Bidirectional Information Flow (BIF) - A Sample Efficient Hierarchical Gaussian Process for Bayesian Optimization

Conference: NeurIPS2026
arXiv: 2505.11294
Code: https://anonymous.4open.science/r/Bidirectional_Information_Flow
Area: Optimization & Theory
Keywords: Bayesian optimization, hierarchical Gaussian processes, credit assignment, structural priors, subtask transfer

TL;DR

BIF constructs a soft parent prior from child Gaussian processes' acquisition maps and allocates real parent responses to children using uncertainty-aware weights for continual learning, improving reconstruction and learning trajectories in low-budget composite tasks; this feedback comprises biased pseudo-responses, not true subtask decomposition.

Background & Motivation

Expensive black-box optimization often returns only an aggregate outcome while retaining useful subtask structure. For example, muscle activation under multichannel neurostimulation depends jointly on electrode channels, and GAN performance depends on both generator and discriminator parameters. Standard Gaussian process Bayesian optimization (GPBO) learns the aggregate response over the full parameter space and estimates uncertainty, but does not explicitly preserve individual subtask knowledge. Hierarchical Gaussian processes instead assign subspaces to child models and let a parent handle the aggregate objective. If children receive only a few real initialization samples and subsequently send information upward without receiving feedback, new expensive parent observations cannot improve them.

This gap cannot be closed simply by copying the parent observation to every child. The aggregate response mixes subtasks and interactions: it is not the true value of any individual subtask, and its composition need not be strictly additive. Copying it contaminates children with the same global signal; requiring the parent to equal a fixed combination of child outputs would also restrict its ability to correct erroneous priors. BIF therefore pursues two goals together: early search guidance from existing child knowledge and the use of every real parent query to train the entire hierarchy rather than the parent alone.

The approach treats upward information as a structural prior that data can correct, and downward information as approximate training feedback requiring independent calibration. Core Idea: child acquisition maps form a soft parent mean prior, aggregate parent responses yield conservative credit assignments, and real and inferred child responses are normalized separately to update the hierarchy in a closed loop.

Method

Overall Architecture

Inputs comprise specified subtasks and parameter subspaces, a few real initialization samples per subtask, and an environment that can evaluate the aggregate objective; outputs include optimization recommendations, a parent surrogate, and reusable child models. BIF does not discover a task decomposition from an unstructured black box. It starts with a specified hierarchy and repeatedly performs upward acquisition-map transfer, parent selection and real querying, downward response credit assignment, and dual-stream calibration. The parent continues to train on real aggregate responses, whereas children receive their own real data and pseudo-responses allocated from parent feedback as separate sources.

At the start of each round, children generate acquisition maps over their subspaces. The parent lifts these maps to the full composite space and averages them to guide its mean prior; its own posterior and acquisition function still determine the actual query. That query returns one real aggregate response, not true labels for every subtask. The downward pathway uses the children's pre-query predictions and uncertainties to allocate this response, producing an approximate label at each child's projected coordinates. Updated children generate the next round's acquisition maps.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    I["Specified subtasks and<br/>real initialization data"] --> U["Acquisition-map<br/>soft prior"]
    U --> P["Parent posterior selection<br/>and real environment query"]
    P -->|Real aggregate response| D["Conservative<br/>response credit"]
    U -->|Child acquisition utilities| D
    D -->|Inferred child responses| R["Independent<br/>dual-stream calibration"]
    I -->|Real child responses| R
    R -->|Update children| U
    P -->|Update parent with real aggregate response| U

Real environment querying and model training must be distinguished: producing multiple child training points does not imply multiple physical experiments or observation of multiple true child responses. The mean prior changes with the children, while the parent corrects it using real historical data. The loop consequently changes both child models and subsequent parent search preferences.

