Decoding with autoregressive language models traditionally occurs sequentially, generating one token after another. Recent attempts to introduce parallelism require a pre-determined structure in the generated content to implement parallel generation, such as by pattern-matching on bullet points. In this work, we present a new technique to automate parallel generation by dynamically exploiting the semantic independence of generation outputs to implement asynchronous decoding. We introduce an annotation language Pasta-Lang for language models to initiate asynchronous decoding at inference time. We also develop an accompanying Pasta-Lang interpreter that performs on-the-fly asynchronous decoding, effectively implementing parallel generation and speeding up inference. We present an instruction-finetuning dataset with Pasta-Lang-annotated responses for teaching LLMs to annotate semantic independence with Pasta-Lang as well as the methodology for creating the dataset. Our evaluation shows using the interpreter with a Pasta-Lang-equipped model achieves significant speedup while maintaining the same generation quality.

We introduce FreshStack, a holistic framework for automatically building information retrieval (IR) evaluation benchmarks by incorporating challenging questions and answers. FreshStack conducts the following steps:(1) automatic corpus collection from code and technical documentation,(2) nugget generation from community-asked questions and answers, and(3) nugget-level support, retrieving documents using a fusion of retrieval techniques and hybrid architectures.We use FreshStack to build five datasets on fast-growing, recent, and niche domains to ensure the tasks are sufficiently challenging. On FreshStack, existing retrieval models, when applied out-of-the-box, significantly underperform oracle approaches on all five domains, denoting plenty of headroom to improve IR quality. In addition, we identify cases where rerankers do not improve first-stage retrieval accuracy (two out of five domains) and oracle context helps an LLM generator generate a high-quality RAG answer.We hope FreshStack will facilitate future work toward constructing realistic, scalable, and uncontaminated IR and RAG evaluation benchmarks.

Advanced probabilistic programming languages (PPLs) using hybrid particle filtering combine symbolic exact inference and Monte Carlo methods to improve inference performance. These systems use heuristics to partition random variables within the program into variables that are encoded symbolically and variables that are encoded with sampled values, and the heuristics are not necessarily aligned with the developer’s performance evaluation metrics. In this work, we present inference plans , a programming interface that enables developers to control the partitioning of random variables during hybrid particle filtering. We further present Siren , a new PPL that enables developers to use annotations to specify inference plans the inference system must implement. To assist developers with statically reasoning about whether an inference plan can be implemented, we present an abstract-interpretation-based static analysis for Siren for determining inference plan satisfiability . We prove the analysis is sound with respect to Siren ’s semantics. Our evaluation applies inference plans to three different hybrid particle filtering algorithms on a suite of benchmarks. It shows that the control provided by inference plans enables speed ups of 1.76 x on average and up to 206 x to reach a target accuracy, compared to the inference plans implemented by default heuristics; the results also show that inference plans improve accuracy by 1.83 x on average and up to 595 x with less or equal runtime, compared to the default inference plans. We further show that our static analysis is precise in practice, identifying all satisfiable inference plans in 27 out of the 33 benchmark-algorithm evaluation settings.

We study how down-scaling large language model (LLM) size impacts LLM capabilities. We begin by measuring the effects of weight pruning – a popular technique for reducing model size – on the two abilities of LLMs: (a) recalling facts presented during pre-training and (b) processing information presented in context. Surprisingly, we find that existing pruning techniques affect these two abilities of LLMs differently. For example, pruning more than 30% of weights significantly decreases an LLM’s ability to recall facts presented during pre-training. Yet pruning 60-70% of weights largely preserves an LLM’s ability to process information in-context, ranging from retrieving answers based on information presented in context to learning parameterized functions such as a linear classifier based on a few examples. Moderate pruning impairs LLM’s ability to recall facts learnt from pre-training. However, its effect on model’s ability to process information presented in context is much less pronounced. The said disparate effects similarly arise when replacing the original model with a smaller dense one with reduced width and depth. This similarity suggests that model size reduction in general underpins the said disparity.

A *neural surrogate* is a neural network that mimics the behavior of a program. Neural surrogates of programs have been used to automatically tune program inputs, adapt programs to new settings, and accelerate computations. Neural surrogates have traditionally been developed by training on input-output examples for a single program. Language models present another approach wherein a model is trained on a single, large dataset then directly consumes program text, to act as a neural surrogate of the program. Having the language model as both the neural surrogate generator and the neural surrogate, however, poses a tradeoff of limited accuracy or excessive resource consumption. We present *neural surrogate compilation*, a technique for producing neural surrogates directly from program text without coupling neural surrogate generation and execution. We implement neural surrogate compilers using hypernetworks trained on a dataset of C programs and find they produce neural surrogates that are $1.91$-$9.50\times$ as data-efficient and train in $4.31$-$7.28\times$ fewer epochs than neural surrogates trained from scratch.

