acceptodds
Under review as a conference paper at ICLR 2027

Joint-Domain Implicit Conditional Density Estimation under Sparse Structured Coverage

Abstract

Conditional density estimation (CDE) aims to estimate the full conditional distribution \(p(y\mid x)\). We study CDE under sparse coverage of ordered or gridded conditioning domains. Structure-aware CDE approaches typically encode cross-condition sharing through prespecified analytic relational functions, such as kernels or covariance functions. While effective when well matched to the data, the chosen functional family can restrict the form of admissible interactions. We introduce Joint-domain Implicit Conditional Density Estimation (JICDE), which represents \({p(y\mid x):x\in\mathcal T}\) as the location-wise marginals of a stochastic generator defined jointly over the conditioning domain. JICDE retains the geometry and locality of the conditioning domain as architectural inductive biases, while parameterizing the detailed cross-condition sharing rule through shared learnable local transformations, allowing the induced sharing behavior to adapt across different structural regimes. Across ordered and gridded structured CDE benchmarks, JICDE consistently compares favorably with representative CDE and structured probabilistic baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conditional Density Estimation, Structured Probabilistic Modeling, Implicit Generative Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_sZmMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses conditional density estimation when observations cover only a small subset of a structured conditioning domain, such as a temporal sequence or a two-dimensional image grid. Existing approaches, including kernel CDE and Gaussian-process-based methods, encode cross-condition relationships through predefined similarity or covariance functions, which may be poorly matched to the underlying data.

The authors propose Joint-Domain Implicit Conditional Density Estimation (JICDE), which generates a stochastic response field over the entire conditioning domain. It combines random noise with Fourier positional encodings and employs shared deconvolutional transformations to model dependencies between neighboring conditions. The model is trained using the Continuous Ranked Probability Score (CRPS), encouraging its generated samples to match the observed conditional distributions.

Critical strengths

(1) Sparse conditioning coverage is a meaningful challenge for CDE. The paper appropriately emphasizes that estimating an entire predictive distribution at poorly observed locations requires effective information sharing across conditions, rather than merely accurate point prediction.

(2) A coherent joint-domain formulation. Representing the family of conditional distributions as location-wise marginals of a shared stochastic generator offers a unified perspective on structured CDE. It avoids treating every conditioning location as an independent estimation problem.

(3) The temporal CDE experiments and image-completion experiments demonstrate that the same general modeling principle can be applied to different structured geometries. This strengthens the case that the approach is not limited to a single data format.

Critical weaknesses

(1) The methodological novelty relative to existing convolutional generative models is insufficiently established. The central contribution is to represent conditional distributions jointly over a structured domain and use shared deconvolutional transformations to learn cross-condition interactions. However, convolutional neural processes (ConvNP, ConvCNP), convolutional generative models, and DDN already exploit convolutional structure and shared transformations to model spatially organized outputs. The paper explains how JICDE differs conceptually from these approaches, but does not sufficiently isolate the genuinely new methodological component. In particular, it is unclear whether the main contribution is the joint-domain probabilistic formulation, the use of a stochastic decoder over the conditioning domain, or simply applying a convolutional generator to sparse CDE.

(2) The claimed adaptive cross-condition sharing mechanism is not directly validated. A central motivation is that JICDE learns its sharing rule from data, whereas kernel CDE and GP-based methods prescribe relationships through kernels or covariance functions. However, the experiments primarily report prediction metrics and qualitative reconstructions. Better performance alone does not establish that the learned interaction mechanism actually adapts to different structural regimes as claimed. For example, the paper does not directly examine the learned effective receptive fields, sensitivity to neighboring conditions, or how information propagates across locations in the Gaussian, Markov, and multimodal settings.

(3) The lack of ablation studies makes the source of the performance gains unclear. JICDE combines several components: stochastic noise embeddings, Fourier positional encodings, a global latent representation, a hierarchical deconvolutional decoder, and CRPS-based training. The paper does not sufficiently establish the individual contribution of each component.

(4) Insufficient comparison with modern conditional generative models. The proposed JICDE employs a stochastic generator with shared deconvolutional layers to estimate conditional distributions over sparsely observed structured domains. However, the relationship between this formulation and modern conditional diffusion or flow-matching models is not sufficiently discussed or empirically investigated. These generative frameworks can also model distributions conditioned on sparse observations, while incorporating spatial or temporal inductive biases through their architectures. It is therefore unclear whether the reported improvements arise from the proposed structured-domain formulation, the shared convolutional architecture, or the choice of a direct stochastic generator trained with CRPS. Comparisons with appropriately adapted conditional diffusion and flow-matching baselines, or at least a clearer discussion of their relative advantages in modeling quality, sampling cost, and cross-location dependencies, would help establish the methodological novelty and practical benefits of JICDE.

Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Weakness 1: Methodological novelty and relation to convolutional models

Thank you for pointing out that the boundary of our methodological contribution was insufficiently clear. We agree that convolutional architectures, stochastic decoders, and parameter sharing across locations are established ideas. Our contribution concerns the parameterization and fitting scheme for conditional distribution estimation on sparsely observed structured domains; the original manuscript did not sufficiently distinguish this contribution from its architectural components.

