--- title: Condensation Point Clustering url: https://www.emergentmind.com/topics/condensation-point-clustering type: topic --- # Condensation Point Clustering Searching arXiv for recent and foundational papers on object condensation / condensation-point clustering. arxiv_search.query({"search_query":"all:\"object condensation\" OR all:\"condensation clustering\"","start":0,"max_results":10}) arxiv_search.search({"query":"object condensation clustering","max_results":10}) arxiv_search({"query":"object condensation clustering","max_results":10}) {"query":"object condensation clustering","max_results":10} Condensation point clustering denotes a family of clustering and condensation formulations in which representative points emerge and act as anchors for grouping. In the machine-learning formulation introduced as object condensation, each vertex predicts object properties $p_i$, a condensation score $\beta_i\in(0,1)$, and coordinates $x_i\in\mathbb{R}^d$ in a learned clustering space; high-$\beta$ vertices act as condensation points, and nearby vertices are assigned to them by a simple thresholded procedure [2002.03605]. Related uses of the same vocabulary appear in graph condensation, where class-wise cluster centroids become synthetic nodes [2502.17614], and in statistical mechanics or random CSPs, where condensation refers to regimes in which a small number of clusters or condensates dominate the state space [2603.00806] [2305.17334]. ## 1. Object-condensation formalism The canonical machine-learning formulation treats pixels, detector hits, graph vertices, or point-cloud points as the primitive carriers of both instance membership and object properties. For each vertex $i=1,\dots,N$, the ground truth comprises membership indicators $M_{ik}\in\{0,1\}$ over objects $k=1,\dots,K$, a background label $n_i\in\{0,1\}$, and target properties $t_i$. The network predicts per-vertex properties $p_i$, a condensation score $\beta_i$, and clustering-space coordinates $x_i$. From $\beta_i$ one defines a charge $$ q_i = \operatorname{arctanh}^2(\beta_i) + q_{\min}, $$ with $q_{\min}>0$ a small hyperparameter [2002.03605]. The total loss is $$ L = L_p + s_c(L_\beta + L_V), $$ where $s_c$ balances condensation against property regression. The property term uses the charge-weighted factor $$ \xi_i = (1-n_i)\operatorname{arctanh}^2(\beta_i), $$ and $$ L_p = \frac{1}{\sum_{i=1}^N \xi_i}\sum_{i=1}^N \xi_i\,L_i(t_i,p_i). $$ The condensation-point assignment term enforces exactly one high-$\beta$ representative per object and suppresses background activations: $$ L_\beta = \frac1K\sum_{k=1}^K [1-\beta_{\alpha_k}] + s_B\cdot \frac{1}{N_B}\sum_{i\,|\,n_i=1}\beta_i, $$ where $\alpha_k=\arg\max_{i:M_{ik}=1}(q_i)$ is the highest-charge vertex in object $k$ [2002.03605]. The geometric part of the loss approximates each object by its highest-charge vertex. For a vertex $j$ belonging to object $k$, the attractive potential is $$ \breve V_k(x_j)=q_{\alpha_k}\|x_j-x_{\alpha_k}\|^2, $$ while for a vertex outside object $k$ the repulsive potential is $$ \hat V_k(x_j)=q_{\alpha_k}\max(0,1-\|x_j-x_{\alpha_k}\|). $$ The resulting pull-push term is $$ L_V = \frac1N\sum_{j=1}^N q_j\sum_{k=1}^K \left[M_{jk}\breve V_k(x_j)+(1-M_{jk})\hat V_k(x_j)\right]. $$ This construction makes the method independent of assumptions on object size, sorting or object density, and it generalises to non-image-like data structures such as graphs and point clouds [2002.03605]. ## 2. Inference, clustering mechanics, and implementation choices After training, losses are no longer evaluated. Inference proceeds by collecting all vertices with $\beta_i>t_\beta$ into a candidate list, sorting them by descending $\beta$, and accepting a candidate as a final condensation point only if it lies farther than $t_d$ from all previously accepted points in clustering space. The accepted set $S$ is then used for nearest-representative assignment: every vertex is assigned to its nearest $s\in S$ if $\|x_i-x_s\|0$ by the sigmoid activation. Batch size and learning-rate scheduling, including cyclic learning rates between $10^{-5}$ and $10^{-4}$, are described as stabilizing choices. In graph or point-cloud applications, any GNN layer such as GravNet can be used to produce $x_i$, $\beta_i$, and $p_i$; for images, the reference implementation uses a small U-Net-style convnet predicting a grid of vertices each with $(x,\beta,p)$ [2002.03605]. The method is positioned against both computer-vision and detector-specific baselines. For image-like data, each pixel is a vertex and the condensation loss merges pixels of the same object without anchors or NMS. For graphs and point clouds, graph edges need not be pre-built because $x_i$ is learned directly. Traditional clustering methods such as DBSCAN or HDBSCAN cluster in input space and require separate seeding or thresholding, whereas object condensation integrates seeding via $\beta$ and clustering into a single end-to-end loss. Likewise, particle-flow algorithms use hard rules for seeding, clustering, and track-cluster linking; object condensation replaces seeding, clustering, track-cluster linking and thresholding by a single differentiable network and loss [2002.03605]. ## 3. Detector reconstruction and empirical realizations As introduced, object condensation was presented as a one-stage grid-free multi-object reconstruction method for physics detectors, graph data, and image data, with proof-of-concept applications to a simple object-classification task in images and to reconstruction of multiple particles from detector signals [2002.03605]. The detector setting is especially natural because sparse calorimeter or tracker readouts are more naturally represented as irregular sets or graphs than as dense images. A detailed later realization is the CLAS12 calorimeter study, which applies object condensation clustering to neutron and photon reconstruction in a hodoscopic detector [2503.11277]. In that formulation, each hit predicts a $2$-D condensation coordinate $m_i=(x_{c,i},y_{c,i})\in\mathbb{R}^2$ and a confidence score $\beta_i\in[0,1]$. For each true cluster $t$, the representative hit is $$ r_t=\arg\max_{i\in C_t}\beta_i, $$ and the latent distance is $d_{i,t}=\|m_i-m_{r_t}\|_2$. The loss splits into $\mathcal{L}=\mathcal{L}_V+\mathcal{L}_\beta$, with an attractive term that pulls hits toward their own condensation point, a repulsive term with margin $\Delta$ that separates different clusters, a “coward” term that forces at least one hit per true cluster to have high $\beta$, and a noise term that penalizes large $\beta$ on padded or background hits [2503.11277]. The CLAS12 inference procedure sorts hits by descending $\beta_i$, repeatedly promotes the highest unassigned hit with $\beta_i\ge t_\beta$ to a new seed, and collects all unassigned hits within latent distance $t_D$ into the cluster; all remaining hits are labeled noise. The reported hyperparameters are $t_\beta=0.5$ and $t_D=0.28$. The model accepts up to $V=150$ hits per event, each with $F=17$ features, constructs per-hit embeddings with BatchNorm and three fully connected layers, computes positional encodings with four GravNet blocks over a fixed tensor for all $H=2448$ strips, and contextualizes the resulting tokens with a four-layer Transformer encoder before predicting $(x_{c,i},y_{c,i},\beta_i)$ [2503.11277]. Evaluation is reported on one million simulated $e+p$ collision events. A reconstructed neutral cluster is defined as trustworthy when there exists exactly one true particle of that type within $\Delta\theta\le 4^\circ$ and $\Delta\phi\le 4^\circ$ of the reconstructed direction, and no other reconstructed cluster of the same type lies within that cone. Under this criterion, the fraction of reliable neutron clusters increases from $8.98\%$ to $30.65\%$, and the photon fraction increases from $51.10\%$ to $63.64\%$ [2503.11277]. The study also states that it is the first application of AI clustering techniques for hodoscopic detectors [2503.11277]. ## 4. Graph-condensation usage of condensation-point clustering A distinct but related use of the term appears in GECC, a graph condensation method designed for large-scale and evolving graph