Computations in physical simulation, computer graphics, and probabilistic inference often require the differentiation of discontinuous processes due to contact, occlusion, and changes at a point in time. Popular differentiable programming languages, such as PyTorch and JAX, ignore discontinuities during differentiation. This is incorrect for parametric discontinuities —conditionals containing at least one real-valued parameter and at least one variable of integration. We introduce Potto, the first differentiable first-order programming language to soundly differentiate parametric discontinuities. We present a denotational semantics for programs and program derivatives and show the two accord. We describe the implementation of Potto, which enables separate compilation of programs. Our prototype implementation overcomes previous compile-time bottlenecks achieving an 88.1x and 441.2x speed up in compile time and a 2.5x and 7.9x speed up in runtime, respectively, on two increasingly large image stylization benchmarks. We showcase Potto by implementing a prototype differentiable renderer with separately compiled shaders.

Numerous quantum algorithms require the use of quantum error correction to overcome the intrinsic unreliability of physical qubits. However, quantum error correction imposes a unique performance bottleneck, known as T -complexity, that can make an implementation of an algorithm as a quantum program run more slowly than on idealized hardware. In this work, we identify that programming abstractions for control flow, such as the quantum if-statement, can introduce polynomial increases in the T -complexity of a program. If not mitigated, this slowdown can diminish the computational advantage of a quantum algorithm. To enable reasoning about the costs of control flow, we present a cost model that a developer can use to accurately analyze the T -complexity of a program under quantum error correction and pinpoint the sources of slowdown. To enable the mitigation of these costs, we present a set of program-level optimizations that a developer can use to rewrite a program to reduce its T -complexity, predict the T -complexity of the optimized program using the cost model, and then compile it to an efficient circuit via a straightforward strategy. We implement the program-level optimizations in Spire, an extension of the Tower quantum compiler. Using a set of 11 benchmark programs that use control flow, we empirically show that the cost model is accurate, and that Spire's optimizations recover programs that are asymptotically efficient, meaning their runtime T -complexity under error correction is equal to their time complexity on idealized hardware. Our results show that optimizing a program before it is compiled to a circuit can yield better results than compiling the program to an inefficient circuit and then invoking a quantum circuit optimizer found in prior work. For our benchmarks, only 2 of 8 tested quantum circuit optimizers recover circuits with asymptotically efficient T -complexity. Compared to these 2 optimizers, Spire uses 54 × − 2400 × less compile time.

Quantum algorithms for tasks such as factorization, search, and simulation rely on control flow such as branching and iteration that depends on the value of data in superposition. High-level programming abstractions for control flow, such as switches, loops, higher-order functions, and continuations, are ubiquitous in classical languages. By contrast, many quantum languages do not provide high-level abstractions for control flow in superposition, and instead require the use of hardware-level logic gates to implement such control flow. The reason for this gap is that whereas a classical computer supports control flow abstractions using a program counter that can depend on data, the typical architecture of a quantum computer does not analogously provide a program counter that can depend on data in superposition. As a result, the complete set of control flow abstractions that can be correctly realized on a quantum computer has not yet been established. In this work, we provide a complete characterization of the properties of control flow abstractions that are correctly realizable on a quantum computer. First, we prove that even on a quantum computer whose program counter exists in superposition, one cannot correctly realize control flow in quantum algorithms by lifting the classical conditional jump instruction to work in superposition. This theorem denies the ability to directly lift general abstractions for control flow such as the λ-calculus from classical to quantum programming. In response, we present the necessary and sufficient conditions for control flow to be correctly realizable on a quantum computer. We introduce the quantum control machine, an instruction set architecture featuring a conditional jump that is restricted to satisfy these conditions. We show how this design enables a developer to correctly express control flow in quantum algorithms using a program counter in place of logic gates.

In reinforcement learning, the classic objectives of maximizing discounted and finite-horizon cumulative rewards are PAC-learnable: There are algorithms that learn a near-optimal policy with high probability using a finite amount of samples and computation. In recent years, researchers have introduced objectives and corresponding reinforcement-learning algorithms beyond the classic cumulative rewards, such as objectives specified as linear temporal logic formulas. However, questions about the PAC-learnability of these new objectives have remained open. This work demonstrates the PAC-learnability of general reinforcement-learning objectives through sufficient conditions for PAC-learnability in two analysis settings. In particular, for the analysis that considers only sample complexity, we prove that if an objective given as an oracle is uniformly continuous, then it is PAC-learnable. Further, for the analysis that considers computational complexity, we prove that if an objective is computable, then it is PAC-learnable. In other words, if a procedure computes successive approximations of the objective's value, then the objective is PAC-learnable. We give three applications of our condition on objectives from the literature with previously unknown PAC-learnability and prove that these objectives are PAC-learnable. Overall, our result helps verify existing objectives' PAC-learnability. Also, as some studied objectives that are not uniformly continuous have been shown to be not PAC-learnable, our results could guide the design of new PAC-learnable objectives.

