting with a supervised contrastive loss to improve class discriminability. The combined loss:

$$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$4

enables both sensitive anomaly detection (MSE) and robust class separation, achieving AUC = 92.5%, pAUC = 86.3% on DCASE2020 Task 2, exceeding discriminative and generative-only baselines (Zeng et al., 2023).

5. GECO and GeCo in Scientific Computing and Semi-Analytic Modeling

5.1 Geometric Conservative (GeCo) ODE Integration Schemes

GeCo methods generalize explicit Runge–Kutta time integration by introducing solution-dependent positive weights (φ) to enforce unconditional positivity and conservation of invariants in stiff or chemically reactive ODEs:

For first order (GeCo1):

$$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$5

where $$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$6, ensuring positivity for all $$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$7. Second order (GeCo2) schemes are constructed analogously. Stability theorems guarantee Lyapunov stability of fixed points when the Jacobian’s eigenvalues meet spectral radius criteria. For GeCo1, the stability domain is unbounded (no step-size restriction); for GeCo2, it is bounded (critical $$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$8) (Izgin et al., 2023).

5.2 Arbitrary-Order GeCo Schemes: Order Conditions

GeCo schemes of order $$C(x, g_\theta(z)) = \| x - g_\theta(z) \|^2 - \kappa^2$$9 are derived by embedding solution-dependent weights into the Butcher tableau and systematically applying NB-series and N-tree expansions:

Stage and update formulas:

$$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$0

General order conditions reduce to weighted sum identities on the coefficients (e.g., $$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$1), providing a path toward construction of third- and fourth-order schemes with exact positivity and mass preservation guarantees (Izgin et al., 2023).

5.3 GECO: Galaxy Evolution COde

GECO (Galaxy Evolution COde) is a semi-analytical model for galaxy formation integrating Monte Carlo-based extended Press–Schechter hierarchical merger trees of dark matter halos and analytic baryonic modules (cooling, star formation, feedback, black hole growth and feedback). The calibrated parameter vector (e.g., $$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$2, $$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$3, $$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$4) allows matching local stellar mass functions and BH-bulge relations. GECO reproduces the observed downsizing of star formation (formation redshifts increasing with halo mass) even in the absence of AGN feedback, while stellar mass assembly follows a hierarchical pattern, consistent with ΛCDM (Ricciardelli et al., 2010).

6. GECo as Interpretable and Counterfactual Explanations

GeCo is also instantiated as an algorithm for real-time, high-quality, plausible and feasible counterfactual explanations. It employs a customized genetic algorithm with a Δ-representation of candidate edits and partial evaluation of classifiers, optimizing:

$$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$5

with domain-aware distance metrics and application of user-provided plausibility and feasibility (PLAF) constraints. GeCo achieves faster (100–300 ms per instance), more concise explanations (average 1.27 features modified) than other constraint-satisfying or heuristic counterfactual systems (Schleich et al., 2021).

7. GECO and GeCo in Low-Shot Detection, Retrieval, and Compatibility

7.1 GeCo: Unified Low-Shot Counting, Detection, and Segmentation

GeCo develops a single-stage architecture for low-shot object counting, leveraging dense per-pixel prototype expansion from few- or zero-shot exemplars, and a novel detection-specific loss:

$$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$6

Substantial MAE and AP50 gains are observed over baselines on FSCD147 and FSCD-LVIS (Pelhan et al., 2024).

7.2 GeCo: Generative Compatibility Model for Fashion Retrieval

GeCo combines conditional GAN-based complementary item generation (Pix2Pix cGAN) and a dual-head ResNet-18-based compatibility model jointly trained by BPR and InfoNCE objectives. The compatibility score is

$$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$7

where $$\log \lambda^{t} \leftarrow \log \lambda^{t-1} + \eta \bar{C}^t$$8 is composed from input and generated images. GeCo significantly outperforms prior compositional and generative methods in AUC (up to 0.8491) and MRR across FashionVC, ExpFashion, and FashionTaobaoTB datasets (Attimonelli et al., 2024).

---

This collection demonstrates that GECO and GeCo, as a family of advanced algorithms and frameworks, share a unifying thread: the use of geometric, structural, constraint-based, or generative modeling to achieve improved fidelity, interpretability, efficiency, or scientific accuracy across a wide array of research domains.