Approximate Bayesian computation (ABC) enables Bayesian inference for models with intractable likelihoods by estimating parameter posterior distributions from simulated data. Sequential ABC methods improve sampling efficiency by iteratively adapting the proposal distribution toward regions of high posterior probability. However, these methods typically operate in generations, where each generation must wait for its slowest simulation before the next can begin. On a parallel machine, this synchronization barrier leaves workers idle and reduces throughput, particularly when simulation runtimes vary substantially. We introduce a generation-free ABC sampler that updates its proposal when simulation results arrive and sends a new candidate to each available worker. The proposal is based on the best particles evaluated so far. To estimate the posterior, we reuse the full history and reweight particles against a reconstructed mixture of sampling proposals. We show how errors in approximating the mixture of sampling proposals affect the estimated posterior and establish consistency for the smooth-ABC posterior when the approximation becomes asymptotically exact. In our experiments, the method achieves well-calibrated posterior estimates where a pyABC baseline with the same kernel and bandwidth schedule under-covers. From the asynchronous run's timing alone, we predict how much a synchronization barrier would slow the sampler. These predictions agree with measured synchronization penalties ranging from 1.2x to 400x. On a Cellular Potts tissue simulator, the asynchronous sampler completes 2.1-2.4x as many evaluations as pyABC in the same hour, and the discrepancy threshold for its best 100 evaluations is 4.1x smaller. With the starting bandwidth set from prior-predictive discrepancies, it also produces a more concentrated two-parameter posterior with a larger effective sample size than pyABC.
Figures & tables
Figure 1: One generation of ABC-PMC (a) beside one update of our sampler (b), step by step on a two-dimensional toy problem. ABC-PMC proposes a whole population from one mixture, simulates all of it, keeps the accepted particles and only then forms the next proposal, so a generation ends with its slowest simulation. Our sampler appends every completed evaluation to a history and rebuilds the archive, bandwidth and proposal from that history at every call. It draws one candidate for the free worker and weights that candidate against the current and stored proposals rather than against a single proposal. Row (a) is a schematic of textbook hard-threshold PMC, not the smooth-kernel pyABC baseline of § 5.2 . Appendix C explains the symbols used in the figure.
Benchmark
d
cost/sim
runtime spread
budget, workers
role
Gaussian mean
1
0.1 ms
injected: straggler 0 – 20× ; lognormal σ≤2
300 s, 48 ( 16 in the straggler study)
calibration; barrier cost under a known runtime distribution; the near-instantaneous end of the cost boundary
g-and-k
4
2 ms
none
600 s, 48
4-D calibration and recovery against a reference posterior
Lotka–Volterra
4
4 ms
trajectory-dependent; 98% extinct
600 s, 48 ; scaling to 288
strong scaling; the cost boundary
Cellular Potts 503
2
13 s
CV 0.05
3600 s, 48 ; scaling to 384
posterior recovery against a known truth; barrier cost on a real simulator
Cellular Potts 803
2
170 s
CV 0.13 – 0.23
3600 s, 48
barrier cost at realistic cost and heterogeneity
Table 1: Benchmark suite. Cost is the mean wall-clock time of one simulation in the asynchronous runs (an evaluation on Cellular Potts is four replicate simulations); runtime spread is the in-run coefficient of variation of that time, or how it is injected. Budgets are per replicate; five replicates throughout except three at 803 .
Figure 2: The barrier’s cost predicted from the asynchronous run’s timing alone (§ 5.3 ) against the cost measured with the synchronized variant, on every configuration the synchronized variant was run on: a persistent straggler at five slowdowns, lognormal runtime heterogeneity at five spreads, and Cellular Potts at 503 ( 48 – 384 workers, earlier parameterization, utilization ratio) and 803 ( 48 workers, against pyABC). Points show replicate medians and ranges. The dashed line indicates equal predicted and measured costs; open markers identify the three configurations outside the predictor’s domain.
Figure 3: Strong scaling at a fixed wall-clock budget against pyABC with population max(100,W) . Left: Lotka–Volterra ( 4 ms; 180 s; medians and inter-quartile ranges over five replicates). Right: Cellular Potts 503 (earlier parameterization, 5.1 s; 1800 s; log–log), with linear scaling dotted.