We consider a predefined conditioning domain with sparse response observations from a single instance and seek predictive distributions at unobserved locations. Without repeated observations at individual locations, these distributions cannot be freely fitted in isolation, motivating their joint estimation through shared structural constraints. ConvCNP and ConvNP already exploit convolutional structure and shared transformations for spatially organized prediction. The original ConvNP framework learns an amortized predictor across task instances. Our setting instead provides only sparse observations from the current instance. Repeated context–target splits can construct training tasks, but do not supply the additional information available from observing other instances, which may limit the performance. This motivates directly fitting the location-wise distribution family on the current domain, without learning a prediction mapping intended for reuse across instances. JICDE combines noise and positional embeddings with shared local deconvolutional transformations along the conditioning domain. Resampling latent variables produces response fields whose location-wise marginals represent the predictive distributions, and the shared parameters are optimized using marginal CRPS at observed locations. Under our data conditions and training configuration, JICDE outperforms the evaluated ConvNP baseline, providing empirical support for this instance-specific estimation scheme.

The distinction from DDN concerns the domain over which deconvolution operates. DDN constructs a density representation over discretized response-space bins and applies deconvolution to that representation. JICDE instead organizes local transformations along conditioning locations, jointly constraining neighboring location-wise distributions through shared transformations. Thus, the deconvolutional structure serves different roles in the two estimation schemes. Convolutional single-instance fitting has precedents in DIP (Ulyanov et al., 2018) and its Bayesian extensions. DIP fits a convolutional network to a single degraded image to obtain a reconstruction, while Bayesian DIP variants based on weight-posterior inference quantify reconstruction uncertainty through approximate inference over network weights. JICDE directly fits a stochastic generator using marginal CRPS at observed locations. After fitting, network parameters remain fixed while latent variables are resampled, yielding location-wise predictive distributions as the marginals of generated response fields. The distinction therefore concerns the probabilistic representation and training objective, rather than convolutional architecture or single-instance optimization alone. We have also added a Bayesian DIP baseline to the image-completion experiments, with competitive results reported in our response to Weakness 4.

Our contribution is therefore the combination of a structured stochastic parameterization over the conditioning domain and direct marginal-score fitting for single-instance, sparse fixed-domain CDE. We will make this contribution explicit in the revised manuscript and distinguish it from the established convolutional architectures and stochastic generation mechanisms on which it builds. We thank the reviewer for this constructive feedback, which has helped us articulate our methodological contribution more precisely and better position the work within the existing literature.

AuthorsOpenReview ↗

Response to Weakness 2: Analysis of cross-condition information sharing

Thank you for requesting a more direct examination of cross-condition sharing. We agree that predictive performance alone does not fully characterize how information is shared across locations. We therefore combine the existing distribution visualizations with additional analyses of receptive fields and the influence of training observations.

First, the original synthetic visualizations provide qualitative evidence of distributional flexibility. On Gaussian, Markov, and Double Sine data, JICDE produces Gaussian-like predictive shapes, discrete-valued structure, and bimodal branches, respectively. Visualizing learned distributions also has precedents in related distribution-learning studies, such as OPT-GAN (Lu et al., 2023) and OPT-VAE (Zhang et al., 2026), which examine how the learned distributions evolve over complex problem landscapes. We use these visualizations to qualitatively assess how the same modeling framework accommodates different distributional structures. We will extend this analysis with additional visualization experiments, including sharply peaked distributions.

Second, we have added a convolutional kernel-size analysis and effective receptive field (ERF) diagnostics. The kernel-size analysis examines the effect of the architectural interaction range, with a kernel size of 7 performing well overall among the tested configurations. The ERF analysis further shows that output sensitivity is concentrated around the corresponding local neighborhood in the intermediate feature representation. Together, these analyses characterize the architectural range and the feature-level sensitivity of the decoder.

Third, we conducted observation-deletion interventions to examine how training observations affect predictions at unobserved locations. Starting from the same original sparse training set on a 200-point sequence, we separately removed three random predefined groups: {25,38,80}, {90,100,101}, and {142,147,154}, using zero-based indices. We refitted the model with paired initializations and matched training settings, and measured the Wasserstein-1 (W1) distance between the resulting predictive distributions and those of the reference model fitted to the original sparse training set. All comparisons use the same originally unobserved evaluation locations. The table reports the average W1 changes on the Gaussian and Markov datasets. In addition to these aggregate values, we have plotted location-wise W1 response curves to show the spatial distribution of the changes. Stronger responses occur primarily in unobserved regions near the deleted observations, and the main response regions shift as the deleted groups move from the earlier to the middle and later parts of the sequence. These results provide empirical evidence of cross-location influence and its locality through interventions on the training data.

W1 measures prediction changes after deletion, not prediction error. Values are reported as mean±SD; indices are zero-based.

