Two-Phase Rapid Simulation-Based Inference with Differentiable Simulators
Authors: Uros Zivanovic, Rahul Srinivasan, Roberto Trotta, Andre Scaffidi
Organizations: Max Planck Institute for Intelligent Systems, Tübingen, Germany · Max Planck Institute for Gravitational Physics (Albert Einstein Institute), Potsdam, Germany · Scuola Internazionale Superiore di Studi Avanzati (SISSA), Trieste, Italy · IFPU - Institute for Fundamental Physics of the Universe, Trieste, Italy · INFN – National Institute for Nuclear Physics, Italy · Italian Research Center on High-Performance Computing, Big Data and Quantum Computing, Casalecchio di Reno, Italy · Imperial College London, United Kingdom
We introduce Rapid Simulation Based Inference (RSBI), a diffusion-based variational approach to likelihood-free Bayesian inference that achieves high sampling efficiency under large prior-to-posterior volumes. RSBI builds on advances in Schrödinger Bridge (SB) diffusion sampling to handle multi-modal posteriors, with an initial likelihood supplied by a distance measure via a differentiable simulator. The key innovation of this initial surrogate likelihood is that it gives a highly efficient proposal distribution, replacing sequential rounds of other methods with an efficient one-shot proposal. Subsequent refinement then transports the proposal to the true posterior as specified by the simulator and the original prior. The SB sampler is warm-started from the proposal, and its energy is corrected by a neural likelihood ratio that is learned only where the proposal places mass. As a variational method, RSBI does not suffer from prior leakage commonly observed in standard sequential posterior estimation methods. To improve mode coverage we optionally utilize the well-tempered meta-dynamics framework, which also encourages exploration of the prior volume. We observe strong performance on standard SBI benchmarks, particularly when the prior volume is increased relative to the ground truth posterior, achieving a performance degradation of only ∼5% on Two Moons under a highly constrained simulation budget. Additionally, we evaluate RSBI's performance for gravitational wave ring-down posterior estimation, in a real-world physics benchmark inspired by black hole spectroscopy.
Figures & tables
Figure 1: Actual diffusion trajectories from the RSBI pipeline on the Two Moons task detailed in Sec. 5 . RSBI is able to produce a high quality posterior despite the large prior size.
Figure 2: Left: Schematic representation of the two-phase procedure of RSBI. Right: Comparison of posteriors from RSBI (our method) and SNPE on a Two Moons task with 104 simulations and Rp∼104.3 , with two sequential steps for both SNPE and Phase 2 of RSBI.
Prior (see Table 21 )
Method
default
wide
log10Rp
≈2.6
≈2.8
SNVI+fKL
0.958(0.032)
0.986(0.022)
SNPE
0.935(0.071)
0.985(0.015)
RSBI (ours)
0.891(0.070)
0.903(0.069)
Table 1: Mean C2ST on gw_ringdown against the emcee reference at Nsim=105 , averaged over 10 simulated observations. For every method the phases φ220 , φ221 are wrapped to [0,2π) before scoring.
Figure 3: 6D gw_ringdown posteriors of RSBI, SNPE and the MCMC reference for observation 10 at Nsim=105 , under the default prior (left, log10Rp≈2.62 ) and the wide prior (right, ≈2.84 ; Table 21 ). Insets: whitened observation (grey), noise-free injection (red) and the 1σ posterior-predictive bands of SNPE and RSBI from 200 re-simulated posterior samples each.
Figure 4: Robustness to prior-to-posterior volume ratio on Two Moons, Lotka–Volterra, Bernoulli GLM and SLCP benchmarks. Mean C2ST ( ±1σ ) calculated against the reference posterior vs. simulation budget Nsim over 10 observations. Legends report log10Rp , full line is the sbibm default. C2ST =0.5 matches the reference, 1.0 is fully distinguishable.
Appendix figures & tables34 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Left: C2ST of RSBI against number of simulations. Right: C2ST of RSBI against the equivalent compute-matched simulation budget (reported in Table 3 ). Calculated for Two Moons, Bernoulli GLM and SLCP benchmarks. Y-axis labels include the compute ratio CR charged per Phase-1 simulation (Table 2 ). Lines are the prior scenarios of Figure 4 with error bars being the standard deviation over 10 observations.