Key Designs

1. Acquisition-map soft prior: transmit subtask search preferences without fixing the parent composition

Each child GP uses a Matérn kernel to predict a mean and standard deviation in its parameter subspace and then computes acquisition utility. An acquisition map contains exploration/exploitation scores at candidate locations, not true subfunction values or merely posterior means. A parent candidate specifies coordinates in each child subspace, allowing corresponding map entries to be read and averaged. An uncertain but promising child region can increase the prior attraction of a composite candidate, explaining why this pathway helps early search rather than deterministically predicting aggregate reward.

The parent mean uses an intercept and the averaged acquisition map as explicit basis functions, with Gaussian priors retained over their two coefficients. Separately, the parent kernel sums child kernels to encode structural covariance. Mean guidance and kernel structure are distinct mechanisms and both are necessary to explain the model.

\[ A(x)=\frac{1}{|S|}\sum_{s\in S}\alpha_s(x_s),\qquad m_p(x)=\theta_{\mathrm{bias}}+\theta_{\mathrm{scale}}A(x),\qquad \boldsymbol{\theta}_p\sim\mathcal{N}([0,1]^\top,I),\qquad k_p(x,x')=\sum_{s\in S}k_s(x_s,x_s'). \]

Here, \(S\) is the child set and \(x_s\) projects a parent candidate into a child subspace. The coefficient prior centers the initial parent mean on the average child map, while its nonzero covariance preserves flexibility: contradictory parent observations need not leave the model bound to the map's original scale and offset. This is soft structural guidance, not a hard constraint that the parent objective obey a known analytic combination of subtasks.

Averaging rather than summing also introduces a scale distinction. Appendix L.1 establishes that, when the true objective is exactly the sum of child functions and child prediction errors and exploration residuals vanish, the average map converges to the objective divided by the number of children. It does not establish unconditional convergence of the average map to the objective itself. The parent mean's scale coefficient is therefore not redundant, and the main text's abbreviated convergence claim should be interpreted with this proportionality in mind.

2. Conservative response credit: exploit each parent query through controlled approximate labels

How aggregate feedback trains children is the central mechanism and the main source of possible misunderstanding. At the composite point just queried by the parent, BIF reads each child's utility at its projected coordinates and divides it by that child's maximum acquisition-map utility to reduce direct scale differences. A softmax then produces positive weights that multiply the real parent response. With UCB, utility combines current predictions with uncertainty discounted by location query counts, so a data-starved child may receive more credit rather than allocation depending only on predicted means.

\[ \alpha_s(x_s)=\mu_s(x_s)+\gamma\frac{\sigma_s(x_s)}{\sqrt{n_s(x_s)}},\qquad c_s(x_s)=\frac{\alpha_s(x_s)}{\max_{z_s}\alpha_s(z_s)},\qquad w_s(x)=\frac{\exp(c_s(x_s))}{\sum_{j\in S}\exp(c_j(x_j))},\qquad \tilde y_s(x_s)=w_s(x)y_p(x). \]

These relations express the source mechanism; \(\tilde y_s\) explicitly marks inferred labels to distinguish them from real child-environment responses. The weights sum to one, so inferred responses before normalization sum to the parent response, including a proportional allocation of its noise. However, they originate from the same noisy observation and are not independent new measurements. Conservation means that the full aggregate signal was not copied multiple times; it establishes neither unbiased labels nor identification of a true, unique latent decomposition.

The paper claims that division by the maximum utility bounds credit scores within \([0,1]\). This requires nonnegative utilities and a strictly positive, finite maximum. Maximum division alone handles neither negative utilities nor a zero denominator. The main text and Appendix L.3 do not specify general shifting, clipping, or zero-denominator protection, so these conditions are retained rather than replacing the mechanism with an unreported min-max credit score. The main-text UCB denominator also lacks an explicit count floor; Appendix L.1 introduces \(\sqrt{n_s(x_s)\vee1}\). This boundary convention requires implementation verification.

Even perfectly learned child functions generally do not make maximum-normalized softmax weights equal the true shares of aggregate response. The zero-bias result in Appendix L.2 requires positive additive subtasks, equal true contributions at the examined point, and equal maxima of the two child functions. General unbalanced cases receive only a bias characterization. This is a useful online credit-assignment heuristic, not an identification theorem recovering every true child label from a single aggregate response.

