CORDS: Continuous Representations of Discrete Structures

Many learning problems require predicting sets of objects when the number of objects is not known beforehand. Examples include object detection, molecular modeling, and scientific inference tasks such as astrophysical source detection. Existing methods often rely on padded representations or must explicitly infer the set size, which often poses challenges. We present a novel strategy for addressing this challenge by casting prediction of variable-sized sets as a continuous inference problem. Our approach, CORDS (Continuous Representations of Discrete Structures), provides an invertible mapping that transforms a set of spatial objects into continuous fields: a density field that encodes object locations and count, and a feature field that carries their attributes over the same support. Because the mapping is invertible, models operate entirely in field space while remaining exactly decodable to discrete sets. We evaluate CORDS across molecular generation and regression, object detection, simulation-based inference, and a mathematical task involving recovery of local maxima, demonstrating robust handling of unknown set sizes with competitive accuracy.

💡 Research Summary

The paper introduces CORDS (Continuous Representations of Discrete Structures), a framework that converts a set of spatial objects with associated features into two continuous fields: a density field ρ(r) and a feature field h(r). Each object is represented by a kernel K(r; r_i) that is normalized to have unit integral α. The density field is defined as ρ(r)= (1/α)∑{i=1}^N K(r; r_i), while the feature field is h(r)= (1/α)∑{i=1}^N x_i K(r; r_i). Because each kernel contributes the same mass α, the total mass ∫Ω ρ(r)dr equals the cardinality N, providing a direct, differentiable estimate of the number of objects without any explicit counting module.

Positions are recovered by solving a kernel‑matching optimization problem that minimizes the L2 distance between the observed density and a superposition of N kernel translates. In the ideal case the original positions are the global optimum; in practice gradient‑based solvers yield sufficiently accurate approximations. Features are recovered by projecting the feature field onto the subspace spanned by the recovered kernels. Forming the Gram matrix G_{ij}=∫Ω K(r; r_i)K(r; r_j)dr and the vector B_i=∫Ω h(r)K(r; r_i)dr, the original feature matrix X can be solved analytically as X = α G^{-1} B, provided the kernel yields a positive‑definite G (which holds for common choices such as Gaussians).

To make the continuous representation tractable for neural networks, the fields are sampled at a finite set of points. For regular grids (images, time series) uniform sampling is used; for sparse 3‑D data (molecules) importance sampling proportional to ρ(r) concentrates points where signal exists, avoiding empty regions and bounding‑box constraints. Sampled tuples (r_m, ρ(r_m), h(r_m)) are fed directly into neural architectures. Grid‑based data are processed with standard 2‑D or 1‑D CNNs, while unordered point clouds are handled by Erwin, a hierarchical permutation‑invariant transformer that scales to thousands of points.

The authors evaluate CORDS on four distinct tasks that naturally involve variable‑size sets:

1. Molecular generation and property regression (QM9 and GeomDrugs). CORDS learns a joint distribution over atom coordinates, atom types, and charges in field space. Unconditional and conditional generation experiments show that the model can accurately reproduce atom counts (via density mass) and generate chemically valid molecules, achieving performance comparable to state‑of‑the‑art graph‑based generators while eliminating the need for explicit graph decoding during training.

2. Object detection in images (MultiMNIST with out‑of‑distribution counts). The model predicts a density map and per‑class feature maps; decoding yields both the number of digits and their positions/classes. Even when the test set contains digit counts unseen during training, CORDS maintains high counting accuracy and precise localization, outperforming traditional heat‑map detectors that lack feature reconstruction.

3. Simulation‑based inference in astrophysics (burst decomposition of light curves). Light‑curve data are encoded as temporal density and feature fields representing burst events. CORDS jointly infers the number of bursts and their parameters, demonstrating that continuous field modeling can replace padding‑based approaches commonly used in neural posterior estimation.

4. Synthetic benchmark for local maxima recovery. A continuous function with multiple peaks is transformed into a density field; CORDS successfully recovers the exact locations and values of the peaks, illustrating the method’s applicability to abstract mathematical problems.

Across all experiments, CORDS shows that a single, mathematically principled representation can handle cardinality, localization, and attribute prediction simultaneously. The main advantages are: (i) implicit, differentiable cardinality via density mass, (ii) exact invertibility guaranteeing lossless decoding, (iii) uniform applicability across modalities, and (iv) elimination of padding or fixed‑size slots. Limitations include sensitivity to kernel choice and bandwidth, computational cost of the position‑recovery optimization for large N, and the need for careful sampling strategies in high‑dimensional spaces.

Future work suggested by the authors includes learning kernel parameters end‑to‑end, developing faster decoding schemes (e.g., direct peak detection), and scaling the approach to massive 3‑D volumetric data or long time‑series. Overall, CORDS offers a compelling alternative to discrete‑set modeling by leveraging continuous fields, opening new possibilities for tasks where the number of objects is unknown a priori.