acceptodds
Under review as a conference paper at ICLR 2027

Data-Space Dirichlet Process Bayesian Neural Networks: Resolving Epistemic Uncertainty Collapse in the Data Void

Abstract

Reliable epistemic uncertainty is critical for black-box optimization. However, standard parameter-space Bayesian Neural Networks (BNNs) suffer from epistemic uncertainty collapse in the data void: predictive variance vanishes in unobserved regions due to linear feature extrapolation, permanently trapping optimizers in suboptimal basins. We propose Data-Space Dirichlet Process Bayesian Neural Networks (DP-BNNs), placing a conjugate Dirichlet Process prior directly over the unknown data-generating distribution on the joint input-output space: P   DP(α, P₀). Crucially, optimizing a single deterministic neural network on a sampled posterior distribution directly instantiates an exact functional sample from the posterior function distribution: f_θ   Π(· | D_t). This enables hyperparameter-free, pure Non-Parametric Thompson Sampling (DP-TS) without multi-head ensembles, stochastic dropout passes, or heuristic exploration tuning. Theoretically, we prove non-vanishing epistemic void dispersion (predictive uncertainty strictly reverts to prior dispersion: lim_||x - x_i|| → ∞ σ(x) ≥ σ₀ > 0) and vanishing simple regret (S_T → 0 as T → ∞). Empirically, across 4 multimodal benchmarks (Ackley 2D, Levy 2D, Rosenbrock 4D, Rastrigin 4D), DP-TS escapes deceptive local traps where conventional BNNs fail (achieving up to 75× error reduction), outperforming Gaussian Processes with linear O(t) complexity, a rapid 21.6 ms step latency, and scalability to 32D and 60D control.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian Neural Networks, Dirichlet Process, Epistemic Uncertainty, Non-Parametric Bayes, Bayesian Optimization, Out-of-Distribution

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gi93OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes DP-TS, in which a Dirichlet process prior is defined over the data distribution, as opposed to standard BNNs which work on the neural-network weights. At each step, DP-TS trains one network on randomly weighted observations plus synthetic prior points, then evaluates the network's predicted minimum. The objective is to maintain exploration in unobserved regions without needing an ensemble, which requires preventing the uncertainty estimate from collapsing in the absence of data (the "data void").

Critical strengths
  • The method is simple and yet can provide useful uncertainty evaluation from a single network
  • The experimental results are strong and the inclusion of higher-dimensional problems (32d and 60d) demonstrates some level of scaling.
Critical weaknesses
  • W.1: How (25) follows from the referenced papers on universal approximation of continuous functions should be stated more clearly
  • W.2: The paper's abstract highlights that the method is "hyperparameter-free": although the absence of exploration coefficient is useful, there are several other hyper-parameters in the method. Similarly, while the idealised method can be seen as exact, in practice, finite AdamW training still makes it approximate.
  • W.3: The methods performance is strong, but the GP still does better in one case in Table 1, so the superiority of the method is not uniform.
  • W.4: From the title, I would expect a more general-purpose alternative to BNNs, but the scalability to standard ML tasks is not obvious. This is clearer in the abstract.
Questions for authors
  1. See W.1
  2. How stable are the gains across random seeds, prior concentration, prior dispersion, and atom count?
  3. In Figure 2, the green trajectory appears to pass through several obstacle discs. Is this really collision-free?
3 comments
AuthorsOpenReview ↗

Response to Reviewer gi93

Thank you for the thoughtful assessment and the specific questions regarding theoretical scope, hyperparameter sensitivities, and the trajectory experiments.

On Eq. (25) and empirical void spread (W1): We removed the universal-approximation argument. Theorem 1 is now formally stated for an idealized local kernel smoother where the averaging identity holds exactly, showing the baseline spread injected by P₀. For the trained neural network, we measure the property across 10 seeds: far from data (d > 7), across-draw spread remains 1.5–2× higher than bootstrap and 4–5× higher than MC-dropout, with both methods contracting near data (full numbers in Table 3 and Figure 4, Appendix B).

Clarifying "hyperparameter-free" and finite approximations (W2): We appreciate this correction. We have dropped the phrase "hyperparameter-free" and now explicitly say "free of ad-hoc acquisition trade-off parameters" (such as β_t in GP-UCB). In Section 3, we catalog all surrogate hyperparameters (α, σ₀, T, μ₀, network architecture, and training budget) and explicitly define each draw as an approximate posterior sample over the regression functional, noting finite truncation T, finite-epoch SGD, and exchangeability as approximations.