DeletionRemoved indicesGaussian dataMarkov data
125,38,800.26±0.070.17±0.06
290,100,1010.11±0.020.06±0.03
3142,147,1540.05±0.020.09±0.03
AuthorsOpenReview ↗

Response to Weakness 3: Component contributions and ablation studies

Thank you for pointing out that the contributions of individual components were not sufficiently clear. The original submission already included approximately parameter-matched MLP, Transformer, and FIXED LOCAL decoder ablations in Appendix B.6, where they were placed due to space constraints. We apologize for not highlighting these analyses sufficiently in the main text. We agree that their coverage was incomplete and have therefore added ablations of the noise embedding encoder, Fourier positional encoding, and training objective. We will present the key findings together in the revised main text.

  1. Noise embedding encoder. Freezing the encoder parameters while retaining latent resampling during both training and testing increases mean CRPS from 0.86, 0.54, and 1.10 to 0.95, 0.60, and 1.69 on BIA, PDB, and ETT, respectively. Mean NLL and MAE also worsen on all three datasets. This supports the utility of learning the noise embedding within the proposed architecture.

  2. Fourier positional encoding. Removing the Fourier positional encoding worsens mean MAE and CRPS on all three datasets. Mean NLL also worsens on BIA and PDB, although it improves slightly on ETT. These results support the value of explicit positional features in the tested sparse-observation settings.

  3. Local decoder structure and learnable spatial mixing. In the approximately parameter-matched comparisons in Appendix B.6, the MLP variant performs consistently worse. FIXED LOCAL retains the local structure and learnable channel transformations but replaces learnable spatial mixing with a fixed Gaussian operator. It performs best on the Gaussian benchmark, whereas JICDE performs better on Markov, PDB, and ETT. These results support the selected decoder structure and the value of learnable spatial mixing, while also showing that a fixed smoothing prior can be effective when well matched to the data.

  4. CRPS training objective. Compared with the MSE-trained variant, the full model achieves lower mean CRPS and NLL on all three datasets, while the MSE variant achieves lower MAE. This observed trade-off between point and distributional prediction supports the use of CRPS for the intended distribution-estimation objective.

We appreciate the reviewer's constructive feedback, which has helped us make the component analysis more systematic and the evidence supporting our design choices clearer.

Additional component ablations

DatasetBIABIABIAPDBPDBPDBETTETTETT
VariantNLLMAECRPSNLLMAECRPSNLLMAECRPS
MSE2.43±0.641.55±0.280.94±0.142.17±0.811.01±0.150.62±0.074.64±0.881.62±0.111.23±0.11
No PE2.07±0.092.67±0.171.11±0.131.64±0.101.53±0.100.62±0.062.25±0.122.74±0.131.23±0.10
Frozen encoder2.17±0.511.86±0.510.95±0.271.59±0.261.21±0.130.60±0.086.18±0.892.08±0.391.69±0.34
JICDE1.84±0.211.81±0.360.86±0.121.39±0.111.16±0.150.54±0.072.33±0.221.99±0.131.10±0.10
Reviewer_wKs9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents JICDE (Joint-domain Implicit Conditional Density Estimation), a conditional density estimator under sparse and ordered/gridded conditioning domains. Traditional Conditional Density Estimators (CDE) encode cross-conditioning sharing through specific family functions, resulting in structural misspecifications that can degrade the conditional density estimation. JICDE proposes the evaluation of the full conditional distribution family {p(y∣x):xT}\{ p(y|x) : x \mathcal{T}\} instead of evaluating p(y∣x)p(y|x) point per point. The model’s architecture passes Gaussian noise and the random Fourier positional encodings to a deconvolutional decoder that serves to propagate feature representations across neighbouring locations in the conditional domain. The training procedure is performed by minimizing the Continuous Ranked Probability Score (CRPS) over the empirical samples, which prevents the network from collapsing to the conditional mean, as would happen with the MSE loss. The method is validated empirically against 6 different methods (Kernel CDE, GP-CDE, DDN, GPs, TDGP, and ConvNP). The datasets used for the comparison consist of one-dimensional synthetic benchmarks, one-dimensional temporal domain (BIA, PDB, and ETT), and two dimensional spatial grid (MNIST, CIFAR, and ImageNet), using NLL and MAE. The results show how JICDE on average achieves the best performances across all the models. JICDE is not uniformly the best as GP is the best in the Gaussian synthetic benchmark and GP-CDE shows a better MAE on ETT. An ablation study against MLP, Transformer and a fixed-Gaussian-kernel variant is done supporting the claim that learning the local interaction helps on non-Gaussian regimes. The method has been tested as a probabilistic surrogate model in Bayesian Optimization on Schwefel and Rastrigin functions, where it converges fastest.

Critical strengths

Adaptive Cross-Condition Information Sharing without prescribed Relational Families Rather than imposing a specific kernel/function, JICDE parametrizes local spatial/temporal interactions via shared deconvolutional filters. JICDE maintains robust performance across smooth, sharp and multimodal examples. This, and the ablation study in Appendix B.6, support the claim that learning the local interaction rule offers a better structural adaptability.