Benchmark
cost/sim
sims (async / sync)
throughput
per-simulation
equal wall clock
Gaussian mean
0.1 ms
1.31 M / 4.61 M
0.28×
0.78× ( 0.77 – 1.12 )
0.24× ( 0.21 – 0.28 )
g-and-k
2 ms
1.75 M / 3.81 M
0.46×
1.06× ( 1.02 – 1.15 )
0.88× ( 0.85 – 0.92 )
Lotka–Volterra
4 ms
1.51 M / 1.56 M
0.96×
1.30× ( 1.08 – 1.51 )
1.30× ( 1.02 – 1.43 )
Cellular Potts
13 s
12,900 / 5,400
2.39×
1.6× ( 1.14 – 1.94 )
4.09× ( 3.1 – 6.6 )
Table 2: Tolerance at a matched budget on every benchmark ( k=100 order statistic of each method’s own discrepancies; five replicates; medians, with the replicate range of each ratio in parentheses; completed simulations only). Per-simulation: ϵ of pyABC over the asynchronous sampler at pyABC’s own simulation count; equal wall clock: the same ratio at the end of both runs. Bold marks the larger simulation count.
Figure 4: Throughput ratio and per-simulation efficiency of the asynchronous sampler over pyABC against the cost of one simulation, one point per benchmark (Table 2 ).
Target
Method
0.50
0.80
0.90
0.95
Gaussian mean (1-D)
asynchronous
0.51
0.80
0.90
0.94
pyABC
0.35
0.63
0.75
0.83
Bimodal N(θ2,σ2) (1-D)
asynchronous
0.50
0.82
0.91
0.96
pyABC
0.45
0.82
0.90
0.95
g-and-k (4-D), A
asynchronous, full history
0.62
0.91
0.96
0.99
pyABC
0.54
0.82
0.91
0.95
Table 3: Simulation-based calibration of the posterior estimate ( 1000 trials; empirical coverage of central intervals at the nominal levels; Monte Carlo standard error ±0.007 – 0.016 ). Nominal is best, and bold marks the method closer to it.
Figure 5: Each method’s own posterior (§ 5.3 ) scored against the analytic posterior (a) and the g-and-k reference posterior (b) over wall-clock prefixes. The upper panels show median W1 and interquartile ranges over five replicates. The lower panels show posterior effective sample size; dotted lines indicate the number of evaluated particles.
Figure 6: Cellular Potts posterior (division rate, cell volume) on the fixed production run, replicate 0 : the asynchronous AMIS estimator over the full history, pyABC’s last completed generation with its population weights, and the 100 best rejection draws, each drawn as a weighted kernel density estimate of the method’s own particles (effective sample sizes in the legend) (Table 4 ). Axes are the log-uniform prior coordinates ticked in physical units; the joint panel outlines each method’s 90% highest-density region, and dashed lines mark the truth.
contraction
sims
util.
ϵ(100) , end
division rate
cell volume
ESS
Asynchronous, ϵ0=10
13,300
99.8%
6.3×10−5
84±6%
13±14%
645
Asynchronous, ϵ0=0.1
12,900
99.8%
7.6×10−5
92±1%
81±1%
303
pyABC, ϵ0=0.1
5,400
42.8%
3.1×10−4
91±1%
72±4%
66
Rejection ABC, best 100 of 13,000
13,000
—
2.9×10−3
91%
67%
100
Table 4: Cellular Potts 503 , five replicates per adaptive method and one rejection run, 3600 s on 48 workers (means ± standard deviation for contraction, medians otherwise). The first row is the asynchronous sampler at the benchmark’s original starting bandwidth ( 10 , twenty times the prior-predictive median discrepancy), the rest at the production value ( 0.1 ). Rejection ABC is one run of 13,000 prior draws on all 48 workers, keeping the best 100 . Contraction is 1−sd/sdprior ; the truth lies inside the central 90% interval of every run. Bold marks the best value per column.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Assumption
Status
Basis
1 kernel, Lipschitz part
by construction
Gaussian and Epanechnikov are peak-normalized and Lipschitz in ϵ on every [a,ϵ0] , with LK(a)=2/ea and 2/a . The hard indicator is not, and is excluded from the theory.
1 Zϵ∞>0
holds for Gaussian
The Gaussian kernel is strictly positive, so Lϵ∞>0 and the limiting target is non-degenerate whatever the discrepancy distribution. For a compactly supported kernel (Epanechnikov) it requires the achievable discrepancies to reach inside the support, which we do not verify.