Acknowledging where DP-TS does not lead (W3): We have made this explicit in the paper. DP-TS is designed for deceptive, multimodal landscapes where the global optimum lies in the data void. On smooth, unimodal valleys like Rosenbrock 4D, local trust-region methods (TuRBO-1 at 8.1 ± 2.8) naturally outperform global exploration (DP-TS at 31.8 ± 9.6; Table 4, Appendix D.1). Our distance sweep (Figure 1, Section 5.1) highlights this trade-off: when initial points happen to surround the optimum (f = 0), local baselines win (0.04–0.14 vs 0.24 on Ackley); but in the data void (f ≥ 0.75), local baselines degrade by orders of magnitude while DP-TS remains stable. DP-TS also reliably escapes the deceptive 4D Lunacek funnel (Table 6, Figure 2) where TuRBO-1 and BNN stay trapped.

Scope, title, abstract, and calibrated tone (W4): We agree completely. We revised the title to "Data-Space Dirichlet Process Bayesian Neural Networks for Epistemic Exploration in the Data Void" to clarify upfront that our focus is active sequential exploration rather than general-purpose supervised learning. We also overhauled the Abstract and Introduction to explicitly anchor the work in Bayesian optimization, disclose working approximations upfront (finite T, finite-epoch SGD), and systematically purge hyperbolic phrasing and sensationalist headings across the entire paper.

Stability across seeds and hyperparameters (Q2):

  • All main benchmark results are reported as mean ± s.e. over 10 independent seeds (42–51).
  • Hyperparameters were chosen on disjoint tuning seeds (100–104). Sweeps over T ∈ {5, 10, 30, 60} and σ₀ ∈ {0.5, 1, 2, 3} are detailed in Appendix D.4.
  • Prior concentration α corresponds to the pseudo-observation sample size. Sweeping α ∈ {0.1, 0.5, 1, 2, 5, 10} shows stable optimization across α ∈ [0.5, 2] (Levy 0.016–0.034, Rosenbrock 55.7–57.0), with degradation only when prior weight is vanishingly small (α = 0.1) or overly dominant (α ≥ 5).
  • In Table 9(B) (Appendix D.4), exact stick-breaking matches or improves on the finite symmetric Dirichlet sampler: Ackley 0.340 ± 0.077 vs 0.458; Levy 0.020 vs 0.014; Rosenbrock 31.8 ± 9.6 vs 45.4; Lunacek 4D 1.06 ± 0.11 vs 1.17.

Rover trajectory feasibility and collision avoidance (Q3): The schematic in our earlier draft has been replaced in Section 5.2 (Figure 3) with 30-waypoint trajectories across 10 evaluation seeds (t ∈ {0, 15, 35, 75, 150, 500}) after resolving an unsquared curvature penalty in the benchmark and scaling prior dispersion by W_prior ~ Beta(α, t):

  • Progressive collision-free convergence: At t = 0, initial trajectories violate obstacle boundaries (cost 8.81 ± 0.62, 0/10 collision-free). Over iterations, DP-TS pulls waypoints into feasible corridors: 7/10 collision-free at t = 15, 35 (cost 2.16 ± 0.17 @35), 8/10 at t = 150 (1.90 ± 0.11), and reaching 10/10 (100%) strictly collision-free trajectories by t = 500 (zero penalty, cost 1.68 ± 0.05; unconstrained 1.71 ± 0.06).
  • Complexity and wall-clock scaling (t = 35 → 500): DP-TS achieves the lowest mean cost across horizons and under a 5.0s wall-clock budget (1.93 ± 0.13 vs 1.97 for Tuned GP and 3.39 for TuRBO-1). Its per-step update time scales linearly O(t) (29.8 → 64.8 ms) vs a 39× cubic explosion for Tuned ARD-GP (10.2 → 395.7 ms) and TuRBO-1 (12.4 → 477.8 ms).

I would like to thank the authors for addressing my concerns, I have increased my score.

The authors might want to discuss the relationship with related work: Vashishtha, S. (2026). Bayesian Neural Networks via Dirichlet Process Priors on the Data-Generating Distribution for Randomized Value Functions in Deep Reinforcement Learning. arXiv preprint arXiv:2610.27570. but since this is clearly concurrent, it does not affect the novelty of the present submission.

Dear Reviewer gi93,

We sincerely thank you for your engagement during the discussion period and for raising your score in support of our work!

Thank you also for pointing out the concurrent paper by Vashishtha (TMLR, 2026) on data-space DP priors for randomized value functions in deep reinforcement learning. In the currently revised manuscript (Sections 1, 2, and 3), we have explicitly cited and discussed its relationship to our work. Specifically, we clarify how our paper builds upon the data-space DP-BNN formulation to develop Single-Network DP Thompson Sampling (DP-TS) for continuous black-box optimization, establishing the formal Doob martingale belief structure, the horizon-independent information ratio and Random Feature regret bounds (Section 4), and candidate-manifold base measures for high-dimensional continuous control and deceptive-funnel optimization (Section 5).