Computational Efficiency and Strong Empirical Scaling Traditional methods (GPs, TDGP or DDN) require longer training time as the spatial resolution increases, and this is a known bottleneck for these types of methods. JICDE demonstrates a better efficiency training under 1 minute in every two-dimensional dataset while achieving the best result across the analyzed metrics.

Critical weaknesses

Transductive Architectural Constraint on Domain Geometry As acknowledged in Appendix E, the deconvolutional generator operates on a fixed, pre-discretized and regular grid scaffold T\mathcal{T}. This makes the framework strictly transductive; it cannot evaluate arbitrary off-grid coordinate queries or extrapolate continuously beyond the defined bounding scaffold without re-specifying and re-training the decoder resolution. In contrast, GP baseline naturally handle continuous, arbitrary query locations. Example: If you train JICDE on a temporal dataset with 360 timestamps and want to query a timestamp at index 180.5, JICDE cannot evaluate them directly without re-specifying the grid geometry or interpolating.

Divergence between Joint Generation Mechanism and Marginal Training Loss Although JICDE's decoder outputs a full joint trajectory in a single forward pass, the training objective (CRPS​) and evaluation metrics (NLL/MAE) evaluate location-wise marginal distributions independently. The authors clarify in Appendix B and E that joint output realizations serve primarily as an architectural vehicle for parameter sharing rather than an identified joint stochastic process. However, without evaluating joint multi-point metrics (e.g., multivariate energy scores or joint path coverage), it remains unverified whether the multi-location correlations produced within a single draw are statistically calibrated or merely an artifact of the convolutional filters.

Evaluation Protocol NLL is computed from a 64-bin histogram of 10.000 samples, with the range set by the minimum and maximum of the model’s own samples. A mode-collapsed/extreme sample can place the ground truth outside the computed range, so the density falls to the numerical floor of ϵ=10−8\epsilon = 10^{-8}. In the GP-CDE result on Table 3 in the Markov benchmarks, the NLL is 18.42±2e−618.42 \pm 2e-6 which equals −ln(10−8)-ln(10^{-8}) and therefore the saturated floor.

Questions for authors

Q1: While the paper correctly notes that JICDE targets location-wise marginals, a key advantage of joint stochastic generators is their ability to produce spatial/temporal trajectories in a single forward pass. Have you evaluated multi-point joint predictive metrics (such as the Energy Score or Variogram Score across pairs/triplets of unobserved locations) to assess whether the joint trajectories generated by JICDE capture realistic spatial/temporal coherence compared to GP or ConvNP?

Q2: The experiments on temporal and image completion use a single sparsity level per dataset (e.g., 10% for time series, 100 pixels for MNIST). How does JICDE perform under extreme observation sparsity (e.g., 1% or 2% observed points)? Is there a critical sparsity threshold where the local receptive fields of the deconvolutional decoder fail to propagate sufficient gradients across unobserved regions?

Q3: In practical spatial applications, observation grids are frequently irregular or irregularly spaced. If unobserved query locations do not align with a regular 1D/2D grid, how significantly does the performance degrade if inputs are mapped onto a regular grid via interpolation before running JICDE?

Q4: Could you report a metric that does not depend on the histogram construction (test time CRPS, calibration and coverage curves), and clarify how the GP-CDE value at the ϵ\epsilon floor should be interpreted?

6 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 3: Fixed-grid design and off-grid prediction

We sincerely thank the reviewer for these constructive comments on the intended scope and practical applicability of JICDE. They highlighted the need to explain our fixed-grid design and its implications for off-grid queries more clearly. In response, we have clarified the task setting and conducted an additional experiment to assess the effects of grid misalignment. We address Weakness 1 and Question 3 together below.

JICDE is designed for probabilistic completion in settings where prediction locations are specified in advance, as in fixed-resolution image completion and regularly sampled time-series completion. In these applications, the objective is to estimate predictive distributions at unobserved locations within a known structured domain. This motivates our grid-based parameterization: aligning condition locations with the decoder's structured axes allows local transformations, weight sharing, and multiscale structure to jointly constrain distribution estimates across the domain.The resulting model directly predicts unobserved grid locations, while its query flexibility is limited by the predefined grid. Arbitrary off-grid queries within the covered domain require additional interpolation or a redefined working grid and model refitting; continuous extrapolation beyond this domain is not supported natively. We will clarify this intended task setting, its architectural motivation, and the associated query limitations in the main text.

To quantify the effect of interpolation-based adaptation, we added a controlled one-dimensional grid-misalignment experiment on BIA, PDB, and ETT. We retain the original observations and train/test split, and shift the working grid by 0.1, 0.3, 0.5, or 0.7 times the original grid spacing. For example, a shift of 0.1 changes [0,...,359] to [0.1,...,359.1]. We preserve the original sparse training mask: the loss is evaluated only at shifted nodes corresponding to the original training indices. Their training responses are interpolated exclusively from the original training observations; no held-out responses are used. We refit the model on each shifted grid and interpolate each generated sample back to the original test coordinates. The unshifted model and all shifted variants are evaluated on the same fixed set of interior test locations against the original measured responses. Mean NLL increases on all three datasets. CRPS exhibits dataset-dependent changes: it decreases from 0.86 to 0.729–0.770 on BIA, but increases by approximately 4.4%–5.9% on PDB and 15.0%–16.0% on ETT.

