skip to main content

Title: Certifying the True Error: Machine Learning in Coq with Verified Generalization Guarantees
We present MLCERT, a novel system for doing practical mechanized proof of the generalization of learning procedures, bounding expected error in terms of training or test error. MLCERT is mechanized in that we prove generalization bounds inside the theorem prover Coq; thus the bounds are machine checked by Coq’s proof checker. MLCERT is practical in that we extract learning procedures defined in Coq to executable code; thus procedures with proved generalization bounds can be trained and deployed in real systems. MLCERT is well documented and open source; thus we expect it to be usable even by those without Coq expertise. To validate MLCERT, which is compatible with external tools such as TensorFlow, we use it to prove generalization bounds on neural networks trained using TensorFlow on the extended MNIST data set.
Authors:
;
Award ID(s):
1657358
Publication Date:
NSF-PAR ID:
10138645
Journal Name:
Proceedings of the AAAI Conference on Artificial Intelligence
Volume:
33
Page Range or eLocation-ID:
2662 to 2669
ISSN:
2159-5399
Sponsoring Org:
National Science Foundation
More Like this
  1. Compositional compiler verification is a difficult problem that focuses on separate compilation of program components with possibly different verified compilers. Logical relations are widely used in proving correctness of program transformations in higher-order languages; however, they do not scale to compositional verification of multi-pass compilers due to their lack of transitivity. The only known technique to apply to compositional verification of multi-pass compilers for higher-order languages is parametric inter-language simulations (PILS), which is however significantly more complicated than traditional proof techniques for compiler correctness. In this paper, we present a novel verification framework for lightweight compositional compiler correctness . Wemore »demonstrate that by imposing the additional restriction that program components are compiled by pipelines that go through the same sequence of intermediate representations , logical relation proofs can be transitively composed in order to derive an end-to-end compositional specification for multi-pass compiler pipelines. Unlike traditional logical-relation frameworks, our framework supports divergence preservation—even when transformations reduce the number of program steps. We achieve this by parameterizing our logical relations with a pair of relational invariants . We apply this technique to verify a multi-pass, optimizing middle-end pipeline for CertiCoq, a compiler from Gallina (Coq’s specification language) to C. The pipeline optimizes and closure-converts an untyped functional intermediate language (ANF or CPS) to a subset of that language without nested functions, which can be easily code-generated to low-level languages. Notably, our pipeline performs more complex closure-allocation optimizations than the state of the art in verified compilation. Using our novel verification framework, we prove an end-to-end theorem for our pipeline that covers both termination and divergence and applies to whole-program and separate compilation, even when different modules are compiled with different optimizations. Our results are mechanized in the Coq proof assistant.« less
  2. We consider settings in which the right notion of fairness is not captured by simple mathematical definitions (such as equality of error rates across groups), but might be more complex and nuanced and thus require elicitation from individual or collective stakeholders. We introduce a framework in which pairs of individuals can be identified as requiring (approximately) equal treatment under a learned model, or requiring ordered treatment such as "applicant Alice should be at least as likely to receive a loan as applicant Bob". We provide a provably convergent and oracle efficient algorithm for learning the most accurate model subject tomore »the elicited fairness constraints, and prove generalization bounds for both accuracy and fairness. This algorithm can also combine the elicited constraints with traditional statistical fairness notions, thus "correcting" or modifying the latter by the former. We report preliminary findings of a behavioral study of our framework using human-subject fairness constraints elicited on the COMPAS criminal recidivism dataset.« less
  3. Given a mixture between two populations of coins, “positive” coins that each have unknown and potentially different—bias ≥ 1 + ∆ and “negative” coins with bias ≤ 2 − ∆, we consider the task of estimating the fraction ρ of positive coins to within additive error E. We achieve an upper and lower bound of Θ( ρ log 1 ) samples for a 1 −δ probability of success, where crucially, our lower bound applies to all fully-adaptive algorithms. Thus, our sample complexity bounds have tight dependence for every relevant problem parameter. A crucial component of our lower bound proof ismore »a decomposition lemma (Lemma 5.2) showing how to assemble partially-adaptive bounds into a fully-adaptive bound, which may be of independent interest: though we invoke it for the special case of Bernoulli random variables (coins), it applies to general distributions. We present sim- ulation results to demonstrate the practical efficacy of our approach for realistic problem parameters for crowdsourcing applications, focusing on the “rare events” regime where ρ is small. The fine-grained adaptive flavor of both our algo- rithm and lower bound contrasts with much previous workin distributional testing and learning.« less
  4. Formally verified correctness is one of the most desirable properties of software systems. But despite great progress made via interactive theorem provers, such as Coq, writing proof scripts for verification remains one of the most effort-intensive (and often prohibitively difficult) software development activities. Recent work has created tools that automatically synthesize proofs or proof scripts. For example, CoqHammer can prove 26.6% of theorems completely automatically by reasoning using precomputed facts, while TacTok and ASTactic, which use machine learning to model proof scripts and then perform biased search through the proof-script space, can prove 12.9% and 12.3% of the theorems, respectively.more »Further, these three tools are highly complementary; together, they can prove 30.4% of the theorems fully automatically. Our key insight is that control over the learning process can produce a diverse set of models, and that, due to the unique nature of proof synthesis (the existence of the theorem prover, an oracle that infallibly judges a proof's correctness), this diversity can significantly improve these tools' proving power. Accordingly, we develop Diva, which uses a diverse set of models with TacTok's and ASTactic's search mechanism to prove 21.7% of the theorems. That is, Diva proves 68% more theorems than TacTok and 77% more than ASTactic. Complementary to CoqHammer, Diva proves 781 theorems (27% added value) that Coq-Hammer does not, and 364 theorems no existing tool has proved automatically. Together with CoqHammer, Diva proves 33.8% of the theorems, the largest fraction to date. We explore nine dimensions for learning diverse models, and identify which dimensions lead to the most useful diversity. Further, we develop an optimization to speed up Diva's execution by 40X. Our study introduces a completely new idea for using diversity in machine learning to improve the power of state-of-the-art proof-script synthesis techniques, and empirically demonstrates that the improvement is significant on a dataset of 68K theorems from 122 open-source software projects.« less
  5. We present calf , a c ost- a ware l ogical f ramework for studying quantitative aspects of functional programs. Taking inspiration from recent work that reconstructs traditional aspects of programming languages in terms of a modal account of phase distinctions , we argue that the cost structure of programs motivates a phase distinction between intension and extension . Armed with this technology, we contribute a synthetic account of cost structure as a computational effect in which cost-aware programs enjoy an internal noninterference property: input/output behavior cannot depend on cost. As a full-spectrum dependent type theory, calf presents a unifiedmore »language for programming and specification of both cost and behavior that can be integrated smoothly with existing mathematical libraries available in type theoretic proof assistants. We evaluate calf as a general framework for cost analysis by implementing two fundamental techniques for algorithm analysis: the method of recurrence relations and physicist’s method for amortized analysis . We deploy these techniques on a variety of case studies: we prove a tight, closed bound for Euclid’s algorithm, verify the amortized complexity of batched queues, and derive tight, closed bounds for the sequential and parallel complexity of merge sort, all fully mechanized in the Agda proof assistant. Lastly we substantiate the soundness of quantitative reasoning in calf by means of a model construction.« less