3. Independent dual-stream calibration: avoid mixing responses with incompatible scales

Each child initially receives real child-environment data and subsequently receives pseudo-responses allocated from the parent. The parent's composition, number of children, and credit weights alter pseudo-response scales. Direct concatenation can therefore assign incompatible values to the same child-space location. BIF applies min-max normalization separately to the two sources before training the same child GP, making within-source relative response levels the primary learning signal.

\[ T(y)=\frac{y-\min(Y)}{\max(Y)-\min(Y)}. \]

Real and inferred streams use their own \(Y\) rather than a range computed after merging. This reconciles numerical scales but cannot remove structural credit-assignment bias or guarantee that normalized pseudo-labels and real labels represent exactly the same function. In particular, a single initialization sample or a constant-response stream gives a zero denominator. The paper does not specify a constant-stream rule, so this transformation should not be presented as automatically valid for all initialization states.

This differs from experimental online z-score normalization or division by the maximum environment response observed so far, which are dataset-level response preprocessing choices. Dual-stream calibration specifically reconciles the two sources arriving at children. After independent normalization, child labels generally no longer sum to the original parent response. Signal conservation thus belongs to the credit-assignment step, not a claim that the entire training pipeline preserves physical quantities.

A Worked Example

Suppose a parent candidate contains coordinates in two child subspaces, its real query returns 10, and the normalized child utilities are 0.8 and 0.2. This illustrates the mechanism and is not an experimental observation from the paper. Softmax gives weights of approximately 0.646 and 0.354, so children receive pseudo-responses of approximately 6.46 and 3.54, while the parent receives the real aggregate response of 10.

The pseudo-responses sum to 10, but true child responses need not equal either value; even an additive objective can admit other contribution combinations with the same aggregate value. Each child next normalizes using its separate real-stream and inferred-stream ranges, adds the projected coordinates to its training data, and updates its posterior. Changed means and uncertainties alter the next child acquisition maps and the parent soft prior, potentially leading to a different composite query.

This loop uses one expensive aggregate query to update multiple surrogates. It neither reduces querying to zero nor provides two additional independent real child samples. Under a strongly nonlinear composition, the same credit rule still applies: the parent may correct misleading guidance, but child labels are more likely to retain composition-induced bias.

Loss & Training

Algorithm 1 first gathers \(r\) real samples for each child and then runs a parent-query loop with budget \(Q\). Each round rebuilds child acquisition maps and the parent prior, maximizes the parent acquisition function, obtains one real parent response, allocates credit, appends data, increments the relevant query counts, and performs \(t\) training steps for all models. Experiments use Adam with learning rate 0.01 and child Matérn kernels with smoothness 0.5. GP updates learn kernel and related parameters; the paper does not introduce a separate neural-network-style supervised loss.

The cached algorithm queries the environment through the parent every round. It does not specify an additional round-robin schedule alternating real queries among the parent and children. Continual child updates primarily use pseudo-responses allocated from parent queries, and incrementing training counts should not be interpreted as equally many child-environment samples. Real child initialization is explicitly required, so applications still need those initial responses or the modular transfer scheme. Reliance on aggregate feedback describes the main loop after initialization.

The main text gives neurostimulation parameters in the order \(\{\kappa,\gamma,r,t\}=\{8,4,1,15\}\). Appendix E Table 3 lists different settings for different normalization choices, while its closing paragraph reports another overall selection of 4.0, 3, 6, and 10. These cannot be merged into one universal configuration. In Appendix G, modular transfer clears previous query data and retains pretrained child hyperparameters before insertion into a new parent task, rather than carrying all historical observations. This improves early parent performance, but children trained from scratch can eventually reconstruct better, revealing the cost of transferred task bias.