2 denominator class
by construction
The prior support is a box and the archive lies in it, so component means do; the covariance carries an additive jitter 10−9tr(Σ)/d+10−12bˉ2 ( bˉ2 the mean squared box width), whose second term is a run-independent floor, giving σ−2>0 , and σ+2 follows from the archive lying in a bounded box. The truncation normalizer is the product of per-axis in-box masses, so with a correlated covariance qˉn does not integrate to one. Assumption 2 does not require that: it constrains a positive function, and the discrepancy is carried by r instead (§ A.1 (c)).
4 execution model
holds; implementation property
(B1), (B2): the propagator draws from its own stream, seeded once per rank, and every simulator call receives a fresh seed derived from the run seed, the rank and a per-rank call counter, so the two streams are independent in the sense in which pseudo-random streams are, and no two calls share a seed. (B3): a worker proposes when its own simulation completes and builds the proposal from the results that have reached it; delivery is asynchronous message passing, exogenous to both streams. Lemma 5 then gives the adapted filtration with q~i the emitted density. What asynchrony leaves in r is the arrival-order reconstruction of the snapshots, a boundary of at most W particles per snapshot (§ A.1 (d)).
3 prior floor
by construction
The retroactive denominator floors the prior component’s mass at δ=0.5/(m+1) (Appendix C ), so qˉn≥δπ and Lemma 6 bounds every posterior weight by 1/δ , independently of what the sampler does.
5 fidelity ( 9 )
assumed; two of its four factors measured
The one place the sampler enters for Corollary 2 ; Theorem 1 does not use it. Decomposed into four factors in § A.1 , two of them estimated from stored histories rather than bounded; since the factors multiply, the measured part is a partial accounting rather than a bound on the whole.
Appendix
Table 5: What the implementation guarantees, what it assumes, and what it violates. “By construction” means the code enforces the property for every run, not that it was checked post hoc.
Property
Classical ABC-SMC
This work
Update style
generation, barrier-synchronized
event-driven (single-arrival), no barrier
Archive
explicit, generational
reconstructed, sliding top- k
ABC likelihood
hard threshold
smooth kernel
Importance weight
against qt−1 only
balance heuristic over S
Archive state
mutable, carried
none; reconstructed per call
Posterior estimate
final population
pure function of the history
Appendix
Table 6: A typical generation-staged ABC-SMC/PMC implementation (the pyABC baseline used here) versus the proposed generation-free method. The listed properties characterize pyABC as configured, not ABC-SMC as a family. The running sampler does carry the online snapshot buffer of § 3.5 ; the posterior estimate rebuilds it from the history.
Symbol
Meaning
Option
Value
k
archive size
k
100 ( 200 in the four-dimensional calibration runs)
ϵ0
starting bandwidth
tol_init
5 (Gaussian mean), 10 (g-and-k), 5×105 (Lotka–Volterra), 0.1 (Cellular Potts; 10 in its earlier parameterization)
Kϵ
ABC kernel
kernel
Gaussian
s
scale of the perturbation covariance Σn
perturbation_scale
0.8
S
kept snapshots
amis_snapshots
20
snapshot cadence, in calls
amis_interval
k
Appendix
Table 7: Settings of the asynchronous sampler. Option names are those of the implementation; values are those of the production experiments, except where a study sweeps them (Appendix F ).
Experiment
Workers
Budget (s)
Reps
k
Methods
Gaussian mean
48
300
5
100
A,S,R
g-and-k
48
600
5
100
A,S,R
Lotka–Volterra
48
600
5
100
A,S
Cellular Potts 503
48
3600
5
100
A,S,R
Cellular Potts 803
48
3600
3
100
A,S,R
Straggler
16
300
5
100
A,S
Appendix
Table 8: Resolved experiment configurations. Workers, per-run wall-clock budget, replicates (trials for SBC), archive size k , and methods (A: asynchronous; S: pyABC (matched kernel); R: rejection ABC; T: synchronized variant). The synchronized variant row covers the straggler, heterogeneity and Cellular Potts 503 studies at 16 , 48 and 48 – 384 workers. † Simulation-limited: the budget is per-trial (SBC) or per-run (kill-and-resume) simulations , not wall-clock seconds; the synchronized variant, and the asynchronous re-run matched to it, stop at a fixed simulation count per level (§ 5.2 , Table 9 ). ‡ Also run with AMIS on/off, a drain-after-deadline variant, and a steeper coupling regime (base delay 0.5 s, cap 8 s). § The Cellular Potts rows of these studies ran the benchmark’s earlier configuration (Appendix E ).