Simulator
Forward (ms)
Forward+Backward (ms)
Ratio
Two Moons
0.1401
0.3528
2.52 ×
SLCP
0.2113
0.8524
4.03 ×
Bernoulli GLM
0.8057
1.3854
1.72 ×
Appendix
Table 2: Effect of backwards pass on per-batch mean compute time across 3 sbibm benchmarks.
Task
Prior
Nsim=103
Nsim=104
Nsim=105
Two Moons
default
724
5 064
70 037
wide10x
800
9 203
99 203
wide50x
800
9 203
99 203
Bernoulli GLM
default
820
8 200
83 455
medium
820
8 200
82 015
large
820
8 200
82 015
Appendix
Table 3: Total simulator calls made by the compute-matched runs of Figure 5 .
Figure 6: Demonstration of posterior leakage and it’s mitigation with RSBI: Posterior samples on the tractable squared-location task ( d=2 , Nsim=104 , one observation). Dashed square: the support of the uniform prior, [−b,b]2 ; crosses: the four true modes at (±c,±c) , which lie close to the corners of the support. Each panel shows the raw sampler output of one method (before rejection for the flows, before removing the metadynamics bias for RSBI), with the fraction of its mass inside the support. The C2ST against the analytic reference ( 0.5 = indistinguishable) is computed on each method’s final posterior restricted to the support: the in-support samples of the flows and NPE, and the in-support reweighted samples of RSBI. The atomic SNPE-C flows and amortised NPE place much of their mass outside the support; RSBI keeps it inside while covering all four modes, like the analytic reference (top left).
Method
Nsim=103
104
105
SNPE (prior, R=2 )
0.997(0.002)
0.991(0.015)
0.985(0.009)
SNVI+fKL (prior, R=5 )
0.987(0.004)
0.943(0.023)
0.993(0.020)
Phase 1 + TSNPE ( R=2 )
0.925(0.032)
0.873(0.083)
0.738(0.107)
Phase 1 + SNVI ( R=2 )
0.938(0.061)
0.770(0.101)
0.848(0.023)
RSBI (Phase 1 + NRE/ASBS)
0.693(0.033)
0.617(0.034)
0.629(0.026)
Appendix
Table 4: Phase-2 ablation on Two Moons with the widest prior ( [−50,50]2 ): mean (sd) C2ST over ten observations at matched simulation budget. Phase 1 + TSNPE uses the trained Phase 1 sampler of each RSBI run as the first truncation of TSNPE [ 9 ] , with the same per-round simulation allocation as RSBI’s ratio estimator ( n1 more simulations, since the Phase 1 pairs are not truncated-prior draws). Phase 1 + SNVI replaces RSBI’s NRE and ASBS sampler by SNVI [ 7 ] (SNRE-B and a normalising flow trained by forward KL), trained on exactly the same simulations in round 1 and the same number in round 2. SNPE and SNVI+fKL from the prior are the main-benchmark configurations. Phase 1 alone (524 simulations) reaches 0.995(0.001) .
Figure 7: C2ST against the reference posterior on Two Moons with the widest prior (mean ± sd over ten observations, lower is better). Values as in Table 4 .
Figure 8: Phase-1 posteriors used for the main results in Fig. 4 for Two Moons, SLCP, gw_ringdown and the Bernoulli GLM (top left to bottom right), each at its default prior with Nsim=105 at observation 1 (see Sec. G for the Phase-1 configuration). We show the 1- and 2- σ contours of the Phase-1 (blue) and reference (grey) posteriors over the prior (dashed red: the prior box for the uniform priors, and the 1- and 2- σ contours of the Gaussian prior for the Bernoulli GLM). For SLCP and gw_ringdown we show the 2-D marginals of the full 5-D and 6-D posteriors for the two parameters with the most pronounced structure. These are (θ3,θ4) , which separate the four SLCP modes, and the mass and spin pair (M,χ) , whose posterior is a single curved ridge (Phase 1 also places small islands of mass off the ridge). For the unimodal Bernoulli GLM we show the most strongly correlated pair of its 10 parameters.
Figure 9: Pilot energy landscapes for Two Moons (left), SLCP (centre) and the Bernoulli GLM (right), computed with the Phase-1 surrogate, bandwidth and temperature used for the benchmark main results at the default prior on observation 1 (see Sec. G for Phase-1 hyperparamters and surrogates used). Colour is log(1+E/T−minE/T) ; white arrows show the direction of −∇θE ; stars are basin sinks (gold: significant); red edges are ridges. SLCP and the Bernoulli GLM are shown on 2-D slices through (θ3,θ4) and (θ1,θ2) . The number in each panel title is the score used to rank bandwidths for that posterior: separation for the multimodal Two Moons and SLCP, structure for the unimodal Bernoulli GLM. Scores for these settings: Two Moons: separation 0.79, structure 0.79; SLCP: separation 0.97, structure 0.21; Bernoulli GLM: separation 0.00, structure 0.74.