The theory analyzes learned-kernel GP-UCB. Its central conditions include a compact action space, a bounded RKHS norm under the true kernel, bounded diagonal variance for true and learned kernels, independent Gaussian noise, and suitable confidence parameters. Kernel mismatch is converted into posterior mean and variance perturbations before entering the cumulative regret bound. Mean error contributes linearly, whereas variance error also enters through square roots. With additive structure, total kernel mismatch is bounded by the sum of child mismatches; this does not mean regret numerically equals that sum.

These bounds explain why reducing child kernel mismatch is desirable, but do not prove that pseudo-label credit assignment necessarily drives mismatch to zero. The appendix's standard GP-UCB selection rule also differs from the location-count-adjusted UCB described for implementation, and prior-map convergence requires strict additivity rather than arbitrary complex composition. The theory is conditional error-propagation analysis, not a general end-to-end convergence or safety guarantee derived from nonlinear robustness experiments or empirical child reconstruction gains.

Key Experimental Results

Main Results

Main-text Table 1 uses 100 training queries and reports means and standard errors over 10 random initializations. The table below retains the source's percentage-scale RO and \(R^2\) values. AUC cumulatively combines RO, parent \(R^2\), and average child \(R^2\) over training; it is not merely a final optimization metric. Across tasks and models with different available child metrics, AUC should not be interpreted as cumulative regret on a common scale.

Task	Model	RO	Parent \(R^2\)	Child \(R^2\)	AUC
Michalewicz	Vanilla GPBO	99.30 ± 20.25	55.29 ± 1.67	—	10,138.74
Michalewicz	ADD GP	97.27 ± 1.43	75.17 ± 1.80	75.74 ± 2.34	18,716.24
Michalewicz	BIF UCB	98.87 ± 0.60	86.10 ± 0.88	61.88 ± 3.25	18,587.90
Synthetic 3D	Vanilla GPBO	80.79 ± 0.28	13.39 ± 0.13	—	7,677.74
Synthetic 3D	BIF UCB	88.24 ± 3.23	61.39 ± 2.55	42.86 ± 5.87	14,986.14
Neurostimulation	Vanilla GPBO	98.46 ± 0.07	70.13 ± 0.11	—	14,119.99
Neurostimulation	BIF UCB	81.83 ± 0.25	79.70 ± 0.12	63.98 ± 0.15	18,229.36
HPO-GAN	Vanilla GPBO	96.04 ± 0.45	9.67 ± 1.00	—	2,035.21
HPO-GAN	BIF UCB	96.50 ± 0.45	12.19 ± 0.86	26.72 ± 1.74	2,596.53
HPO-Lasso	Deep GP	99.86 ± 0.04	89.27 ± 5.89	—	14,655.16
HPO-Lasso	BIF UCB	99.98 ± 0.02	75.35 ± 2.84	50.96 ± 5.79	17,906.14
HPO-Lasso	BIF PI	99.83 ± 0.07	79.96 ± 0.94	71.63 ± 3.40	20,249.31

The source defines RO below, where \(\hat x\) is the location recommended by the acquisition function and \(x^*\) and \(\tilde x\) are the true maximum and minimum locations.

\[ RO=\frac{y(\hat x)+y(\tilde x)}{y(x^*)+y(\tilde x)}. \]

The original expression adds the minimum value rather than using conventional minimum subtraction. With a negative minimum, the denominator can vanish and the metric is not generally guaranteed to lie in \([0,1]\). The definition is retained and flagged rather than silently replaced. The neurostimulation results prose attributes parent \(R^2\) of 81.43 to EI, whereas Table 1 attributes it to PI. This note follows the table configuration and preserves the discrepancy.

Ablation Study

Appendix I Table 5 separates removal of the acquisition-map soft prior (No Up) from removal of conservative response credit (No Down). The selection below emphasizes child reconstruction and learning trajectories. Full BIF does not achieve the best RO on every task.