Figure 7: Inference per unit wall clock under runtime heterogeneity (Gaussian mean, analytic posterior, 48 workers, 60 s budget, five replicates; bands are inter-quartile ranges). Left: error of the archive mean (the unweighted top- k archive, not πn ; § 6.4 ) to the analytic posterior mean, against the runtime spread σ . Right: simulations completed within the budget. The barrier starves pyABC as σ grows; the asynchronous sampler’s error stays flat.
Straggler slowdown
0×
1×
5×
10×
20×
Asynchronous, sims/s
3888
3112
3241
3218
3205
Synchronized variant, barrier every W
48.7
65.3
31.6
15.9
8.0
Synchronized variant, barrier every 112
162.8
148.4
31.8
15.9
8.0
Synchronized variant, predicted
(3870)
(154)
31.8
15.9
7.98
Ratio, measured (every W )
80×
48×
103×
202×
402×
Ratio, predicted
—
—
102×
202×
401×
Appendix
Table 9: Synchronized variant on the straggler benchmark ( W=16 , five replicates, medians). Throughput in simulations per second; predicted as in § 5.3 , and at 0× and 1× the collective’s latency, absent from the model, dominates. Lower block: W1 of each sampler’s posterior to the analytic posterior at the synchronized variant’s simulation count and seeds, at the bandwidth its schedule reached and, last row, re-reported at ϵ(100) (§ 6.4 ).
Figure 8: pyABC against the asynchronous sampler under injected runtime heterogeneity on the Gaussian mean (§ 6.1 ). (a) Throughput under a persistent straggler: one of 16 workers carries a permanent post-evaluation delay of 0.1 s scaled by the factor ( 300 s; medians and inter-quartile ranges over five replicates; logarithmic axis). (b) Fraction of worker time spent idle under a lognormal runtime multiplier of spread σ ( 48 workers, 60 s).
Heterogeneity σ
0
0.5
1.0
1.5
2.0
Asynchronous, sims/s
47.6
42.1
28.9
17.6
11.4
Synchronized variant, barrier every W
44.8
15.1
4.37
1.06
0.37
Synchronized variant, barrier every 96
45.1
19.8
6.50
1.75
0.45
pyABC
23.1
18.9
9.91
4.35
2.39
Ratio, measured (every W )
1.06×
2.8×
6.6×
16.6×
29.6×
Ratio, predicted
1.00×
2.77×
6.34×
14.0×
35.4×
Appendix
Table 10: Synchronized variant under runtime heterogeneity ( W=48 , lognormal runtime multiplier of scale σ , five replicates, medians of simulations per second of each sampler’s own elapsed wall clock). Predicted is from the injected runtime distribution and the asynchronous run’s measured overhead. The σ=0 control costs almost nothing because these evaluations take 1 s and barrier latency is negligible against them; the cost is the wait for the slowest draw and grows with σ . pyABC degrades less than the synchronized variant because it discards latecomers instead of waiting for them. The lower block scores the posteriors against the analytic posterior at the bandwidth each schedule reached, then re-reported at the 100 th order statistic of the same history (§ 6.4 ); no rank made k calls in the asynchronous run, so its schedule never ran at σ≤0.5 , and the synchronized variant’s ran only at σ≤0.5 .
Workers ( = cores)
48
96
192
384
Throughput ratio, async / synchronized variant
1.91×
1.87×
1.86×
2.82×
= utilization ratio
1.18×
1.19×
1.21×
1.21×
× simulation-duration ratio
1.63×
1.59×
1.52×
2.33×
Straggler factor of the synchronized variant’s generations
1.21
1.28
1.21
1.29
Predicted from asynchronous timing
1.26×
1.30×
1.32×
( 2.34× )
Appendix
Table 11: The synchronized variant’s slowdown on Cellular Potts 503 (earlier parameterization, Appendix E ), factorized (five replicates, medians). Throughput is utilization over mean simulation time, so the ratio is the product of the two columns that follow. The utilization ratio is what the barrier costs; it agrees with the expected-maximum-to-mean runtime ratio measured within the synchronized variant’s generations and with the prediction from the asynchronous run’s runtime distribution. The duration ratio (the synchronized variant’s simulations running longer, with replicate means from 6 to 36 s against 5.1 s on every asynchronous replicate) is not a barrier effect and is not attributed here.