Task
wells
cohesion
ridge
depth
structure
separation
Two Moons
2
1.00
0.79
1.00
0.79
0.79
SLCP
4
0.22
0.97
1.00
0.21
0.97
Bernoulli GLM
1
0.74
–
1.00
0.74
0.00
Appendix
Table 5: Landscape scores of the pilot batches in Fig. 9 . Wells : significant basins. Cohesion : median cosine between a draw’s drift and the direction to its basin’s sink. Ridge : share of cross-basin neighbour pairs whose drifts point apart (– for a single well, where the ridge term is 1 ). Depth : mean over basins of min(1,ΔE/10σE) , or min(1,(E95−Emin)/10σE) for a single well. Structure = cohesion × ridge × depth. Separation = ridge × depth for two or more wells and 0 otherwise.
Surrogate
103
104
105
Gaussian-NLL
0.981±0.005
0.955±0.017
0.941±0.011
Energy distance
0.982±0.008
0.891±0.047
0.875±0.026
MMD
0.981±0.007
0.934±0.026
0.922±0.012
Appendix
Table 6: Phase-1 C2ST on Two Moons (baseline prior) vs. the simulation budget (capped), mean ± std over 10 observations. Lower is better ( 0.5 = indistinguishable from the reference).
Surrogate
103
104
105
Gaussian-NLL
1.560±1.092
10.560±6.981
19.030±11.230
Energy distance
1.270±1.328
57.630±24.147
101.430±37.410
MMD
1.480±1.282
21.500±10.500
38.990±12.119
Appendix
Table 7: Phase-1 mode fidelity on Two Moons (baseline prior) vs. the simulation budget (capped), mean ± std over 10 observations. Values ×1000 .
Surrogate
103
104
105
Gaussian-NLL
0.114±0.048
0.260±0.105
0.323±0.181
Energy distance
0.093±0.063
0.527±0.114
0.639±0.234
MMD
0.107±0.055
0.350±0.115
0.530±0.132
Appendix
Table 8: Phase-1 mode discovery on Two Moons (baseline prior) vs. the simulation budget (capped), mean ± std over 10 observations. Higher is better. Samples drawn from the prior score 0.04 .
Figure 10: Phase-1 samples on Two Moons for each surrogate’s profile (observation 1, 104 simulations; colour) over the reference posterior (grey). Dashed circles: the regions used for mode discovery ( 1.5× each moon’s 95th-percentile reference radius). All three surrogates place a cluster on each moon; discovery is limited by how unevenly the mass is split between the moons and by clusters wider than the moons themselves.
Surrogate
103
104
105
Gaussian-NLL
0.979±0.011
0.976±0.013
0.971±0.016
Energy distance
0.983±0.007
0.973±0.015
0.970±0.015
MMD
0.981±0.011
0.981±0.008
0.979±0.014
Appendix
Table 9: Phase-1 C2ST on SLCP (baseline prior) vs. the simulation budget (capped), mean ± std over 10 observations. Lower is better ( 0.5 = indistinguishable from the reference).
Surrogate
103
104
105
Gaussian-NLL
68.642±75.682
111.676±133.676
138.893±158.338
Energy distance
41.317±46.915
130.302±131.649
112.418±109.132
MMD
41.959±40.802
28.202±28.300
36.760±52.349
Appendix
Table 10: Phase-1 mode fidelity on SLCP (baseline prior) vs. the simulation budget (capped), mean ± std over 10 observations. Values ×1000 . Mode fidelity is computed on (θ3,θ4) only, the coordinates that separate the four SLCP modes (and carry the metadynamics bias).
Surrogate
103
104
105
Gaussian-NLL
0.287±0.241
0.338±0.298
0.377±0.332
Energy distance
0.173±0.172
0.358±0.266
0.389±0.273
MMD
0.164±0.163
0.224±0.184
0.241±0.183
Appendix
Table 11: Phase-1 mode discovery on SLCP (baseline prior) vs. the simulation budget (capped), mean ± std over 10 observations. Higher is better. Samples drawn from the prior score 0.23 .