Task	Metric	No Up	No Down	Full BIF
Sub 2D	Child \(R^2\)	74.76 ± 2.55	10.99 ± 1.67	81.65 ± 1.94
Sub 2D	AUC	20,251.86	13,813.10	22,650.28
Michalewicz	Child \(R^2\)	47.43 ± 4.38	32.08 ± 4.74	61.88 ± 3.25
Michalewicz	AUC	18,079.62	16,782.96	18,587.90
HPO-Lasso	Child \(R^2\)	37.68 ± 4.44	21.57 ± 3.94	50.96 ± 5.79
HPO-Lasso	AUC	16,549.95	16,708.63	17,906.14

For HPO-Lasso, No Down reduces child \(R^2\) from 50.96 to 21.57, approximately 57.7% relative to the full model; for Sub 2D, the reduction from 81.65 to 10.99 is approximately 86.5%. The main text claims that removing upward flow reduces AUC by 14% on HPO-Lasso and 6% on Michalewicz, whereas Table 5 implies approximately 7.6% and 2.7% relative to the full model. The prose percentages do not replace the table values.

Key Findings

• In 3D, parent \(R^2\) increases from 13.39 to 61.39 and AUC from 7,677.74 to 14,986.14, supporting low-budget reconstruction and trajectory improvements but not arbitrary high-dimensional scalability.
• Neurostimulation BIF UCB improves reconstruction and AUC but has lower RO than Vanilla GPBO. Tuning favors aggregate AUC, so these results do not establish universal superiority in final-optimum search.
• Exact-GP computation remains costly: Appendix D reports 100-query Michalewicz runtimes of BIF 1:14, Vanilla 0:07, and Deep GP 5:11. The method targets environment sample efficiency, not minimum surrogate computation time.

Highlights & Insights

• Acquisition utility serves as cross-level search knowledge rather than only choosing the next query. Placing it in a soft mean prior instead of a hard constraint lets parent data correct unsuitable composition assumptions.
• One real aggregate observation updates the parent and children together. Credit assignment supplies a useful learning direction, not new independent evidence; this distinction determines how pseudo-label quality should be evaluated.

Limitations & Future Work

• Specified subtasks and real child initialization are prerequisites. Without accessible child environments, initialization still needs a source; the method cannot be claimed to rely exclusively on aggregate feedback throughout.
• Nonnegative utilities, positive maxima, nonzero normalization ranges, and count floors require boundary handling. Verifiable implementation rules are needed rather than assuming the formulas work for arbitrary rewards.
• Strong nonlinear or negative interactions undermine credit interpretability. Appendix H's exponential-nonlinearity test reports a 37% reduction in parent \(R^2\) and 24% in child \(R^2\), showing that good RO need not imply successful latent-subtask recovery.
• Learned-kernel regret bounds require additional assumptions. Connecting the theory to count-adjusted UCB, soft nonzero means, and dual-stream normalization needs care, while exact GPs retain cubic computational growth.

Related Work & Insights

• vs Laferrière et al.'s hierarchical GPBO: both use children to guide a parent. BIF adds downward parent-observation feedback to address post-initialization child stagnation, at the cost of credit bias and calibration requirements.
• vs ADD GP / Deep GP: ADD GP supplies additive structure and Deep GP hierarchical nonlinear expressiveness. BIF emphasizes explicit reusable subtasks and a feedback loop; it neither dominates additive models on every metric nor identifies arbitrary nonlinear compositions.
• vs BILBO: BILBO constrains upper-level feasibility using lower-level solutions. BIF retains parent correction through soft priors, fitting settings where structure is useful but imperfect.

Rating

• Novelty: 4/5 — Connects acquisition-map soft priors, response credit, and dual-stream calibration in an explicit online hierarchical loop.
• Experimental Thoroughness: 4/5 — Five benchmarks, module ablations, and robustness tests offer broad coverage, but initialization costs and metric choices affect comparisons.
• Writing Quality: 3/5 — The mechanism is understandable, but normalization boundaries, configurations, metric attribution, and some percentages are inconsistent.
• Value: 4/5 — Useful for expensive black-box optimization with available subtasks, not a substitute for child-label identification or safety validation.