Mean±SD; lower is better. Offset is measured in units of the original grid spacing; 0 denotes the original grid.

DatasetBIABIAPDBPDBETTETT
OffsetNLLCRPSNLLCRPSNLLCRPS
0.12.288±0.6450.770±0.1741.865±0.4090.565±0.0503.679±0.5681.269±0.096
0.32.342±0.7140.736±0.1291.873±0.4220.572±0.0603.712±0.6541.265±0.101
0.52.328±0.6840.729±0.1261.901±0.4250.564±0.0593.756±0.5011.271±0.107
0.72.313±0.6550.741±0.1551.862±0.4010.571±0.0533.713±0.5431.276±0.084
01.840±0.2100.860±0.1201.390±0.1100.540±0.0702.330±0.2201.100±0.100
AuthorsOpenReview ↗

Response to Weakness 2 and Question 1: Joint generation and joint predictive evaluation

Thank you for highlighting the relationship between joint generation and marginal training. We address Weakness 2 and Question 1 together. We agree that location-wise CRPS constrains the marginal predictive distributions and does not, by itself, ensure that cross-location dependence is correctly identified. The original pointwise NLL and MAE evaluations are also insufficient to assess this property. JICDE's joint-domain generator primarily enables structural and parameter sharing across condition locations, while also inducing a joint predictive distribution. To assess the empirical quality of this induced distribution beyond pointwise prediction, we added multi-point evaluations without changing JICDE's marginal CRPS training objective.

On BIA, PDB, and ETT, we evaluate triplets At,ℓ=t,t+ℓ,t+2ℓA_{t,\ell}={t,t+\ell,t+2\ell} with ℓ∈1,5,10\ell\in{1,5,10}. Lags are measured in the original time indices, and all three locations are held out. JICDE, GP, and ConvNP are evaluated on identical triplets. For each method, the three responses are extracted from the same native joint draw: GP sampling retains the posterior covariance, ConvNP uses its native joint sampling mechanism, and JICDE extracts the corresponding values from a single generated response field. We report two complementary metrics, Energy Score (ES) and Variogram Score (VS), both lower being better. ES assesses overall multivariate predictive quality, while VS evaluates predicted cross-location differences. JICDE achieves the lowest mean ES in all nine dataset–lag settings and the lowest mean VS in eight: all lags on BIA and PDB, and lags 5 and 10 on ETT. The exception is ETT at lag 1, where GP achieves a lower mean VS (0.36) than JICDE (0.43).

These results extend the original marginal evaluation and support the empirical quality of JICDE's predictions on the tested triplets, including their cross-location variation. We will include these results and clarify this distinction in the revised manuscript.

Mean±SD; lower is better. Lags are measured in the original time indices.

Energy Score (ES)

DatasetBIABIABIAPDBPDBPDBETTETTETT
MethodLag 1Lag 5Lag 10Lag 1Lag 5Lag 10Lag 1Lag 5Lag 10
GP2.89±1.012.84±1.012.83±1.022.48±0.572.49±0.572.57±0.584.14±0.733.98±0.534.13±0.64
ConvNP3.83±0.153.76±0.153.69±0.142.69±0.802.72±0.842.69±0.883.66±0.133.78±0.173.79±0.19
JICDE1.77±0.231.76±0.251.78±0.251.14±0.151.14±0.161.16±0.172.36±0.232.39±0.212.40±0.25

Variogram Score (VS)

DatasetBIABIABIAPDBPDBPDBETTETTETT
MethodLag 1Lag 5Lag 10Lag 1Lag 5Lag 10Lag 1Lag 5Lag 10
GP0.24±0.190.81±0.500.88±0.370.16±0.050.63±0.220.76±0.190.36±0.061.00±0.311.14±0.29
ConvNP1.48±0.410.98±0.270.72±0.191.13±1.100.84±1.120.68±0.971.06±0.160.71±0.090.60±0.06
JICDE0.12±0.010.26±0.040.29±0.040.10±0.010.17±0.020.19±0.030.43±0.030.48±0.040.44±0.03
AuthorsOpenReview ↗

Response to Question 2: Performance under extreme observation sparsity

Thank you for pointing out that a single observation ratio is insufficient to characterize sensitivity to sparsity. We have added experiments on BIA, PDB, and ETT with observation ratios of 1%, 2%, 3%, 5%, 7%, 10%, 20%, and 50%, reporting the mean and standard deviation of NLL, MAE, and CRPS.

Both point and distributional prediction generally deteriorate as the observation ratio decreases. In particular, errors under 1%–2% observation are substantially higher than under the original 10% setting. At 1% observation, mean CRPS increases from 0.86 to 2.32 on BIA, from 0.54 to 1.90 on PDB, and from 1.10 to 2.33 on ETT, corresponding to approximately 2.70, 3.52, and 2.12 times the respective 10% values. MAE follows the same overall trend, while NLL generally exhibits greater variability at low observation ratios. These results show that structural sharing does not prevent substantial predictive deterioration when observations are extremely scarce.

We thank the reviewer again for this helpful suggestion, which has strengthened our empirical evaluation and provided a clearer picture of JICDE's performance and limitations across different levels of observation sparsity.

Mean±SD; lower is better. Ratio denotes the percentage of observed locations.

DatasetBIABIABIAPDBPDBPDBETTETTETT
RatioNLLMAECRPSNLLMAECRPSNLLMAECRPS
1%5.62±1.512.93±0.422.32±0.337.07±2.872.33±0.341.90±0.4710.01±2.822.64±0.362.33±0.32
2%5.20±1.502.82±0.362.12±0.305.48±1.762.03±0.181.56±0.217.41±3.432.45±0.422.00±0.55
3%4.30±1.272.61±0.391.85±0.373.36±1.371.72±0.271.15±0.254.28±1.442.20±0.311.54±0.33
5%3.09±1.122.20±0.471.37±0.392.01±0.401.41±0.210.78±0.142.80±0.532.10±0.181.26±0.10
7%2.20±0.391.95±0.370.99±0.151.58±0.161.28±0.230.63±0.092.59±0.322.10±0.221.23±0.13
10%1.84±0.211.81±0.360.86±0.121.39±0.111.16±0.150.54±0.072.33±0.221.99±0.131.10±0.10
20%1.51±0.211.54±0.340.61±0.101.15±0.151.00±0.140.41±0.052.01±0.101.87±0.060.93±0.04
50%1.38±0.251.38±0.330.49±0.141.01±0.140.91±0.140.34±0.031.77±0.031.75±0.050.80±0.02
Reviewer_7LpjOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposed a method to estimate the conditional density p(Y|X). The setting is very different from normal conditional density estimation (CDE) task, where a training set that contains points sampled i.i.d from a joint distribution of P_{XY} is used. To me, it is more like an interpolation/super-resolution method, where

  1. x is on a fixed grid, and y is a numeric value (think about images, where x is the position of the pixel, and y is the pixel value); thus (x,y) samples are not really i.i.d;
  2. Given an image or 1D time series with only a subset of values known, the model is trained on those known points, and then predict the conditional densities of the remaining pixels/time series values.
  3. The model is fit separately on each series or image.
Critical strengths

The method itself is interesting (see my questions though).

Critical weaknesses
  1. There is a serious mismatch between what the paper claimed to solve (CDE) and what the proposed method actually do (interpolation). This is to me very misleading, and hence I don't fully agree how the authors position the paper in the context of CDE. (The appendix does acknowledge a fixed-design setting, but the title, abstract and main text still frame the work as CDE, which to me is misleading).

  2. The empirical evaluation has the following issues as well:

  • It's a mix of "real CDE methods" (e.g., KCDE, which, like standard CDE methods, assumes (x, y) pairs sampled i.i.d. from a joint distribution, an assumption that does not hold in their fixed-grid setting with one dependent observation per location, and hence, to me, are not suitable baselines to compare against in their setting) and interpolation methods (e.g., Gaussian process), yet no interpolation methods that are specifically designed for images are included.

  • When evaluating the NLL, all methods use histograms fitted to samples drawn from each method’s predictive distribution, yet some baselines methods are capable of directly predicting the CDEs (this is different from the proposed method, which can only generate samples, which is also a shortcoming for a CDE method).

Questions for authors
  • What's the status of the interpolation/CDE methods for image pixels? Are there other methods also using convolutional generator to the observed pixels of a single image and can provide per-pixel uncertainty?

  • Can the authors kindly provide some intuitions of the proposed method? It's a bit unclear to me why the specific method is proposed, i.e., which design choices are novel, and what motivates these design choices (e.g., to overcome a certain shortcoming of existing method).

6 comments
AuthorsOpenReview ↗

Response to Weakness 1: CDE framing and its relationship to interpolation

Thank you for pointing out this issue with the positioning of our work. We agree that our fixed-design setting differs from conventional CDE based on i.i.d. random-design samples. The original title, abstract, and main text did not sufficiently emphasize this distinction, and we apologize for the resulting confusion. However, a fixed design does not in itself preclude CDE. Efromovich (2007) explicitly studies conditional density estimation under fixed designs; Fong et al. (2023) consider families of conditional densities indexed by fixed design points; and Foster et al. (2024) include CDE within a fixed-design estimation framework. We cite these works to clarify that i.i.d. sampling from a joint distribution under a random design is not a necessary condition for formulating a CDE problem.

Our intended scope is distribution estimation at fixed conditioning locations from sparse observations. JICDE fits a shared stochastic generator to the current instance using marginal CRPS, and repeated latent sampling yields a predictive distribution at each location. To make the observed-data context explicit, we will denote the fitted distributions by F^θ^(D)(⋅∣xi)\widehat F_{\widehat\theta(\mathcal D)}(\cdot\mid x_i), where D\mathcal D is the instance's observed subset. This notation distinguishes the conditioning location from the data used to fit the estimator. These are model-implied marginal predictions; their estimation relies on structural sharing across locations. Accordingly, our use of CDE refers specifically to estimating location-wise response distributions in this fixed-domain setting, with probabilistic completion/interpolation as its application. This terminology does not imply that our setting is interchangeable with conventional random-design CDE.

We will revise the scope stated in the title and abstract, specify the setting at the beginning of the main text, and align the experimental interpretations and conclusions accordingly. We have added DIP and Bayesian DIP as the baselines of image interpolation task to address your suggestion regarding image-specific restoration methods; their results are discussed in our combined response to Weakness 2 and Question 1.

References:

[1] Efromovich, S. (2007). Conditional density estimation in a regression setting. The Annals of Statistics, 35(6), 2504-2535.

[2] Fong, E., Holmes, C., and Walker, S. G. (2023). Martingale posterior distributions. Journal of the Royal Statistical Society: Series B, 85(5), 1357-1391.

[3] Foster, D. J., Han, Y., Qian, J., and Rakhlin, A. (2024). Online Estimation via Offline Estimation: An Information-Theoretic Framework. NeurIPS.

AuthorsOpenReview ↗

Response to Weakness 2 (first point) and Question 1: Baseline selection and image completion

Thank you for pointing out the differences between the statistical assumptions underlying KCDE and our task setting. We retain KCDE as a local-smoothing baseline directly relevant to our motivation: it allows us to assess distance-based sharing across conditioning locations under the same observation conditions. We would like to clarify that i.i.d. sampling is a common assumption in classical analyses of KCDE, rather than a prerequisite for evaluating its kernel-weighted predictive density. Kernel conditional density estimation has also been studied for dependent data, for example by Hyndman and Yao (2002). We therefore consider KCDE an informative baseline for evaluating distance-based sharing under sparse observations. Nevertheless, it is not an image-specific restoration method and cannot replace comparisons with such methods.

Regarding Question 1, yes: existing methods fit convolutional generators to single partially observed images and provide pixel-wise uncertainty estimates. Our original manuscript did not sufficiently discuss these directly related approaches or include the corresponding reconstruction baselines. We have therefore added the relevant comparisons and will expand the related-work discussion and clarify the scope of the image experiments. Deep Image Prior (DIP; Ulyanov et al., 2018) reconstructs an image by optimizing a convolutional network on a single degraded observation. Cheng et al. (2019) subsequently used stochastic gradient Langevin dynamics for approximate posterior sampling, providing reconstructions and uncertainty estimates for tasks including inpainting. Further extensions include Bayesian DIP with Monte Carlo dropout (Laves et al., 2020), mean-field variational inference (Tölle et al., 2021), and linearised DIP (Antorán et al., 2023). DIP primarily optimizes a single reconstruction, while Bayesian DIP variants that infer network weights construct predictions through a weight posterior. JICDE instead resamples latent variables with the fitted network parameters held fixed, using the location-wise marginals of generated outputs as predictive distributions.

Deep Probabilistic Imaging (DPI; Sun and Bouman, 2021) is more closely related in that it also represents image distributions through latent-variable sampling. DPI uses an invertible generative model and a variational objective incorporating the imaging likelihood, image prior, and distributional entropy to approximate a joint image posterior. JICDE uses shared local transformations over the conditioning domain and marginal CRPS to directly fit location-wise predictive distributions, without evaluating the generator density. Another related line of work concerns single-image generation and diverse completion. SinGAN (Shaham et al., 2019) learns multiscale patch distributions from a single complete image, while IDC (Alkobi et al., 2023) and PSinGAN (Miyata, 2024) generate diverse completions from partially observed images. These approaches are relevant to our image tasks, although diverse completion and the statistical evaluation of pixel-wise predictive distributions are distinct objectives. We will clarify their training objectives, observation settings, and evaluation priorities in the related-work discussion.

We have added comparisons with DIP and Bayesian DIP, together with DDPM and conditional flow matching (CFM). We report MAE, CRPS, and NLL at unobserved locations for probabilistic methods. DIP achieves the lowest mean MAE on CIFAR and ImageNet, with JICDE ranking second, whereas JICDE achieves the lowest mean MAE on MNIST. For distributional prediction, JICDE achieves the lowest mean CRPS across all three image tasks. Bayesian DIP achieves the lowest mean NLL on MNIST, while JICDE achieves the lowest mean NLL on CIFAR and ImageNet. JICDE also outperforms the evaluated DDPM and CFM baselines in terms of mean performance on all reported metrics across these three tasks. These results support the competitiveness of JICDE relative to the image reconstruction methods evaluated in our setting.

Mean±SD; lower is better. BDIP=Bayesian DIP; N/A=not applicable.

DataMethodNLLMAECRPS
MNISTDIPN/A0.064±0.039N/A
MNISTBDIP-6.17±0.540.051±0.0310.045±0.032
MNISTDDPM-0.64±0.540.154±0.0320.114±0.018
MNISTCFM-1.48±0.620.102±0.0080.094±0.008
MNISTJICDE-2.64±0.350.050±0.0100.038±0.008
CIFARDIPN/A0.023±0.003N/A
CIFARBDIP-1.72±0.090.039±0.0010.022±0.002
CIFARDDPM-0.78±0.320.080±0.0110.041±0.005
CIFARCFM-1.03±0.200.053±0.0030.036±0.003
CIFARJICDE-1.84±0.100.030±0.0100.021±0.002
ImageNetDIPN/A0.031±0.001N/A
ImageNetBDIP-1.45±0.070.045±0.0010.028±0.001
ImageNetDDPM-1.11±0.120.113±0.0100.053±0.003
ImageNetCFM-1.35±0.080.087±0.0030.044±0.002
ImageNetJICDE-2.14±0.040.040±0.0010.022±0.001
AuthorsOpenReview ↗

Response to Weakness 2 (second point): Predictive density evaluation and histogram-based NLL

Thank you for highlighting this limitation. JICDE provides an implicit conditional predictive distribution through samples, rather than a directly evaluable predictive density. We will state this limitation explicitly in the main text and distinguish sampling from density evaluation. Our original histogram-based NLL evaluation provided a common procedure for explicit and implicit models. However, replacing a baseline's predictive density with a sample histogram introduces additional sampling and density-reconstruction errors.

We have therefore added two complementary evaluations. First, we report model-based predictive NLL for GP, TDGP, KCDE, GP-CDE, and ConvNP, and will specify the numerical approximations used where applicable. JICDE's NLL remains a histogram-based estimate; it should not be interpreted as an exact evaluation of its predictive density. Second, we report test CRPS computed directly from predictive samples for all methods, without histogram construction or density reconstruction.

The synthetic experiments assess predictive performance across different distributional structures. On Gaussian data, GP achieves the lowest mean NLL and CRPS, consistent with its well-matched prior, while JICDE ranks second on both metrics. On Markov data, JICDE achieves the lowest reported mean NLL, whereas its CRPS (0.34) is higher than those of TDGP (0.28) and GP (0.31). On Double Sine, JICDE achieves the lowest reported mean NLL and CRPS. On BIA, PDB, and ETT, JICDE also achieves the lowest reported mean NLL and CRPS among the evaluated methods. The CRPS results thus support competitive distributional prediction independently of histogram construction, while showing that JICDE does not uniformly outperform all baselines across synthetic settings.

Mean±SD; lower is better. †Model-based predictive NLL; other NLLs are histogram estimates. All CRPS values are computed from predictive samples.

Synthetic tasks:

DatasetGaussianGaussianMarkovMarkovDouble SineDouble Sine
MethodNLLCRPSNLLCRPSNLLCRPS
GP†0.51±0.860.24±0.058.99±13.490.31±0.049.31±0.000.63±0.0003
TDGP†1.24±0.110.38±0.050.69±0.040.28±0.022.89±0.000.50±0.0002
ConvNP†25.35±24.750.80±0.186.43±5.690.47±0.0925.63±28.550.45±0.01
KCDE†2.08±1.430.35±0.060.33±0.580.34±0.040.47±5e-50.36±5e-4
GP-CDE†1.71±0.520.44±0.130.73±0.0020.50±0.0012.42±0.300.69±0.04
DDN10.91±1.260.42±0.093.13±3.350.46±0.061.41±0.460.41±0.0002
DDPM2.92±0.872.23±1.882.44±0.930.98±0.331.20±0.100.45±0.02
CFM2.14±0.490.72±0.072.69±0.820.66±0.101.07±0.010.42±0.00
CSDI1.69±1.050.29±0.035.83±1.500.40±0.029.93±0.780.79±0.03
JICDE0.99±0.560.25±0.03-0.03±0.740.34±0.04-0.08±0.100.34±0.0003

Time-series tasks:

DatasetBIABIAPDBPDBETTETT
MethodNLLCRPSNLLCRPSNLLCRPS
GP†2.85±0.671.29±0.482.64±0.651.09±0.262.81±1.851.32±0.12
TDGP†3.34±0.023.75±0.022.52±0.491.85±0.982.34±0.241.23±0.16
ConvNP†2.96±0.272.06±0.162.47±0.351.46±0.393.33±1.691.83±0.15
KCDE†2.27±0.061.37±0.101.84±0.040.87±0.032.35±0.051.35±0.11
GP-CDE†2.81±0.031.96±0.262.31±0.041.17±0.223.27±0.201.33±0.17
DDN2.21±0.291.14±0.111.88±0.150.91±0.092.55±0.241.34±0.09
DDPM2.83±0.770.97±0.232.88±0.750.69±0.083.83±0.691.43±0.13
CFM2.35±0.600.93±0.232.74±0.660.68±0.083.88±0.621.45±0.18
CSDI2.97±1.671.12±0.252.16±1.210.68±0.083.13±1.491.56±0.37
JICDE1.84±0.210.86±0.121.39±0.110.54±0.072.33±0.221.10±0.10

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.