Figure 11: Phase-1 samples on SLCP in the mode-separating coordinates (θ3,θ4) for each surrogate’s profile (observation 1, 104 simulations; colour) over the reference posterior (grey). Dashed circles: the regions used for mode discovery ( 1.5× each mode’s 95th-percentile reference radius). The Gaussian-NLL puts its mass beside all four modes ( ∣θ3∣≈2 ), spread along θ4 ; the energy distance spreads its mass in a ring around the prior box that passes through them; MMD concentrates on the box edges.
Surrogate
103
104
105
Gaussian-NLL
0.969±0.022
0.983±0.011
0.988±0.010
Energy distance
0.810±0.066
0.725±0.069
0.687±0.059
MMD
1.000±0.000
1.000±0.000
1.000±0.000
Appendix
Table 12: Phase-1 C2ST on Bernoulli GLM (baseline prior) vs. the simulation budget (capped), mean ± std over 10 observations. Lower is better ( 0.5 = indistinguishable from the reference).
Surrogate
103
104
105
Gaussian-NLL
86.070±84.689
231.940±129.145
184.920±175.774
Energy distance
430.270±200.013
737.610±71.645
767.610±54.824
MMD
0.000±0.000
0.000±0.000
0.000±0.000
Appendix
Table 13: Phase-1 mode fidelity on Bernoulli GLM (baseline prior) vs. the simulation budget (capped), mean ± std over 10 observations. Values ×1000 .
Figure 12: Phase-1 sensitivity to the surrogate bandwidth ϵ on Two Moons (4,512 simulations, one observation). Left: C2ST against the reference (lower is better; dotted: 0.5 ). Right: mode fidelity ( = precision × worst-mode recall, log scale; zero drawn at 10−5 ). Solid: each surrogate swept over its own bandwidth parameter (Gaussian-NLL variance, energy-distance length scale, MMD kernel bandwidth); dashed: the surrogate’s profile setting.
Surrogate
profile
ϵ⋆
C2ST (ϵ⋆)
C2ST (0.1ϵ⋆)
C2ST (10ϵ⋆)
Gaussian-NLL
0.894
0.0003
0.809
0.978 ( +0.170 )
0.967 ( +0.158 )
Energy distance
0.812
0.3
0.788
0.914 ( +0.127 )
0.951 ( +0.163 )
MMD
0.843
0.3
0.810
0.987 ( +0.176 )
0.987 ( +0.177 )
Appendix
Table 14: Phase-1 sensitivity to the surrogate bandwidth ϵ on Two Moons (4,512 simulations, one observation). profile : C2ST at each surrogate’s bandwidth and temperature as chosen, without the reference, from the energy-landscape diagnostics. ϵ⋆ : the best fixed bandwidth on the sweep (identified using the reference), with its C2ST, and the C2ST when ϵ⋆ is mis-set by one order of magnitude in either direction (change relative to ϵ⋆ in brackets). Lower is better ( 0.5 = indistinguishable from the reference).
Surrogate
non-trvial band
ϵpeak
fidelity
precision
min-recall
profile
Gaussian-NLL
[0.0003,0.00102]
0.0003
0.204
0.308
0.664
0.050
Energy distance
[0.3,3]
0.3
0.248
0.411
0.604
0.167
MMD
[0.1,0.783]
0.3
0.189
0.371
0.509
0.079
Appendix
Table 15: Phase-1 mode fidelity ( = precision × worst-mode recall; higher is better) versus the surrogate bandwidth ϵ on Two Moons (4,512 simulations, one observation). Non-trivial band: the range of ϵ with mode fidelity above 0.01 (– if none); ϵpeak : the bandwidth with the highest mode fidelity, with its precision and worst-mode recall; profile : mode fidelity at the surrogate’s profile setting.
Figure 13: Phase-1 sensitivity to the surrogate bandwidth ϵ on SLCP (10,000 simulations, one observation). Left: C2ST against the reference (lower is better; dotted: 0.5 ). Right: mode fidelity ( = precision × worst-mode recall, log scale; zero drawn at 10−5 ). Solid: each surrogate swept over its own bandwidth parameter (Gaussian-NLL variance, energy-distance length scale, MMD kernel bandwidth); dashed: the surrogate’s profile setting. Mode fidelity is computed on (θ3,θ4) only, the coordinates that separate the four SLCP modes (and carry the metadynamics bias).