Programmers and researchers are increasingly developing surrogates of programs, models of a subset of the observable behavior of a given program, to solve a variety of software development challenges. Programmers train surrogates from measurements of the behavior of a program on a dataset of input examples. A key challenge of surrogate construction is determining what training data to use to train a surrogate of a given program. We present a methodology for sampling datasets to train neural-network-based surrogates of programs. We first characterize the proportion of data to sample from each region of a program's input space (corresponding to different execution paths of the program) based on the complexity of learning a surrogate of the corresponding execution path. We next provide a program analysis to determine the complexity of different paths in a program. We evaluate these results on a range of real-world programs, demonstrating that complexity-guided sampling results in empirical improvements in accuracy.

In recent years, researchers have made significant progress in devising reinforcement-learning algorithms for optimizing linear temporal logic (LTL) objectives and LTL-like objectives. Despite these advancements, there are fundamental limitations to how well this problem can be solved. Previous studies have alluded to this fact but have not examined it in depth. In this paper, we address the tractability of reinforcement learning for general LTL objectives from a theoretical perspective. We formalize the problem under the probably approximately correct learning in Markov decision processes (PAC-MDP) framework, a standard framework for measuring sample complexity in reinforcement learning. In this formalization, we prove that the optimal policy for any LTL formula is PAC-MDP-learnable if and only if the formula is in the most limited class in the LTL hierarchy, consisting of formulas that are decidable within a finite horizon. Practically, our result implies that it is impossible for a reinforcement-learning algorithm to obtain a PAC-MDP guarantee on the performance of its learned policy after finitely many interactions with an unconstrained environment for LTL objectives that are not decidable within a finite horizon.

Practitioners frequently observe that pruning improves model generalization. A long-standing hypothesis based on bias-variance trade-off attributes this generalization improvement to model size reduction. However, recent studies on over-parameterization characterize a new model size regime, in which larger models achieve better generalization. Pruning models in this over-parameterized regime leads to a contradiction -- while theory predicts that reducing model size harms generalization, pruning to a range of sparsities nonetheless improves it. Motivated by this contradiction, we re-examine pruning’s effect on generalization empirically.We show that size reduction cannot fully account for the generalization-improving effect of standard pruning algorithms. Instead, we find that pruning leads to better training at specific sparsities, improving the training loss over the dense model. We find that pruning also leads to additional regularization at other sparsities, reducing the accuracy degradation due to noisy examples over the dense model. Pruning extends model training time and reduces model size. These two factors improve training and add regularization respectively. We empirically demonstrate that both factors are essential to fully explaining pruning's impact on generalization.

A streaming probabilistic program receives a stream of observations and produces a stream of distributions that are conditioned on these observations. Efficient inference is often possible in a streaming context using Rao-Blackwellized particle filters (RBPFs), which exactly solve inference problems when possible and fall back on sampling approximations when necessary. While RBPFs can be implemented by hand to provide efficient inference, the goal of streaming probabilistic programming is to automatically generate such efficient inference implementations given input probabilistic programs. In this work, we propose semi-symbolic inference, a technique for executing probabilistic programs using a runtime inference system that automatically implements Rao-Blackwellized particle filtering. To perform exact and approximate inference together, the semi-symbolic inference system manipulates symbolic distributions to perform exact inference when possible and falls back on approximate sampling when necessary. This approach enables the system to implement the same RBPF a developer would write by hand. To ensure this, we identify closed families of distributions – such as linear-Gaussian and finite discrete models – on which the inference system guarantees exact inference. We have implemented the runtime inference system in the ProbZelus streaming probabilistic programming language. Despite an average 1.6× slowdown compared to the state of the art on existing benchmarks, our evaluation shows that speedups of 3×-87× are obtainable on a new set of challenging benchmarks we have designed to exploit closed families.

Emerging quantum algorithms for problems such as element distinctness, subset sum, and closest pair demonstrate computational advantages by relying on abstract data structures. Practically realizing such an algorithm as a program for a quantum computer requires an efficient implementation of the data structure whose operations correspond to unitary operators that manipulate quantum superpositions of data. To correctly operate in superposition, an implementation must satisfy three properties --- reversibility, history independence, and bounded-time execution. Standard implementations, such as the representation of an abstract set as a hash table, fail these properties, calling for tools to develop specialized implementations. In this work, we present Core Tower, the first language for quantum programming with random-access memory. Core Tower enables the developer to implement data structures as pointer-based, linked data. It features a reversible semantics enabling every valid program to be translated to a unitary quantum circuit. We present Boson, the first memory allocator that supports reversible, history-independent, and constant-time dynamic memory allocation in quantum superposition. We also present Tower, a language for quantum programming with recursively defined data structures. Tower features a type system that bounds all recursion using classical parameters as is necessary for a program to execute on a quantum computer. Using Tower, we implement Ground, the first quantum library of data structures, including lists, stacks, queues, strings, and sets. We provide the first executable implementation of sets that satisfies all three mandated properties of reversibility, history independence, and bounded-time execution.

Quantum programming languages enable developers to implement algorithms for quantum computers that promise computational breakthroughs in classically intractable tasks. Programming quantum computers requires awareness of entanglement , the phenomenon in which measurement outcomes of qubits are correlated. Entanglement can determine the correctness of algorithms and suitability of programming patterns. In this work, we formalize purity as a central tool for automating reasoning about entanglement in quantum programs. A pure expression is one whose evaluation is unaffected by the measurement outcomes of qubits that it does not own, implying freedom from entanglement with any other expression in the computation. We present Twist, the first language that features a type system for sound reasoning about purity. The type system enables the developer to identify pure expressions using type annotations. Twist also features purity assertion operators that state the absence of entanglement in the output of quantum gates. To soundly check these assertions, Twist uses a combination of static analysis and runtime verification. We evaluate Twist’s type system and analyses on a benchmark suite of quantum programs in simulation, demonstrating that Twist can express quantum algorithms, catch programming errors in them, and support programs that existing languages disallow, while incurring runtime verification overhead of less than 3.5%.

Vector instructions are ubiquitous in modern processors. Traditional compiler auto-vectorization techniques have focused on targeting single instruction multiple data (SIMD) instructions. However, these auto-vectorization techniques are not sufficiently powerful to model non-SIMD vector instructions, which can accelerate applications in domains such as image processing, digital signal processing, and machine learning. To target non-SIMD instruction, compiler developers have resorted to complicated, ad hoc peephole optimizations, expending significant development time while still coming up short. As vector instruction sets continue to rapidly evolve, compilers cannot keep up with these new hardware capabilities.

The computer vision world has been re-gaining enthusiasm in various pre-trained models, including both classical ImageNet supervised pre-training and recently emerged self-supervised pre-training such as simCLR [10] and MoCo [40]. Pre-trained weights often boost a wide range of downstream tasks including classification, detection, and segmentation. Latest studies suggest that pre-training benefits from gigantic model capacity [11]. We are hereby curious and ask: after pre-training, does a pre-trained model indeed have to stay large for its downstream transferability? In this paper, we examine supervised and self-supervised pre-trained models through the lens of the lottery ticket hypothesis (LTH) [31]. LTH identifies highly sparse matching subnetworks that can be trained in isolation from (nearly) scratch yet still reach the full models' performance. We extend the scope of LTH and question whether matching subnetworks still exist in pre-trained computer vision models, that enjoy the same downstream transfer performance. Our extensive experiments convey an overall positive message: from all pre-trained weights obtained by ImageNet classification, simCLR, and MoCo, we are consistently able to locate such matching subnetworks at 59.04% to 96.48% sparsity that transfer universally to multiple downstream tasks, whose performance see no degradation compared to using full pre-trained weights. Further analyses reveal that subnetworks found from different pre-training tend to yield diverse mask structures and perturbation sensitivities. We conclude that the core LTH observations remain generally relevant in the pre-training paradigm of computer vision, but more delicate discussions are needed in some cases. Codes and pre-trained models will be made available at: https://github.com/VITA-Group/CV_LTH_Pre-training.

We show that the error of iteratively magnitude-pruned networks empirically follows a scaling law with interpretable coefficients that depend on the architecture and task. We functionally approximate the error of the pruned networks, showing it is predictable in terms of an invariant tying width, depth, and pruning level, such that networks of vastly different pruned densities are interchangeable. We demonstrate the accuracy of this approximation over orders of magnitude in depth, width, dataset size, and density. We show that the functional form holds (generalizes) for large scale data (e.g., ImageNet) and architectures (e.g., ResNets). As neural networks become ever larger and costlier to train, our findings suggest a framework for reasoning conceptually and analytically about a standard method for unstructured pruning.

Modern software applications are increasingly configurable, which puts a burden on users to tune these configurations for their target hardware and workloads. To help users, machine learning techniques can model the complex relationships between software configuration parameters and performance. While powerful, these learners have two major drawbacks: (1) they rarely incorporate prior knowledge and (2) they produce outputs that are not interpretable by users. These limitations make it difficult to (1) leverage information a user has already collected (e.g., tuning for new hardware using the best configurations from old hardware) and (2) gain insights into the learner’s behavior (e.g., understanding why the learner chose different configurations on different hardware or for different workloads). To address these issues, this paper presents two configuration optimization tools, GIL and GIL+, using the proposed generalizable and interpretable learning approaches. To incorporate prior knowledge, the proposed tools (1) start from known configurations, (2) iteratively construct a new linear model, (3) extrapolate better performance configurations from that model, and (4) repeat. Since the base learners are linear models, these tools are inherently interpretable. We enhance this property with a graphical representation of how they arrived at the highest performance configuration. We evaluate GIL and GIL+ by using them to configure Apache Spark workloads on different hardware platforms and find that, compared to prior work, GIL and GIL+ produce comparable, and sometimes even better performance configurations, but with interpretable results.

Probabilistic programming languages aid developers performing Bayesian inference. These languages provide programming constructs and tools for probabilistic modeling and automated inference. Prior work introduced a probabilistic programming language, ProbZelus, to extend probabilistic programming functionality to unbounded streams of data. This work demonstrated that the delayed sampling inference algorithm could be extended to work in a streaming context. ProbZelus showed that while delayed sampling could be effectively deployed on some programs, depending on the probabilistic model under consideration, delayed sampling is not guaranteed to use a bounded amount of memory over the course of the execution of the program. In this paper, we the present conditions on a probabilistic program’s execution under which delayed sampling will execute in bounded memory. The two conditions are dataflow properties of the core operations of delayed sampling: the m -consumed property and the unseparated paths property . A program executes in bounded memory under delayed sampling if, and only if, it satisfies the m -consumed and unseparated paths properties. We propose a static analysis that abstracts over these properties to soundly ensure that any program that passes the analysis satisfies these properties, and thus executes in bounded memory under delayed sampling.

Deep learning is moving towards increasingly sophisticated optimization objectives that employ higher-order functions, such as integration, continuous optimization, and root-finding. Since differentiable programming frameworks such as PyTorch and TensorFlow do not have first-class representations of these functions, developers must reason about the semantics of such objectives and manually translate them to differentiable code. We present a differentiable programming language, λ S , that is the first to deliver a semantics for higher-order functions, higher-order derivatives, and Lipschitz but nondifferentiable functions. Together, these features enableλ S to expose differentiable, higher-order functions for integration, optimization, and root-finding as first-class functions with automatically computed derivatives. λ S ’s semantics is computable, meaning that values can be computed to arbitrary precision, and we implement λ S as an embedded language in Haskell. We use λ S to construct novel differentiable libraries for representing probability distributions, implicit surfaces, and generalized parametric surfaces – all as instances of higher-order datatypes – and present case studies that rely on computing the derivatives of these higher-order functions and datatypes. In addition to modeling existing differentiable algorithms, such as a differentiable ray tracer for implicit surfaces, without requiring any user-level differentiation code, we demonstrate new differentiable algorithms, such as the Hausdorff distance of generalized parametric surfaces.

A Reduction – an accumulation over a set of values, using an associative and commutative operator – is a common computation in many numerical computations, including scientific computations, machine learning, computer vision, and financial analytics. Contemporary polyhedral-based compilation techniques make it possible to optimize reductions, such as prefix sums, in which each component of the reduction’s output potentially shares computation with another component in the reduction. Therefore an optimizing compiler can identify the computation shared between multiple components and generate code that computes the shared computation only once. These techniques, however, do not support reductions that – when phrased in the language of the polyhedral model – span multiple dependent statements. In such cases, existing approaches can generate incorrect code that violates the data dependences of the original, unoptimized program. In this work, we identify and formalize the optimization of dependent reductions as an integer bilinear program. We present a heuristic optimization algorithm that uses an affine sequential schedule of the program to determine how to simplfy reductions yet still preserve the program’s dependences. We demonstrate that the algorithm provides optimal complexity for a set of benchmark programs from the literature on probabilistic inference algorithms, whose performance critically relies on simplifying these reductions. The complexities for 10 of the 11 programs improve siginifcantly by factors at least of the sizes of the input data, which are in the range of 10 4 to 10 6 for typical real application inputs. We also confirm the significance of the improvement by showing speedups in wall-clock time that range from 1.1x to over 10 6 x.

We study whether a neural network optimizes to the same, linearly connected minimum under different samples of SGD noise (e.g., random data order and augmentation). We find that standard vision models become stable to SGD noise in this way early in training. From then on, the outcome of optimization is determined to a linearly connected region. We use this technique to study iterative magnitude pruning (IMP), the procedure used by work on the lottery ticket hypothesis to identify subnetworks that could have trained in isolation to full accuracy. We find that these subnetworks only reach full accuracy when they are stable to SGD noise, which either occurs at initialization for small-scale settings (MNIST) or early in training for large-scale settings (ResNet-50 and Inception-v3 on ImageNet).

CPU simulators are useful tools for modeling CPU execution behavior. However, they suffer from inaccuracies due to the cost and complexity of setting their fine-grained parameters, such as the latencies of individual instructions. This complexity arises from the expertise required to design benchmarks and measurement frameworks that can precisely measure the values of parameters at such fine granularity. In some cases, these parameters do not necessarily have a physical realization and are therefore fundamentally approximate, or even unmeasurable. In this paper we present DiffTune, a system for learning the parameters of x86 basic block CPU simulators from coarse-grained end-to-end measurements. Given a simulator, DiffTune learns its parameters by first replacing the original simulator with a differentiable surrogate, another function that approximates the original function; by making the surrogate differentiable, DiffTune is then able to apply gradient-based optimization techniques even when the original function is non-differentiable, such as is the case with CPU simulators. With this differentiable surrogate, DiffTune then applies gradient-based optimization to produce values of the simulator's parameters that minimize the simulator's error on a dataset of ground truth end-to-end performance measurements. Finally, the learned parameters are plugged back into the original simulator. DiffTune is able to automatically learn the entire set of microarchitecture-specific parameters within the Intel x86 simulation model of llvm-mca, a basic block CPU simulator based on LLVM's instruction scheduling model. DiffTune's learned parameters lead llvm-mca to an average error that not only matches but lowers that of its original, expert-provided parameter values.

In natural language processing (NLP), enormous pre-trained models like BERT have become the standard starting point for training on a range of downstream tasks, and similar trends are emerging in other areas of deep learning. In parallel, work on the lottery ticket hypothesis has shown that models for NLP and computer vision contain smaller matching subnetworks capable of training in isolation to full accuracy and transferring to other tasks. In this work, we combine these observations to assess whether such trainable, transferrable subnetworks exist in pre-trained BERT models. For a range of downstream tasks, we indeed find matching subnetworks at 40% to 90% sparsity. We find these subnetworks at (pre-trained) initialization, a deviation from prior NLP research where they emerge only after some amount of training. Subnetworks found on the masked language modeling task (the same task used to pre-train the model) transfer universally; those found on other tasks transfer in a limited fashion if at all. As large-scale pre-training becomes an increasingly central paradigm in deep learning, our results demonstrate that the main lottery ticket observations remain relevant in this context. Codes available at https://github.com/VITA-Group/BERT-Tickets.

Synchronous modeling is at the heart of programming languages like Lustre, Esterel, or Scade used routinely for implementing safety critical control software, e.g., fly-by-wire and engine control in planes. However, to date these languages have had limited modern support for modeling uncertainty --- probabilistic aspects of the software's environment or behavior --- even though modeling uncertainty is a primary activity when designing a control system.

Computer programs are increasingly being deployed in partially-observable environments. A partially observable environment is an environment whose state is not completely visible to the program, but from which the program receives partial observations. Developers typically deal with partial observability by writing a state estimator that, given observations, attempts to deduce the hidden state of the environment. In safety-critical domains, to formally verify safety properties developers may write an environment model. The model captures the relationship between observations and hidden states and is used to prove the software correct. In this paper, we present a new methodology for writing and verifying programs in partially observable environments. We present belief programming , a programming methodology where developers write an environment model that the program runtime automatically uses to perform state estimation. A belief program dynamically updates and queries a belief state that captures the possible states the environment could be in. To enable verification, we present Epistemic Hoare Logic that reasons about the possible belief states of a belief program the same way that classical Hoare logic reasons about the possible states of a program. We develop these concepts by defining a semantics and a program logic for a simple core language called BLIMP. In a case study, we show how belief programming could be used to write and verify a controller for the Mars Polar Lander in BLIMP. We present an implementation of BLIMP called CBLIMP and evaluate it to determine the feasibility of belief programming.

Modern probabilistic programming languages aim to formalize and automate key aspects of probabilistic modeling and inference. Many languages provide constructs for programmable inference that enable developers to improve inference speed and accuracy by tailoring an algorithm for use with a particular model or dataset. Unfortunately, it is easy to use these constructs to write unsound programs that appear to run correctly but produce incorrect results. To address this problem, we present a denotational semantics for programmable inference in higher-order probabilistic programming languages, along with a type system that ensures that well-typed inference programs are sound by construction. A central insight is that the type of a probabilistic expression can track the space of its possible execution traces, not just the type of value that it returns, as these traces are often the objects that inference algorithms manipulate. We use our semantics and type system to establish soundness properties of custom inference programs that use constructs for variational, sequential Monte Carlo, importance sampling, and Markov chain Monte Carlo inference.

Predicting the number of clock cycles a processor takes to execute a block of assembly instructions in steady state (the throughput) is important for both compiler designers and performance engineers. Building an analytical model to do so is especially complicated in modern x86-64 Complex Instruction Set Computer (CISC) machines with sophisticated processor microarchitectures in that it is tedious, error prone, and must be performed from scratch for each processor generation. In this paper we present Ithemal, the first tool which learns to predict the throughput of a set of instructions. Ithemal uses a hierarchical LSTM–based approach to predict throughput based on the opcodes and operands of instructions in a basic block. We show that Ithemal is more accurate than state-of-the-art hand-written tools currently used in compiler backends and static machine code analyzers. In particular, our model has less than half the error of state-of-the-art analytical models (LLVM’s llvm-mca and Intel’s IACA). Ithemal is also able to predict these throughput values just as fast as the aforementioned tools, and is easily ported across a variety of processor microarchitectures with minimal developer effort.

Modern microprocessors are equipped with single instruction multiple data (SIMD) or vector instruction sets which allow compilers to exploit fine-grained data level parallelism. To exploit this parallelism, compilers employ auto-vectorization techniques to automatically convert scalar code into vector code. Larsen & Amarasinghe (2000) first introduced superword level parallelism (SLP) based vectorization, which is one form of vectorization popularly used by compilers. Current compilers employ hand-crafted heuristics and typically only follow one SLP vectorization strategy which can be suboptimal. Recently, Mendis & Amarasinghe (2018) formulated the instruction packing problem of SLP vectorization by leveraging an integer linear programming (ILP) solver, achieving superior runtime performance. In this work, we explore whether it is feasible to imitate optimal decisions made by their ILP solution by fitting a graph neural network policy. We show that the learnt policy produces a vectorization scheme which is better than industry standard compiler heuristics both in terms of static measures and runtime performance. More specifically, the learnt agent produces a vectorization scheme which has a 22.6% higher average reduction in cost compared to LLVM compiler when measured using its own cost model and achieves a geometric mean runtime speedup of 1.015× on the NAS benchmark suite when compared to LLVM’s SLP vectorizer.

Algorithms for solid modeling, i.e., Computer-Aided Design (CAD) and computer graphics, are often specified on real numbers and then implemented with finite-precision arithmetic, such as floating-point. The result is that these implementations do not soundly compute the results that are expected from their specifications. We present a new library, StoneWorks, that provides sound and robust solid modeling primitives. We implement StoneWorks in MarshallB, a pure functional programming language for exact real arithmetic in which types denote topological spaces and functions denote continuous maps, ensuring that all programs are sound and robust. We developed MarshallB as an extension of the Marshall language. We also define a new shape representation, compact representation ( K-rep ), that enables constructions such as Minkowski sum and analyses such as Hausdorff distance that are not possible with traditional representations. K-rep is a nondeterminism monad for describing all the points in a shape. With our library, language, and representation together, we show that short StoneWorks programs can specify and execute sound and robust solid modeling algorithms and tasks.

Though many safety-critical software systems use floating point to represent real-world input and output, the mathematical specifications of these systems' behaviors use real numbers. Significant deviations from those specifications can cause errors and jeopardize safety. To ensure system safety, some programming systems offer exact real arithmetic, which often enables a program's computation to match its mathematical specification exactly. However, exact real arithmetic complicates decision-making: in these systems, it is impossible to compute (total and deterministic) discrete decisions based on connected spaces such as R. We present programming-language semantics based on constructive topology with variants allowing nondeterminism and/or partiality. Either nondeterminism or partiality suffices to allow computable decision making on connected spaces such as R. We then introduce pattern matching on spaces, a language construct for creating programs on spaces, generalizing pattern matching in functional programming, where patterns need not represent decidable predicates and also may overlap or be inexhaustive, giving rise to nondeterminism or partiality, respectively. Nondeterminism and/or partiality also yield formal logics for constructing approximate decision procedures. We extended the Marshall language for exact real arithmetic with these constructs and implemented some programs with it.

Researchers have recently designed a number of application-specific fault tolerance mechanisms that enable applications to either be naturally resilient to errors or include additional detection and correction steps that can bring the overall execution of an application back into an envelope for which an acceptable execution is eventually guaranteed. A major challenge to building an application that leverages these mechanisms, however, is to verify that the implementation satisfies the basic invariants that these mechanisms require---given a model of how faults may manifest during the application's execution. To this end we present Leto, an SMT-based automatic verification system that enables developers to verify their applications with respect to an execution model specification. Namely, Leto enables software and platform developers to programmatically specify the execution semantics of the underlying hardware system as well as verify assertions about the behavior of the application's resulting execution. In this paper, we present the Leto programming language and its corresponding verification system. We also demonstrate Leto on several applications that leverage application-specific fault tolerance

Convolutional Neural Networks (CNN) are a class of Ar- tificial Neural Networks (ANN) that are highly efficient at the pattern recognition tasks that underlie difficult AI prob- lems in a variety of domains, such as speech recognition, object recognition, and natural language processing. CNNs are, however, computationally intensive to train. This paper presents the first characterization of the per- formance optimization opportunities for training CNNs on CPUs. Our characterization includes insights based on the structure of the network itself (i.e., intrinsic arithmetic inten- sity of the convolution and its scalability under parallelism) as well as dynamic properties of its execution (i.e., sparsity of the computation).

The accuracy of an approximate computation is the distance between the result that the computation produces and the corresponding fully accurate result. The reliability of the computation is the probability that it will produce an acceptably accurate result. Emerging approximate hardware platforms provide approximate operations that, in return for reduced energy consumption and/or increased performance, exhibit reduced reliability and/or accuracy.

Emerging high-performance architectures are anticipated to contain unreliable components that may exhibit soft errors, which silently corrupt the results of computations. Full detection and masking of soft errors is challenging, expensive, and, for some applications, unnecessary. For example, approximate computing applications (such as multimedia processing, machine learning, and big data analytics) can often naturally tolerate soft errors.

Approximate computations (for example, video, audio, and image processing, machine learning, and many scientific computations) have the freedom to generate a range of acceptable results. Approximate program transformations (for example, task skipping and loop perforation) exploit this freedom to produce computations that can execute at a variety of points in an underlying accuracy versus performance trade-off space. One potential concern is that these transformations may change the semantics of the program and therefore cause the program to crash, perform an illegal operation, or otherwise violate its integrity.

We present a novel technique, automatic input rectification, and a prototype implementation, SOAP. SOAP learns a set of constraints characterizing typical inputs that an application is highly likely to process correctly. When given an atypical input that does not satisfy these constraints, SOAP automatically rectifies the input (i.e., changes the input so that it satisfies the learned constraints). The goal is to automatically convert potentially dangerous inputs into typical inputs that the program is highly likely to process correctly. Our experimental results show that, for a set of benchmark applications (Google Picasa, ImageMagick, VLC, Swfdec, and Dillo), this approach effectively converts malicious inputs (which successfully exploit vulnerabilities in the application) into benign inputs that the application processes correctly. Moreover, a manual code analysis shows that, if an input does satisfy the learned constraints, it is incapable of exploiting these vulnerabilities. We also present the results of a user study designed to evaluate the subjective perceptual quality of outputs from benign but atypical inputs that have been automatically rectified by SOAP to conform to the learned constraints. Specifically, we obtained benign inputs that violate learned constraints, used our input rectifier to obtain rectified inputs, then paid Amazon Mechanical Turk users to provide their subjective qualitative perception of the difference between the outputs from the original and rectified inputs. The results indicate that rectification can often preserve much, and in many cases all, of the desirable data in the original input.

We present Bolt, a novel system for escaping from infinite and long-running loops. Directed by a user, Bolt can attach to a running process and determine if the program is executing an infinite loop. If so, Bolt can deploy multiple strategies to escape the loop, restore the responsiveness of the program, and enable the program to deliver useful output.

Approximate program transformations such as skipping tasks [29, 30], loop perforation [21, 22, 35], reduction sampling [38], multiple selectable implementations [3, 4, 16, 38], dynamic knobs [16], synchronization elimination [20, 32], approximate function memoization [11],and approximate data types [34] produce programs that can execute at a variety of points in an underlying performance versus accuracy tradeoff space. These transformed programs have the ability to trade accuracy of their results for increased performance by dynamically and nondeterministically modifying variables that control their execution.

We present PowerDial, a system for dynamically adapting application behavior to execute successfully in the face of load and power fluctuations. PowerDial transforms static configuration parameters into dynamic knobs that the PowerDial control system can manipulate to dynamically trade off the accuracy of the computation in return for reductions in the computational resources that the application requires to produce its results. These reductions translate directly into performance improvements and power savings.

Applications that process complex inputs often react in different ways to changes in different regions of the input. Small changes to forgiving regions induce correspondingly small changes in the behavior and output. Small changes to critical regions, on the other hand, can induce disproportionally large changes in the behavior or output. Identifying the critical and forgiving regions in the input and the corresponding critical and forgiving regions of code is directly relevant to many software engineering tasks.

We present ClearView, a system for automatically patching errors in deployed software. ClearView works on stripped Windows x86 binaries without any need for source code, debugging information, or other external information, and without human intervention.

While parallel programmers find it easier to reason about large atomic regions, the conventional mutual exclusion-based primitives for synchronization force them to interleave many small operations to achieve performance. Transactional memory promises that programmer scan use large atomic regions while achieving similar performance. However, these large transactions can conflict when operating on shared data structures, even for logically independent operations. Transactional collection classes address this problem by allowing long-running transactions to operate on shared data while eliminating unnecessary conflicts. Transactional collection classes wrap existing data structures, without the need for custom implementations or knowledge of data structure internals.

Runtime reflection facilities, as present in Java and.NET, are powerful mechanisms for inspecting existing code and metadata, as well as generating new code and metadata on the fly. Such power does come at a high price though. The runtime reflection support in Java and.NET imposes a cost on all programs, whether they use reflection or not, simply by the necessity of keeping all metadata around and the inability to optimize code because of future possible code changes. A second—often overlooked—cost is the difficulty of writing correct reflection code to inspect or emit new metadata and code and the risk that the emitted code is not well-formed. In this paper we examine a subclass of problems that can be addressed using a simpler mechanism than runtime reflection, which we call compile-time reflection. We argue for a high-level construct called a transform that allows programmers to write inspection and generation code in a pattern matching and template style, avoiding at the same time the complexities of reflection APIs and providing the benefits of staged compilation in that the generated code and metadata is known to be well-formed and type safe ahead of time.

Program analysis has been increasingly used in software engineering tasks such as auditing programs for security vulnerabilities and finding errors in general. Such tools often require analyses much more sophisticated than those traditionally used in compiler optimizations. In particular, context-sensitive pointer alias information is a prerequisite for any sound and precise analysis that reasons about uses of heap objects in a program. Context-sensitive analysis is challenging because there are over 1014 contexts in a typical large program, even after recursive cycles are collapsed. Moreover, pointers cannot be resolved in general without analyzing the entire program.