data [2502.17614]. GECC replaces gradient- or trajectory-matching approaches with a two-stage, clustering-based procedure. At each time step $t$, it first applies an SGC-style propagation over the normalized adjacency, $$ \bar A_t=\tilde D_t^{-1/2}\tilde A_t\tilde D_t^{-1/2}, \qquad \tilde A_t=A_t+I, $$ and forms propagated features $$ F_t=\sum_{k=0}^{K}\alpha_k \bar A_t^k X_t. $$ It then performs class-wise clustering on the propagated representations, representing class $k$ by $M_k=\lceil n_{t,k}\,r\rceil$ synthetic nodes and producing a condensed graph $G_t'=(I_M,C_t)$ with node features $C_t\in\mathbb{R}^{M\times d}$ and identity adjacency [2502.17614]. The clustering stage uses an assignment matrix $P_t$ and centroids $C_t$ such that $P_tC_t$ approximates $F_t$, with centroid update $$ C_t = D_{P_t}^{-1}P_t^\top F_t. $$ To control both representation distortion and parameter-matching error, GECC minimizes the balanced objective $$ J(P_t,C_t)=\|F_t-P_tC_t\|^2+\|P_t^\top \mathbf 1-u\|^2, $$ where the second term penalizes deviations from ideal cluster sizes. Optimization is performed with hard or soft EM, and in evolving settings previous centroids are inherited through an incremental k-means++ scheme that augments $C_{t-1}$ with new seeds from arriving data [2502.17614]. The paper provides theoretical support through three bounds. The training-stage bound shows that matching propagated features and model parameters suffices to match loss gradients; the test-stage bound implies that limiting the final parameter shift yields generalization guarantees on unseen data; and the third theorem gives $$ \|W-W'\|\le \mathcal C\cdot (\max \operatorname{diag}(P^\top P))^2, $$ so balanced cluster sizes directly tighten the parameter-distance bound [2502.17614]. Complexity is reported as $O(K e_t d)$ for feature propagation and $O(n_t U M d)$ per restart for hard or soft k-means, plus $O(n_tM)$ for the balance term. The method is described as scaling linearly in $n_t$ and $e_t$ with $M\ll n_t$ [2502.17614]. Empirically, GECC is evaluated on transductive datasets including Citeseer, Cora, Pubmed, Ogbn-arxiv, and Ogbn-products, and inductive datasets including Flickr and Reddit. The abstract reports an around $1{,}000\times$ speedup on large datasets [2502.17614]. The detailed Reddit example at reduction rate $r=0.001$ gives $89.8\%$ accuracy and $6{,}100$ s condensation for GCond, $91.3\%$ and $1{,}815$ s for GEOM, and $91.4\%$ and $4.9$ s for GECC; removing feature propagation causes a $3$–$5\%$ absolute accuracy drop, omitting centroid reuse leads to $10\times$–$20\times$ more k-means iterations at large $t$, and selecting the best of $50$ restarts by lowest $J$ yields $+2.7\%$ on Citeseer relative to random-initialized k-means [2502.17614]. ## 5. Condensation and clustering in statistical mechanics Outside machine learning, condensation point clustering describes physical cluster formation under cooling, exclusion constraints, or size-dependent stationary weights. In the two-dimensional Lennard-Jones study under controlled exponential cooling, particles are considered bonded when their separation satisfies $r\le 1.5r_0$, and clusters are connected sets of such bonds. The temperature follows $$ T(S)=T_i e^{-aS}, $$ with $T_i=0.2\,\epsilon/k_B$ and $T_f=0.06$. The final-equilibrium RMS displacement exhibits a broad maximum at $a\approx a_c\approx 10^{-3}$, the cluster-size distribution is well fit by a two-parameter Gamma form, and for $c\ge c_c$ with $c_c\sim 0.5$ all particles coalesce into a single macroscopic cluster [1308.6655]. In the driven pair exclusion process, condensation is mesoscopic rather than macroscopic. Above the threshold density $$ \rho_c=\frac{1}{b-2}, $$ the system develops multiple condensates with scaling $$ m_{\rm con}\sim N^\beta,\qquad N_{\rm con}\sim N^\alpha,\qquad N_{\rm con}m_{\rm con}\sim N. $$ The reported exponents are approximately $(\alpha,\beta)_{\rm 1D}\approx(0.51,0.56)$ and $(\alpha,\beta)_{\rm 2D}\approx(0.34,0.62)$. The drive induces spatial correlations among condensates, so a cluster of condensates appears; in