Surrogate
profile
ϵ⋆
C2ST (ϵ⋆)
C2ST (0.1ϵ⋆)
C2ST (10ϵ⋆)
Gaussian-NLL
0.984
20
0.988
0.991 ( +0.003 )
0.988 ( +0.000 )
Energy distance
0.983
0.001
0.978
0.978 ( +0.000 )
0.984 ( +0.006 )
MMD
0.986
10
0.983
0.988 ( +0.005 )
0.986 ( +0.003 )
Appendix
Table 16: Phase-1 sensitivity to the surrogate bandwidth ϵ on SLCP (10,000 simulations, one observation). profile : C2ST at each surrogate’s bandwidth and temperature as chosen, without the reference, from the energy-landscape diagnostics. ϵ⋆ : the best fixed bandwidth on the sweep (identified using the reference), with its C2ST, and the C2ST when ϵ⋆ is mis-set by one order of magnitude in either direction (change relative to ϵ⋆ in brackets). Lower is better ( 0.5 = indistinguishable from the reference).
Surrogate
non-trivial
ϵpeak
fidelity
precision
min-recall
profile
Gaussian-NLL
–
3
0.003
0.027
0.124
0.007
Energy distance
[0.001,20]
20
0.038
0.132
0.288
0.043
MMD
[3,20]
10
0.032
0.083
0.378
0.005
Appendix
Table 17: Phase-1 mode fidelity ( = precision × worst-mode recall; higher is better) versus the surrogate bandwidth ϵ on SLCP (10,000 simulations, one observation). Non-trivial band: the range of ϵ with mode fidelity above 0.01 (– if none); ϵpeak : the bandwidth with the highest mode fidelity, with its precision and worst-mode recall; profile : mode fidelity at the surrogate’s profile setting. Mode fidelity is computed on (θ3,θ4) only, the coordinates that separate the four SLCP modes (and carry the metadynamics bias).
Surrogate
profile
best fixed ϵ
discovery (ϵ)
range over fixed ϵ
Gaussian-NLL
0.23 (adaptive)
10
0.14
[0.00,0.14]
Energy distance
0.38 (adaptive)
0.03
0.37
[0.11,0.37]
MMD
0.13 (adaptive)
10
0.26
[0.14,0.26]
Appendix
Table 18: Phase-1 mode discovery versus the surrogate bandwidth ϵ on SLCP (10,000 simulations, one observation). profile : discovery at the surrogate’s profile setting (adaptive or fixed bandwidth); best fixed ϵ : the swept bandwidth with the highest discovery; range: lowest and highest discovery ov er the fixed bandwidths. Mode discovery: each Phase-1 sample is assigned to the nearest reference-mode centre in (θ3,θ4) if it lies within 1.5× that mode’s 95th-percentile reference radius (at most half the distance between mode centres); discovery =4× the smallest per-mode share ( 1 = all mass split equally over the 4 modes, 0 = a mode missed). Samples drawn from the prior score 0.16 .
Figure 14: Phase-1 sensitivity to the surrogate bandwidth ϵ on Bernoulli GLM (10,000 simulations, one observation). Left: C2ST against the reference (lower is better; dotted: 0.5 ). Right: mode fidelity ( = precision × worst-mode recall, log scale; zero drawn at 10−5 ). Solid: each surrogate swept over its own bandwidth parameter (Gaussian-NLL variance, energy-distance length scale, MMD kernel bandwidth); dashed: the surrogate’s profile setting.
Surrogate
profile
ϵ⋆
C2ST (ϵ⋆)
C2ST (0.1ϵ⋆)
C2ST (10ϵ⋆)
Gaussian-NLL
0.957
0.05
0.957
0.998 ( +0.041 )
0.985 ( +0.028 )
Energy distance
0.770
0.914
0.770
0.981 ( +0.211 )
0.986 ( +0.216 )
MMD
1.000
0.54
1.000
1.000 ( +0.000 )
1.000 ( +0.000 )
Appendix
Table 19: Phase-1 sensitivity to the surrogate bandwidth ϵ on Bernoulli GLM (10,000 simulations, one observation). profile : C2ST at each surrogate’s bandwidth and temperature as chosen, without the reference, from the energy-landscape diagnostics. ϵ⋆ : the best fixed bandwidth on the sweep (identified using the reference), with its C2ST, and the C2ST when ϵ⋆ is mis-set by one order of magnitude in either direction (change relative to ϵ⋆ in brackets). Lower is better ( 0.5 = indistinguishable from the reference).