Figure 9: Tolerance against simulations spent, ϵ(100)(n) , asynchronous against pyABC on every benchmark (medians and replicate range over five replicates; the slope over the last decade of n in the legend). The vertical gap at equal n is the per-simulation efficiency, the horizontal extent is the extra simulations the wall clock allowed. On Lotka–Volterra 98% of simulations return the extinction fallback, so ϵ(100) sits on that plateau until about 5000 draws have produced 100 survivors; the plateau is cut off.
Figure 10: Parameter-coupled runtime on the Gaussian-mean benchmark (analytic posterior; 48 workers, 60 s budget, five replicates; bands are inter-quartile ranges). The post-evaluation delay grows with the inferred mean, so one region of parameter space is sampled far more often than its mirror image. (a) Aggregate throughput falls as the coupling steepens (monotonically for pyABC, and from a peak at c=0.5 for the asynchronous sampler), consistent with the fast region dominating the raw sample counts. (b) Posterior-mean error to the analytic posterior nonetheless does not increase with coupling and stays comparable between the asynchronous sampler and pyABC, which discards latecomers, so the over-representation does not measurably distort the posterior estimate at these coupling levels.
Workers
1
4
16
48
144
192
240
288
Nodes
< 1
< 1
< 1
1
3
4
5
6
Synchronous
105
96
437
2952
3757
4322
4539
6034
Asynchronous ( k=100 )
149
369
1983
6240
4195
2615
2088
1729
Asynchronous ( k=1000 )
106
133
620
3804
2804
1957
1587
1336
Appendix
Table 12: Strong scaling on Lotka–Volterra: median simulation throughput (sims/s) at a 180 s budget, over five replicates (Fig. 3 , left, shows the inter-quartile spread). Asynchronous (ours) at both archive sizes vs. pyABC. Multi-node points use fully packed nodes (worker counts that are exact multiples of 48 ); at counts above 100 pyABC is given population size equal to the worker count so it can occupy every core (a population of 100 leaves the surplus idle), which is why it keeps scaling here. The asynchronous sampler peaks on one node and then declines as multi-node coordination dominates, so pyABC overtakes it beyond three nodes at k=100 . Enlarging the archive tenfold costs throughput at every scale, because the per-arrival proposal reconstruction is O(k) , and moves that crossover in from beyond three nodes to three.
Figure 11: Measured decomposition of asynchronous worker wall-time on Lotka–Volterra ( k=100 ; 180 s budget) from a dedicated instrumented run. The productive simulation fraction (blue, annotated), the share of worker wall-clock spent inside the simulator, collapses from 55% at 16 workers to 2% at 256 , while coordination and idle (MPI exchange, per-arrival proposal reconstruction, serialization, waiting) dominate as the job spreads beyond one node (16/48 = 1 node; 128 = 3 nodes; 256 = 6 nodes). The strong-scaling decline of Fig. 3 is therefore a measured property of coordinating low-cost evaluations.
archive size k (at S=20 )
50
100
200
400
800
d=1
-0.014
-0.003
+0.006
+0.006
+0.074
d=2
-0.020
-0.006
+0.009
+0.030
+0.072
d=4
+0.025
+0.064
+0.094
+0.133
+0.063
d=8
-0.250
+0.055
+0.180
+0.195
+0.061
d=16
-0.047
+0.028
+0.049
+0.053
+0.055
Appendix
Table 13: Calibration sensitivity to the archive size k and the AMIS snapshot buffer S on the d -dimensional Gaussian-mean target ( 1000 trials per cell; entries are the mean signed deviation of empirical from nominal coverage over the 0.50/0.80/0.90/0.95 levels and the d parameters). Negative is over-confident, positive is conservative; 0.000 is exactly calibrated. Left: varying k at S=20 . Right: varying S at k=100 .
k
sims/s at W=48
vs. k=100
worst ∣Δcov∣
at
pyABC overtakes
50
6684
+7%
0.250
d=8
beyond 3 nodes
100
6240
—
0.064
d=4
beyond 3 nodes
200
5904
−5%
0.180
d=8
beyond 3 nodes
400
5393
−14%
0.195
d=8
3 nodes
800
4331
−31%
0.074
d=1
3 nodes
1000
3804
−39%
—
—
3 nodes
Appendix
Table 14: The archive size on both of its axes at once. Throughput : median asynchronous throughput at the single-node peak ( 48 workers) on Lotka–Volterra, 180 s wall cap, five replicates ( k∈{100,1000} from the main scaling run, k∈{50,200,400,800} from the matched k -frontier run), and its change relative to the k=100 we use throughout. Calibration : worst absolute deviation of empirical from nominal coverage over the dimensions of the d -dimensional Gaussian-mean sweep at S=20 (Table 13 ), with the dimension attaining it. Crossover : fully packed nodes at which pyABC (population = worker count) overtakes the asynchronous sampler. The sweep did not include k=1000 . Bold marks the best value per column.
Figure 12: Simulation-based calibration on Gaussian-mean inference (1000 trials; the full-history AMIS posterior): empirical coverage versus nominal level (left) and rank histogram (right). The asynchronous estimator is close to calibrated.
Figure 13: Cellular Potts 503 worker utilization, earlier parameterization (fraction of worker wall-clock spent inside the simulator) versus worker count ( 1800 s budget; error bars are ±1 standard deviation); pyABC uses a population equal to the worker count (Fig. 3 , right). The asynchronous sampler stays ≈98.9 – 99.8% utilized at every scale. pyABC fills every core within a generation but then idles at the barrier waiting for the slowest simulation, so its utilization falls to ≈24 – 29% ( 41%→24%→26%→29% at 48/96/192/384 workers, i.e. ≈70 – 76% idle at every multi-node scale). The figure is the measured utilization behind pyABC’s sub-linear scaling in Fig. 3 (right).
Neural posterior estimation (NPE) enables Bayesian inference of the posterior parameter in agent-based models (ABMs) with intractable likelihoods, but simulating training data can be computationally expensive. We propose Double-Flow Neural Posterior Estimation (DF-NPE), which combines learned time-series summaries with two conditional normalizing flows for synthetic-data generation and posterior estimation. The framework supplements direct simulations with synthetic training pairs and supports both pooled and ordered training strategies. Experiments in two settings of the Brock-Hommes asset-pricing model show that DF-NPE can reduce direct simulations by 90 percent while achieving lower mean posterior discrepancies and runtime than standard NPE. Performance varies with the allocation of directly simulated and synthetic data and the training strategy. These results establish learned summary augmentation as a practical approach to reducing simulation costs for Bayesian posterior inference of parameters in ABMs.
Junyao Zhang, Jinglai Li, Junqi Tang
School of Mathematics, University of Birmingham, Birmingham, United Kingdom
Asynchronous Bayesian optimization is widely used for gradient-free optimization in domains with independent parallel experiments and varying evaluation times. Existing methods posit that standard acquisitions lead to redundant and repeated queries, proposing complex solutions to enforce diversity in queries. Challenging this fundamental premise, we show that methods, like the Upper Confidence Bound, can in fact achieve theoretical guarantees essentially equivalent to those of sequential Thompson sampling. A conceptual analysis of asynchronous Bayesian optimization reveals that existing works neglect intermediate posterior updates, which we find to be generally sufficient to avoid redundant queries. Further investigation shows that by penalizing busy locations, diversity-enforcing methods can over-explore in asynchronous settings, reducing their performance. Our extensive experiments demonstrate that simple standard acquisition functions match or outperform purpose-built asynchronous methods across synthetic and real-world tasks.
Ben Riegler, James Odgers, Vincent Fortuin
Technical University of Munich Helmholtz AI MCML · UTN Helmholtz AI MCML
Efficient sampling from Bayesian posterior distributions in high dimensions remains a hard problem, despite advances in theory and algorithms. Multiple steps are usually required to draw unbiased samples, for example, iterations of an MCMC algorithm or integration steps in a diffusion-based amortised sampler. However, scalable one-step amortised inference remains a challenging open problem. In this work, we build upon recent progress in data-driven generative modelling to offer a possible solution for one-step amortised sampling. Our one-step implicit sampler (OSIS) uses short-time MCMC dynamics to define a local target for the evolution of an implicitly modelled distribution. In the asymptotic limit, the modelled distribution is a fixed point of the MCMC kernel and thus samples from the target. We analyse the convergence properties of OSIS and show connections with other generative modelling and sampling algorithms. We study various design choices and show that OSIS performs well compared to multi-step amortised samplers and non-amortised MCMC baselines while being far more efficient at sampling time.
Vincent Pauline, Kirill Tamogashev, Arran Carter +3
Technical University of Munich · Munich Center for Machine Learning · University of Edinburgh +1