two dimensions the cluster is anisotropic, with $\delta_\parallel\approx 0.25$ and $\delta_\perp\approx 0.10$ [1203.0867]. A more recent rigorous treatment considers stochastic lattice gases with size-dependent stationary weights $$ w_L(n)=w(n)+\theta L^{-\gamma}(n+1)^\kappa[1+o(1)]. $$ For $\rho>\rho_c$, the system phase-separates into a bulk at the critical grand-canonical measure and a condensed phase containing the excess mass. In the mesoscopic regime, the characteristic condensate scale is $$ C_L=\left(\frac{(\rho-\rho_c)L^\gamma}{\theta\,\Gamma(\kappa+2)}\right)^{1/(\kappa+2)}, $$ and the size-biased sampled cluster size obeys $$ \tilde\eta_1/C_L \Rightarrow Z $$ conditionally on being in the condensate, where $Z\sim \Gamma(\kappa+2,1)$. When $\gamma>\kappa+2$, one instead obtains a single macroscopic condensate with $\max_x \eta_x/L \Rightarrow \rho-\rho_c$ [2603.00806]. A related one-dimensional hopping model with attractive logarithmic interactions has equilibrium gap distribution $$ P(\vec y)=Z^{-1}\delta\!\left(\sum_i y_i-L\right)\prod_{i=1}^N y_i^{-\beta}, \qquad \beta=J/T. $$ The critical coupling is $\beta_c=1$, and in the scaling limit $$ 1-\beta(N)=\frac{b-1}{N-1}, $$ the rescaled gap distribution for the gap containing a uniformly chosen point converges to $$ \hat P_g(\hat y)=(b-1)(1-\hat y)^{b-2},\qquad 0<\hat y<1. $$ The model thereby exhibits multiple correlated clusters and a self-similar spatial structure [1608.04304]. ## 6. Distinctions across domains and broader theoretical interpretations The same terminology covers formally different clustered objects. In learned object condensation, the clustered entity is a set of pixels, hits, or vertices organized around a high-confidence latent representative. In GECC, the clustered entity is a class-wise centroid that becomes a synthetic node in a condensed training graph. In statistical mechanics, the clustered entity is a particle aggregate or condensate, while in random CSPs it is a dominant solution cluster. | Domain | Representative papers | Clustered object | |---|---|---| | Object condensation | [2002.03605], [2503.11277] | Vertices or hits around learned condensation points | | Graph condensation | [2502.17614] | Propagated node embeddings around class-wise centroids | | Physical or combinatorial condensation | [1308.6655], [1203.0867], [2603.00806], [2305.17334], [1307.6374] | Particle clusters, condensates, vortices, or solution clusters | This distinction is especially important in the two theoretical literatures that use “condensation” without a learned latent geometry. In decaying two-dimensional quantum turbulence, negative-temperature states exhibit both macroscopic vortex clustering and kinetic-energy condensation, termed an Onsager-Kraichnan condensate; the clustered fraction, maximum cluster charge, pair-sign correlations, and the large-scale excess in the incompressible kinetic-energy spectrum serve as order parameters [1307.6374]. In random regular NAE-SAT, the condensation regime is the interval $(\alpha_{\rm cond},\alpha_{\rm sat})$ in which $O(1)$ clusters carry almost all solutions, and the local weak limit becomes explicitly non-Markovian because the solution measure is dominated by a small number of large clusters [2305.17334]. Across these settings, the common structural theme is concentration onto a reduced set of representatives: latent representatives in object condensation, centroids in graph condensation, and dominant physical or combinatorial clusters in condensed phases. The technical content, however, remains domain-specific. In detector reconstruction the decisive quantities are $\beta_i$, learned coordinates, and pull-push losses; in graph condensation they are propagated embeddings, balanced assignments, and centroid evolution; in statistical mechanics and CSPs they are cluster-size distributions, critical densities or temperatures, and the transition from many small clusters to a small number of dominant ones [2002.03605] [2502.17614] [2603.00806] [2305.17334]. Source: https://www.emergentmind.com/topics/condensation-point-clustering