Surrogate
non-trivial band
ϵpeak
fidelity
precision
min-recall
profile
Gaussian-NLL
[0.005,0.16]
0.016
0.466
0.984
0.473
0.363
Energy distance
[0.09,2.9]
0.29
0.813
0.972
0.837
0.716
MMD
–
0.54
0.000
0.000
0.000
0.000
Appendix
Table 20: Phase-1 mode fidelity ( = precision × worst-mode recall; higher is better) versus the surrogate bandwidth ϵ on Bernoulli GLM (10,000 simulations, one observation). Non-trivial band: the range of ϵ with mode fidelity above 0.01 (– if none); ϵpeak : the bandwidth with the highest mode fidelity, with its precision and worst-mode recall; profile : mode fidelity at the surrogate’s profile setting.
Task
Scenario
Prior
log10Rpmarg.
log10Rpell.
Two moons
default
U([−1,1]2)
0.89
1.69
wide10x
U([−10,10]2)
2.89
3.65
wide50x
U([−50,50]2)
4.28
5.04
SLCP
default
U([−3,3]5)
1.57
3.46
wide5x
U([−5,5]5)
1.93
4.05
wide10x
U([−10,10]5)
2.69
4.99
Appendix
Table 21: Prior scenarios and mean log prior-to-posterior volume ratios, averaged over the ten observations. Posterior volumes are estimated from the reference samples as the product of per-dimension 95% marginal widths and as the volume of the sample-covariance ellipsoid. For Gaussian priors, the prior volume is measured in the same way from its covariance (marginal widths 2×1.96σd ). Lotka–Volterra volumes are computed in logθ . Σ0 denotes the default sbibm Bernoulli GLM prior covariance and μ0=(−0.125,−3,−0.125,−3) .
default
wide10x
wide50x
103
104
105
103
104
105
103
104
105
Simulation budget
Phase-1 simulations
536
4768
54 272
524
524
NRE sims added, first round
64
232
728
76
476
4476
76
476
4476
NRE sims added, later rounds
400
5000
45 000
400
9000
95 000
400
9000
95 000
NRE buffer size
1000
10 000
100 000
1000
10 000
100 000
1000
10 000
100 000
Appendix
Table 22: RSBI hyperparameters for the Two Moons results in Figure 4 . One column per prior × budget cell. A value spanning several columns applies to all of them, and — marks a setting that is not used in that cell.
103
104
105
default
medium
large
default
medium
large
default
medium
large
Simulation budget
Phase-1 simulations
700
7000
49 725
50 000
80 000
NRE sims added, first round
300
3000
275
0
20 000
NRE sims added, later rounds
—
—
25 000
25 000
—
NRE buffer size
—
—
100 000
100 000
—
Appendix
Table 23: RSBI hyperparameters for the Lotka–Volterra results in Figure 4 . One column per prior × budget cell. A value spanning several columns applies to all of them, and — marks a setting that is not used in that cell. † the Phase-1 simulations are discarded and the first NRE round trains on fresh draws from the Phase-1 sampler.
103
104
105
default
medium
large
default
medium
large
default
medium
large
Simulation budget
Phase-1 simulations
250
2500
22 980
24 980
24 980
NRE sims added, first round
0
0
2020
20
20
NRE sims added, later rounds
750
7500
75 000
NRE buffer size
1000
10 000
100 000
Appendix
Table 24: RSBI hyperparameters for the Bernoulli GLM results in Figure 4 . One column per prior × budget cell. A value spanning several columns applies to all of them, and — marks a setting that is not used in that cell.
103
104
105
default
wide5x
wide10x
default
wide5x
wide10x
default
wide5x
wide10x
Simulation budget
Phase-1 simulations
576
4896
13 248
NRE sims added, first round
424
104
752
NRE sims added, later rounds
—
5000
86 000
NRE buffer size
—
8096
50 000
Appendix
Table 25: RSBI hyperparameters for the SLCP results in Figure 4 . One column per prior × budget cell. A value spanning several columns applies to all of them, and — marks a setting that is not used in that cell.
Department of Computer Science, Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic · Institute of Plasma Physics of the Czech Academy of Sciences, Prague, Czech Republic
Department of Neurophysics, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany · International Max Planck Research School on Cognitive NeuroImaging, Leipzig, Germany · Methods and Development Group Neural Data Science and Statistical Computing, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany +3