SLICE
A likelihood-free sequential inference framework that combines conformal inference with e-values. By constructing test martingales, it provides exact and anytime-valid confidence sequences for model parameters whose likelihoods are intractable.
Timeline (last 90 days)
SLICE (Sequential Likelihood-Free Inference via Conformal E-Values) combines conformal inference with e-values to construct exact and anytime-valid confidence sequences for parameters of models with intractable likelihoods.
Upon each new observation, SLICE generates simulated data from the corresponding model for each candidate null hypothesis parameter value, and uses it to conformalize positive-predictable scores, converting them into conditional e-values.
The validity of the confidence sequences constructed by SLICE is independent of the accuracy of the scoring function; the quality of the score only affects the width of the interval (statistical efficiency) and does not compromise the coverage guarantee.
The oracle likelihood ratio score maximizes the expected logarithmic growth rate and grows the test martingale fastest among all valid scores, thereby most quickly ruling out incorrect parameters and tightening the confidence sequence.
SLICE uses an amortized neural ratio estimator to approximate the true likelihood ratio, which is the mainstream tool in the simulation-based inference field in recent years.
In synthetic benchmark experiments, the confidence sequences of SLICE remain valid under continuous monitoring, while traditional fixed-sample-size methods fail under the same continuous testing.
SLICE can make statistically guaranteed inferences about the model routing behavior of large language model APIs, and can be used to audit the behavior of many commercial APIs that route requests to different models in the background.
SLICE is validated on real astrophysical data from microlensing.
Experiments show that the efficiency of neural scoring is on par with the exact likelihood ratio and scores tailored to specific applications.
All Facts (9)
SLICE (Sequential Likelihood-Free Inference via Conformal E-Values) combines conformal inference with e-values to construct exact and anytime-valid confidence sequences for parameters of models with intractable likelihoods.
50%UnverifiedUpon each new observation, SLICE generates simulated data from the corresponding model for each candidate null hypothesis parameter value, and uses it to conformalize positive-predictable scores, converting them into conditional e-values.
50%UnverifiedThe validity of the confidence sequences constructed by SLICE is independent of the accuracy of the scoring function; the quality of the score only affects the width of the interval (statistical efficiency) and does not compromise the coverage guarantee.
50%UnverifiedThe oracle likelihood ratio score maximizes the expected logarithmic growth rate and grows the test martingale fastest among all valid scores, thereby most quickly ruling out incorrect parameters and tightening the confidence sequence.
50%UnverifiedSLICE uses an amortized neural ratio estimator to approximate the true likelihood ratio, which is the mainstream tool in the simulation-based inference field in recent years.
50%UnverifiedIn synthetic benchmark experiments, the confidence sequences of SLICE remain valid under continuous monitoring, while traditional fixed-sample-size methods fail under the same continuous testing.
50%UnverifiedSLICE can make statistically guaranteed inferences about the model routing behavior of large language model APIs, and can be used to audit the behavior of many commercial APIs that route requests to different models in the background.
50%UnverifiedSLICE is validated on real astrophysical data from microlensing.
50%UnverifiedExperiments show that the efficiency of neural scoring is on par with the exact likelihood ratio and scores tailored to specific applications.
50%