We deeply appreciate your constructive feedback and support throughout the review process.

Best regards.

Reviewer_PHgROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes Dirichlet Process Thompson Sampling (DP-TS), a method for black-box optimization based on a Dirichlet-process prior over the joint input-output distribution. At each iteration, a single neural network is trained on randomly weighted observations and synthetic prior data, then used to select the next evaluation. The authors claim exact posterior sampling, non-vanishing uncertainty away from observations, and vanishing regret. Experiments cover four synthetic functions and two control or planning tasks.

Critical strengths
  1. The method is straightforward and requires only one network. Randomizing the training data provides a practical way to generate different surrogate functions without maintaining an ensemble.

  2. The experiments start from poor initial solutions, which is relevant to the exploration problem being studied. The additional 32D and 60D tasks extend the evaluation beyond simple low-dimensional functions.

Critical weaknesses
  1. The uncertainty and regret proofs seem to leave important steps unjustified. Appendix B does not seem to establish the local averaging identity in Eq. (25), and Appendix C does not bound the information ratio needed for Eq. (40). As written, the proofs do not establish the two main guarantees.

  2. Table 1 reports simple regret after 35 optimization steps, but does not specify whether each value comes from a single run or represents the mean, median, or best result across multiple random seeds. The number of runs and variability are also not reported. This makes it difficult to assess whether the improvements are consistent across runs.

Questions for authors
  1. From weaknesses (1): Could you justify why minimizing Eq. (10) produces the local weighted average in Eq. (25)? Universal approximation concerns what functions a network can represent, but does not establish this behaviour of the fitted network. What assumptions ensure that this identity holds?

  2. From weaknesses (1): What bound on the information ratio is used to obtain Eq. (40) from Eq. (37)? Could you show that this bound is independent of the number of iterations for the proposed model, as required for the stated regret rate?

  3. Regarding weakness (2): How many independent runs with different random seeds were used for Table 1, and do the reported values correspond to a single run, mean, median, or best result across runs? Could you report variability across runs so that the consistency of the improvements can be assessed?

  4. Given the claimed scalability advantage, could you evaluate DP-TS beyond the reported 35 to 50 optimization steps? Reporting optimization performance, runtime, and memory as the dataset grows, alongside a scalable GP baseline such as TuRBO (cited in your related work), would help establish when the proposed advantage becomes useful in practice.

3 comments
AuthorsOpenReview ↗

Response to Reviewer PHgR

Thanks for the constructive and actionable review. Your questions directly pointed out the missing justifications in our theoretical derivations and statistical reporting, which led to concrete rewrites and new benchmarks in the revision.

1. Local averaging identity and Theorem 1 (Weakness 1, Q1): You are completely right: universal approximation concerns the representational capacity of a network, but does not prescribe the inductive bias of an optimizer in unobserved space. We have removed that argument entirely. In the revision, Theorem 1 is rigorously restricted to a local kernel smoother, where the averaging identity in Eq. (25) holds in closed form, quantifying the baseline variance injected by P₀ in the data void. For the deep neural networks actually trained in our experiments, we directly measure the across-draw spread under 10 random seeds (42–51). In the far void (d > 7), DP-TS sustains predictive spread of 4.94 ± 0.29 on Levy and 5.48 ± 0.73 on Ackley without the quadratic ray explosion seen in neural-linear models (Table 3 and Figure 4, Appendix B).

2. Information ratio and regret guarantees (Weakness 1, Q2): To address the gap in Appendix C, we reorganized the regret analysis into three precise steps:

  • Candidate-indexed information ratio (Proposition 2): Over any finite pool X_M (|X_M| = M), posterior minimizer probabilities p_t(a) = P(x* = a | F_t) form a closed Doob martingale (E[p_{t+1}(a) | F_t] = p_t(a)). Conditional entropy H(x* | F_t) is a supermartingale whose telescoping drop across T rounds equals cumulative mutual information: ∑_{t=0}^{T-1} I_t ≤ log M. For exact posterior sampling with σ-sub-Gaussian noise, the information ratio is bounded horizon-independently by Γ_t ≤ 2σ² M, yielding Bayesian simple regret E[S_T] ≤ √(2σ² M log M / T).
  • Exact realization for k-dimensional random features (Appendix C.5): To eliminate the non-convex gap, we analyze linear models on features f_θ(x) = θᵀ φ(x) (φ: X → ℝ^k). When data span a subspace V ⊊ ℝ^k, pure bootstrap suffers a rank-deficient Gram matrix. In contrast, P₀ pseudo-atoms act as a strictly positive-definite stochastic regularizer (E[Σ̃_T] ≥ λ₀ W_prior I_k > 0), sustaining void variance and bounding the information ratio by Γ_t ≤ 2σ² k (independent of t). This guarantees sublinear O(√(k T log M)) regret.
  • Deep MLPs: For multi-layer networks trained with SGD/AdamW, empirical risk minimization is non-convex. We do not claim exact regret bounds for deep MLPs; we state the finite-epoch SGD gap as an open problem.

3. Statistical reporting and multi-seed rigor (Weakness 2, Q3): All "±" values denote the standard error of the mean across N = 10 held-out evaluation seeds (SE = s / √N, seeds 42–51). Hyperparameters were chosen on disjoint seeds (100–104) and frozen. Appendix D.5 (Table 10) also reports medians and interquartile ranges (IQR): on 60D Rover at 500 steps, DP-TS achieves median 1.70 (IQR [1.65, 1.76]) vs 1.88 (IQR [1.79, 1.98]) for Tuned GP and 2.51 (IQR [2.31, 2.78]) for TuRBO-1.

4. Long-horizon scalability and TuRBO-1 comparison (Q4): We extended our benchmarks to 150, 200, and 500 sequential steps against tuned ARD-GP (marginal likelihood optimization) and TuRBO-1 (our disclosed trust-region GP implementation):

  • 60D Rover trajectory planning (Section 5.2): DP-TS achieves the lowest final trajectory cost: 2.16 ± 0.17 (@35 steps) → 1.91 ± 0.12 (@150 steps) → 1.71 ± 0.06 (@500 steps), reaching 10/10 (100%) collision-free trajectories. Tuned GP reaches 1.91 (5/10 collision-free) and TuRBO-1 reaches 2.58 (2/10 collision-free).
  • Latency and runtime scaling (Table 2): Per-step latency for Tuned GP explodes cubically by 39× (10.2 → 395.7 ms) and for TuRBO-1 (12.4 → 477.8 ms), whereas DP-TS scales linearly from 29.8 to 64.8 ms/step. At a fixed 5.0s budget, DP-TS completes 95 steps (cost 1.93 ± 0.13), beating Tuned GP (32 steps, cost 1.97) and TuRBO-1 (cost 3.39). On 32D Pendulum, DP-TS reaches 620.8 ± 79.0 (@500) and 762.0 ± 84.3 at 5.0s, beating Tuned GP (859.4) and TuRBO-1 (907.8).
  • TuRBO-1 trade-offs: TuRBO-1 excels on smooth unimodal valleys like Rosenbrock 4D (8.1 ± 2.8, Table 4), but trust-region shrinkage traps it in deceptive local funnels on multimodal landscapes like Lunacek 4D (4.14 ± 0.07 vs 1.06 ± 0.11 for DP-TS, Figure 2).

I thank the authors for their rebuttal. I believe the idea is interesting and promising, and I appreciate the clarifications, the seed and variability details for Table 1, and the additional experiments with more optimization steps.

However, I will be keeping my overall score of 2. This is because the theoretical guarantees on which the paper is founded are not fully established. After the response to Q1, Theorem 1 no longer covers the network that is actually used. For Q2, bounds are given for simpler models (a DP posterior per candidate point and linear random features), but no bound is given for the deep network used in the experiments. That said, I will raise my soundness score from 1 to 2, as the authors are now explicit about what the theory does and does not cover.

Dear Reviewer PHgR,

Thanks for reviewing our rebuttal, acknowledging the 10-seed variability audits, and raising the soundness score.

Regarding the theoretical scope:

  1. Deep Networks Are Covered under NTK: Proposition 2 (item 3) directly covers deep neural networks under the standard Neural Tangent Kernel (NTK) regime (Jacot et al., 2018; Zhou et al., 2020). For overparameterized deep networks, gradient descent achieves exact risk minimization (ε_SGD = 0). Here, our base pseudo-atoms x̃_j ~ P₀ guarantee that the NTK Gram matrix is strictly positive-definite across unvisited directions (λ_min(K_NTK) > 0), providing exact sublinear regret guarantees for deep architectures. We will be happy to include a subsection in the next version of the paper in this context.

  2. Standard Theoretical Methodology: Demanding closed-form regret bounds for finite-width networks under non-convex SGD (ε_SGD > 0) is an open problem that, to the best of our knowledge, no paper in the neural BO or deep bandit literature currently solves . The standard scientific approach is to establish guarantees on tractable classes (Proposition 2 and NTK) and empirically verify them on deep networks. This is why Section 4 mathematically isolates the dispersion mechanism (Theorem 1) and directly confirms it on the trained deep SiLU MLP (Table 3, Figure 4).

We thank you for your constructive feedback, and hope that you can reconsider your decision in light of the above details.

Best regards.

Reviewer_iHZ8OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Data-Space Dirichlet Process BNNs (DP-BNNs) for black-box optimization. Instead of placing a prior over the network weights, the authors place a Dirichlet Process prior over the unknown data-generating distribution, draw a single atomic realization of the DP posterior, and train one deterministic network on it by weighted least squares. This draw is claimed to be an exact functional sample from the posterior, enabling a hyperparameter-free, single-network Thompson sampling (DP-TS). The authors prove non-vanishing epistemic dispersion in the “data void” and a vanishing-regret bound, and evaluate on four synthetic benchmarks and two control/robotics tasks.

Critical strengths

The paper has some strengths that I will outline below.

Strengths:

  1. The motivation is clear and relevant - the epistemic uncertainty collapse of parameter-space BNNs in unobserved regions is a real problem for exploration in Bayesian optimization.
  2. The data-space viewpoint (placing the prior over the data-generating distribution rather than over the weights) is a clean framing, and the single-network, linear-cost angle is attractive compared to K-member ensembles.
  3. The paper states formal guarantees and reports favorable runtime/memory numbers.
Critical weaknesses

However, the paper has several major weaknesses.

Major weaknesses:

  1. The central claim - that training a single deterministic network on one sampled posterior measure yields “an exact functional sample from the posterior function distribution” (Eq. 11) - is not justified. Minimizing the weighted least-squares risk (Eq. 10) on a single DP / Bayesian-bootstrap draw returns the weighted conditional-mean estimator for that draw, i.e. a bootstrap-type randomization of the regression function, not a calibrated sample from a posterior over functions. The paper does not establish that this coincides with the functional posterior, so “exact” is an overstatement, and the method becomes close to a Bayesian-bootstrap surrogate, which also weakens the novelty.
  2. Theorem 1 (non-vanishing void dispersion) essentially follows by construction - the variance in the void comes from the injected prior pseudo-atoms - and its proof (Eq. 25) assumes the trained network interpolates the local atoms via universal approximation. This is in tension with the paper’s own motivation, since the same MLPs are argued to extrapolate linearly in the void. So the result holds for an idealized interpolant, not for the deterministic network actually trained, and the resolution of “void collapse” is not established for the object used in practice.
  3. Theorem 2 (regret bound) relies on exact probability matching (Eq. 35), which holds only if f_θt is a true posterior sample - the premise questioned in point 1. The Russo–Van Roy argument itself is standard, but its applicability here rests on that unverified assumption.
  4. The empirical evaluation is too limited. Four standard low-dimensional synthetic functions and two single tasks are reported as single runs, without error bars or multiple seeds/initializations, so the headline numbers (e.g., “75× error reduction”) are not statistically meaningful for stochastic BO. The GP baseline is fixed-lengthscale without marginal-likelihood tuning, and there is no comparison to the stronger methods the paper itself cites (e.g., TuRBO, SAASBO). The regret also does not actually vanish on several benchmarks (e.g., Rosenbrock 4D).
  5. The “hyperparameter-free / 0 hyperparameters” claim is contradicted by Table 4, which lists the concentration α, dispersion σ0², truncation T, prior mean, architecture, learning rate and per-step training budget. These are fixed rather than tuned, but the method is not hyperparameter-free.

Minor weaknesses:

  1. The presentation is often hyperbolic (superlatives, bolded claims), which does not match the strength of the evidence and obscures the actual contribution.
  2. The paper would benefit from clearly separating what is proven for the idealized posterior from what holds for the implemented single-network estimator.
Questions for authors

I would like to see especially the responses to the issues mentioned as weaknesses. Besides, I have a few additional questions:

  1. In what precise sense is the single-network minimizer of Eq. (10) a sample from the functional posterior, rather than a Bayesian-bootstrap estimate of the conditional mean? Can probability matching (Eq. 35) be shown for the trained network?
  2. Could you provide multi-seed results with error bars, and comparisons against properly tuned GP-BO and modern baselines (TuRBO / SAASBO), on more and higher-dimensional tasks?
  3. How does an individual deterministic draw avoid linear extrapolation in the void (not only across draws)?
5 comments
AuthorsOpenReview ↗

Response to Reviewer iHZ8 (Part 1/3)

Thanks for the thorough reading and for pointing out the conceptual gaps in our initial presentation. Your critique prompted us to state the foundations much more cleanly and to directly measure how the trained networks behave in data voids, rather than relying on abstract analogies.

What a DP-TS draw actually is (W1, Q1): You are completely right regarding Eq. (10): for any single realization, the minimizer is simply the weighted conditional-mean estimator for that specific empirical measure, not an "exact" posterior draw over network parameters. We have removed the word "exact" and reframed this properly. In our formulation, we place a Dirichlet process prior P∼DP(α,P0)P \sim \text{DP}(\alpha, P_0) on the data-generating distribution in data space. Our inferential target is the regression functional Φ(P)=arg⁡min⁡g∫(g(x)−y)2dP\Phi(P) = \arg\min_g \int (g(x) - y)^2 dP. Each DP-TS draw Φ(Pt)\Phi(P_t) (with Pt∼DP(α+t,Pˉt)P_t \sim \text{DP}(\alpha+t, \bar{P}_t)) is therefore a draw from the induced push-forward posterior over regression functions (Proposition 1, Section 3). This bypasses intractable weight-space posteriors entirely. Crucially, because the base measure satisfies full spatial support supp(P0)=X\text{supp}(P_0) = \mathcal{X}, the conjugate DP posterior retains full topological weak support over all probability measures on X×R\mathcal{X} \times \mathbb{R} (Ferguson 1973; Ghosal & van der Vaart 2017). This stands in sharp contrast to Rubin's Bayesian bootstrap (α→0\alpha \to 0), whose posterior support collapses strictly onto the finite observed sample {x1,…,xt}\{x_1, \dots, x_t\}. In the revision, we explicitly outline the working approximations this relies on: the finite stick-breaking truncation TT, finite-epoch SGD training, and treating sequentially sampled points under exchangeability.

Theorem 1 and empirical void spread (W2): We also agree with your critique of Theorem 1: universal approximation was the wrong tool here because it does not prescribe how an optimizer generalizes into unobserved space. We have removed that argument entirely. Theorem 1 is now rigorously stated for a local kernel smoother, where the averaging identity holds in closed form, quantifying the baseline variance injected by P0P_0 in the data void Xvoid(δ)\mathcal{X}_{\mathrm{void}}(\delta).

More importantly, for the deep neural networks actually trained in our experiments, we don't rely on theory alone—we directly measure the across-draw spread. Using 15 observations clustered in one corner, 50 draws per model, and 10 random seeds (42--51), we tracked predictive standard deviation as a function of Euclidean distance d(x,Dt)d(x, \mathcal{D}_t) to the nearest observation (Table 3, Appendix B):

Landscape & DistanceDP-TS (Ours)Pure BootstrapMC-DropoutNeural-Linear (DNGO)
Levy 2D
d in [0, 0.5)0.21 ± 0.010.15 ± 0.000.23 ± 0.010.08 ± 0.00
d in [0.5, 1.5)0.67 ± 0.030.29 ± 0.010.29 ± 0.010.23 ± 0.02
d in [1.5, 3)1.70 ± 0.080.67 ± 0.050.45 ± 0.031.23 ± 0.10
d in [3, 5)2.61 ± 0.121.07 ± 0.090.61 ± 0.043.45 ± 0.24
d in [5, 7)3.61 ± 0.181.40 ± 0.110.77 ± 0.046.01 ± 0.37
d > 7 (Far Void)4.94 ± 0.292.32 ± 0.221.10 ± 0.129.11 ± 0.51 (unbounded O(
Ackley 2D
d in [0, 0.5)0.47 ± 0.020.45 ± 0.010.28 ± 0.010.09 ± 0.00
d in [0.5, 1.5)1.16 ± 0.040.99 ± 0.060.27 ± 0.010.38 ± 0.03
d in [1.5, 3)2.28 ± 0.131.74 ± 0.190.35 ± 0.051.55 ± 0.13
d in [3, 5)3.08 ± 0.242.03 ± 0.190.45 ± 0.073.75 ± 0.29
d in [5, 7)3.86 ± 0.302.24 ± 0.190.65 ± 0.096.36 ± 0.43
d > 7 (Far Void)5.48 ± 0.733.42 ± 0.191.25 ± 0.1710.03 ± 0.58 (unbounded O(

Near data (d<0.5d < 0.5), all methods contract. Far in the void (d>7d > 7), DP-TS sustains 1.5--2× the spread of the bootstrap and 4--5× that of MC-dropout without quadratic ray explosion (Figure 4, Appendix B).

AuthorsOpenReview ↗

Response to Reviewer iHZ8 (Part 2/3)

Continuing our response:

Why single draws do not collapse into linear extrapolation (Q3): Standard piecewise-linear networks (ReLU, SiLU) extrapolate linearly only outside the convex hull of their training points. In standard BNNs or pure bootstrap, the training points are confined to the observed sample D_t, so unobserved space lies outside that convex hull. In DP-TS, every single draw trains on the observed sample D_t together with T pseudo-atoms {x̃_j, ỹ_j}{j=1}^T drawn from P₀. Because P₀ has full spatial support supp(P₀) = X, the pseudo-atoms x̃_j are distributed across the entire domain, so unobserved regions become interpolation regions within the convex hull of {x_i}{i=1}^t ∪ {x̃_j}_{j=1}^T rather than extrapolation regions. Each draw anchors the network to randomized targets ỹ_j ~ N(μ₀, σ₀²), preventing linear extrapolation into the void. Across draws, varying pseudo-atom targets generate the non-vanishing spread quantified in Table 3. This pathology of naive bootstrapping in unvisited space directly parallels the fundamental challenge of deep exploration in sequential decision-making: as established by Osband et al. (2016, 2018) and Vashishtha (TMLR 2026), pure bootstrap models merely reshuffle observed data and suffer epistemic collapse in unvisited regions without an explicit prior mechanism. While Osband et al. (2018) inject prior variance via multi-network ensembles with additive randomized prior functions (f_θ(x) + p(x) at O(K) compute), DP-TS provides this prior anchor directly in data space via Dirichlet pseudo-atoms P₀, sustaining epistemic exploration with a single network (Vashishtha, TMLR 2026). In higher dimensions, fixed T covers space more sparsely; scaling T with dimensionality or adapting pseudo-atom locations is a natural next step, which we now discuss in Section 6.

Theoretical scope and regret guarantees (W3): You rightly noted that our original submission asserted a regret guarantee without formalizing the candidate-indexed information ratio or addressing the non-convex optimization gap highlighted by Phan, Abbasi-Yadkori & Domke (2019). We have reorganized the theoretical results into three precise steps:

  1. Candidate-indexed information ratio (Proposition 2, Appendix C): For any finite pool X_M (|X_M| = M), the posterior minimizer probabilities p_t(a) = P(x* = a | F_t) form a closed Doob martingale (E[p_{t+1}(a) | F_t] = p_t(a)). The conditional Shannon entropy H(x* | F_t) is a supermartingale whose telescoping drop across T rounds equals cumulative mutual information: ∑_{t=0}^{T-1} I_t ≤ log M. For exact posterior sampling with σ-sub-Gaussian predictive noise, the information ratio is bounded horizon-independently by Γ_t ≤ 2σ² M, yielding Bayesian simple regret E[S_T] ≤ √(2σ² M log M / T).

  2. Exact realization for k-dimensional random features (Appendix C.5): To eliminate the non-convex SGD gap (ε_SGD = 0), we analyze linear models on k-dimensional features f_θ(x) = θᵀ φ(x) (φ: X → ℝ^k). Here, weighted empirical risk minimization has an exact closed-form solution θ_t = V_t⁻¹ b_t. When data span a strict subspace V ⊊ ℝ^k, pure bootstrap suffers a rank-deficient Gram matrix and collapses along V^⟂. In contrast, P₀ pseudo-atoms act as a strictly positive-definite stochastic regularizer (E[Σ̃_T] ≥ λ₀ W_prior I_k > 0), sustaining void variance and bounding the information ratio by Γ_t ≤ 2σ² k. This guarantees sublinear O(√(k T log M)) cumulative regret.

  3. Deep MLPs and open questions: For multi-layer neural networks trained with finite-step SGD or AdamW, non-convex empirical risk minimization does not admit closed-form solutions. We do not claim exact regret bounds for deep MLPs; we state the finite-epoch SGD approximation gap as an open problem and instead verify the necessary void spread empirically in Table 3 and Figure 4.

AuthorsOpenReview ↗

Response to Reviewer iHZ8 (Part 3/3)

Concluding our response:

Experimental rigor, baselines, and negative results (W4, W5): All reported benchmarks now report mean ± s.e. across 10 held-out evaluation seeds (42–51). To prevent data snooping, all hyperparameters were fixed beforehand on disjoint tuning seeds (100–104). We benchmarked against a tuned ARD-GP (marginal likelihood optimization; both LCB and TS) and TuRBO-1 (our own disclosed implementation). Under the protocol where initial evaluations are confined to a far corner (f = 1.0) with 35 sequential steps over a shared 3,000-candidate pool, we report results in Table 4 (Appendix D.1):

MethodSurrogateAckley 2DLevy 2DRosenbrock 4DLatency
Random Search (uniform)None3.73 ± 0.290.32 ± 0.08102.6 ± 20.00 ms
Standard BNN1 Net (MC-Dropout)2.11 ± 1.090.20 ± 0.0973.2 ± 16.483 ms
Tuned ARD GP, LCBMarginal-Likelihood GP1.27 ± 0.250.056 ± 0.01751.2 ± 13.613 ms
Tuned ARD GP, TSMarginal-Likelihood GP2.13 ± 0.450.025 ± 0.01069.7 ± 7.3433 ms
TuRBO-1 (our impl.)Trust-Region ARD GP0.58 ± 0.230.48 ± 0.308.1 ± 2.821 ms
SAASBO (Sparse HMC)ARD Matérn-5/20.72 ± 0.100.004 ± 0.00115.6 ± 4.0117 ms
Pure Bayesian Bootstrap1 Net (alpha -> 0)1.49 ± 0.950.40 ± 0.1345.6 ± 23.235 ms
DP-TS (Ours)1 Net (Thompson, re-init)0.46 ± 0.110.014 ± 0.00431.8 ± 9.633 ms

We do not claim uniform superiority across all settings: TuRBO-1 naturally excels on smooth unimodal valleys like Rosenbrock 4D, and SAASBO is strong on low-dimensional isotropic functions. DP-TS is specifically designed for multimodal and deceptive landscapes where the optimum lies in the data void.

Controlled distance sweep: To test this directly, we varied the initial 5-point box across f ∈ {0, 0.25, 0.5, 0.75, 1} from the true optimum x* to the far corner. At f = 0 (data placed near the optimum), local trust-region and BNN baselines lead (0.04–0.14 on Ackley). But as data shift into the void (f ≥ 0.75), local baselines degrade by 10–900× while DP-TS remains stable (0.24 → 0.46 on Ackley; 0.006 → 0.014 on Levy), demonstrating resilience to initial placement (Figure 1, Section 5.1; Table 5 in Appendix D.2).

Long horizons, deceptive landscapes, and high-dimensional robotics:

  • 4D Lunacek: Over 200 iterations, DP-TS successfully escapes the deceptive local funnel (1.06 ± 0.11), whereas TuRBO-1 (4.14 ± 0.07), BNN (4.38 ± 0.04), and pure bootstrap (4.65 ± 0.06) stay trapped in the local well (Figure 2).
  • 10D Lunacek with deceptive domain knowledge: When supplied with an adversarial prior centered on the deceptive funnel, DP-TS recovers to 3.88 ± 0.34, outperforming BOHAMIANN (6.84 ± 0.42), bootstrap (8.11), and standard BNN (8.54), confirming that empirical data properly override incorrect priors as sample size grows (Table 7, Appendix D.3).
  • 60D Rover & 32D Pendulum (Section 5.2): Under exact stick-breaking weights and beta prior scaling (W_prior ~ Beta(α, t)), DP-TS achieves the lowest final cost on 60D Rover trajectory planning: 2.16 ± 0.17 (@35) → 1.91 ± 0.12 (@150) → 1.71 ± 0.06 (@500), reaching 10/10 (100%) collision-free trajectories (Figure 3, Table 1). Tuned ARD-GP reaches 1.91 and TuRBO-1 reaches 2.58. Per-step latency explodes cubically by 39× for Tuned GP (10.2 → 395.7 ms) and TuRBO-1 (12.4 → 477.8 ms) vs linear O(t) for DP-TS (29.8 → 64.8 ms). At a 5.0s budget, DP-TS achieves 1.93 ± 0.13, beating Tuned GP (1.97) and TuRBO-1 (3.39). On 32D Pendulum, DP-TS progresses from 926.4 ± 68.7 (@35) to 620.8 ± 79.0 (@500) and 762.0 ± 84.3 at 5.0s, beating Tuned GP (859.4; 508 ms) and TuRBO-1 (907.8; 634 ms).
  • Runtime scaling (Table 2): At t ∈ {50, 200, 500, 1000, 2000}, DP-TS scales as 22, 35, 65, 86, 130 ms/step vs Tuned GP's 9, 45, 387, 2765, 20,998 ms/step (160× slower).
  • Hyperparameter stability (Q2): Sweeps over α ∈ {0.1, 0.5, 1, 2, 5, 10}, T ∈ {5, 10, 30, 60}, and σ₀ ∈ {0.5, 1, 2, 3} show robust convergence across α ∈ [0.5, 2] (Appendix D.4).
  • Tone, hyperbole, title, and abstract (Minor Weakness 1): We systematically removed superlatives and unhedged assertions, replaced sensationalist headings (e.g., "Shattering the GP Cubic Bottleneck" → "Overcoming the O(t³) GP Bottleneck"), and retitled the paper ("Data-Space Dirichlet Process Bayesian Neural Networks for Epistemic Exploration in the Data Void"). We overhauled the Abstract and Introduction to explicitly scope the method to sequential exploration, state approximations upfront

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.