Michael Carbin

Cambridge Yang, Michael Littman, Michael Carbin · cambridge

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.

Tian Jin, Michael Carbin, Daniel M. Roy et al. · utoronto

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.

Yi Ding, Alex Renda, Ahsan Pervaiz et al. · mit

Sample-efficient machine learning (SEML) has been widely applied to find optimal latency and power tradeoffs for configurable computer systems. Instead of randomly sampling from the configuration space, SEML reduces the search cost by dramatically reducing the number of configurations that must be sampled to optimize system goals (e.g., low latency or energy). Nevertheless, SEML only reduces one component of cost -- the total number of samples collected -- but does not decrease the cost of collecting each sample. Critically, not all samples are equal; some take much longer to collect because they correspond to slow system configurations. This paper present Cello, a computer systems optimization framework that reduces sample collection costs -- especially those that come from the slowest configurations. The key insight is to predict ahead of time whether samples will have poor system behavior (e.g., long latency or high energy) and terminate these samples early before their measured system behavior surpasses the termination threshold, which we call it predictive early termination. To predict the future system behavior accurately before it manifests as high runtime or energy, Cello uses censored regression to produces accurate predictions for running samples. We evaluate Cello by optimizing latency and energy for Apache Spark workloads. We give Cello a fixed amount of time to search a combined space of hardware and software configuration parameters. Our evaluation shows that compared to the state-of-the-art SEML approach in computer systems optimization, Cello improves latency by 1.19X for minimizing latency under a power constraint, and improves energy by 1.18X for minimizing energy under a latency constraint.

Alex Renda, Yi Ding, Michael Carbin · mit

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.

Yi Ding, Aijia Gao, Thibaud Ryden et al. · mit

Datacenter operators ensure fair and regular server maintenance by using automated processes to schedule maintenance jobs to complete within a strict time budget. Automating this scheduling problem is challenging because maintenance job duration varies based on both job type and hardware. While it is tempting to use prior machine learning techniques for predicting job duration, we find that the structure of the maintenance job scheduling problem creates a unique challenge. In particular, we show that prior machine learning methods that produce the lowest error predictions do not produce the best scheduling outcomes due to asymmetric costs. Specifically, underpredicting maintenance job duration has results in more servers being taken offline and longer server downtime than overpredicting maintenance job duration. The system cost of underprediction is much larger than that of overprediction. We present Acela, a machine learning system for predicting maintenance job duration, which uses quantile regression to bias duration predictions toward overprediction. We integrate Acela into a maintenance job scheduler and evaluate it on datasets from large-scale, production datacenters. Compared to machine learning based predictors from prior work, Acela reduces the number of servers that are taken offline by 1.87-4.28X, and reduces the server offline time by 1.40-2.80X.

Tian Jin, Nolan Clement, Xin Dong et al.

How does scaling the number of parameters in large language models (LLMs) affect their core capabilities? We study two natural scaling techniques -- weight pruning and simply training a smaller or larger model, which we refer to as dense scaling -- and their effects on two core capabilities of LLMs: (a) recalling facts presented during pre-training and (b) processing information presented in-context during inference. By curating a suite of tasks that help disentangle these two capabilities, we find a striking difference in how these two abilities evolve due to scaling. Reducing the model size by more than 30\% (via either scaling approach) significantly decreases the ability to recall facts seen in pre-training. Yet, a 60--70\% reduction largely preserves the various ways the model can process in-context information, ranging from retrieving answers from a long context to learning parameterized functions from in-context exemplars. The fact that both dense scaling and weight pruning exhibit this behavior suggests that scaling model size has an inherently disparate effect on fact recall and in-context learning.

Ellie Y. Cheng, Eric Atkinson, Guillaume Baudart et al.

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 1.76x on average and up to 206x to reach a target accuracy, compared to the inference plans implemented by default heuristics; the results also show that inference plans improve accuracy by 1.83x on average and up to 595x 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.

Hyunji Kim, Ahsan Pervaiz, Henry Hoffmann et al.

Modern computer systems need to execute under strict safety constraints (e.g., a power limit), but doing so often conflicts with their ability to deliver high performance (i.e. minimal latency). Prior work uses machine learning to automatically tune hardware resources such that the system execution meets safety constraints optimally. Such solutions monitor past system executions to learn the system's behavior under different hardware resource allocations before dynamically tuning resources to optimize the application execution. However, system behavior can change significantly between different applications and even different inputs of the same applications. Hence, the models learned using data collected a priori are often suboptimal and violate safety constraints when used with new applications and inputs. To address this limitation, we introduce the concept of an execution space, which is the cross product of hardware resources, input features, and applications. To dynamically and safely allocate hardware resources from the execution space, we present SCOPE, a resource manager that leverages a novel safe exploration framework. We evaluate SCOPE's ability to deliver improved latency while minimizing power constraint violations by dynamically configuring hardware while running a variety of Apache Spark applications. Compared to prior approaches that minimize power constraint violations, SCOPE consumes comparable power while improving latency by up to 9.5X. Compared to prior approaches that minimize latency, SCOPE achieves similar latency but reduces power constraint violation rates by up to 45.88X, achieving almost zero safety constraint violations across all applications.

Elliot Bolton, Abhinav Venigalla, Michihiro Yasunaga et al. · stanford

Models such as GPT-4 and Med-PaLM 2 have demonstrated impressive performance on a wide variety of biomedical NLP tasks. However, these models have hundreds of billions of parameters, are computationally expensive to run, require users to send their input data over the internet, and are trained on unknown data sources. Can smaller, more targeted models compete? To address this question, we build and release BioMedLM, a 2.7 billion parameter GPT-style autoregressive model trained exclusively on PubMed abstracts and full articles. When fine-tuned, BioMedLM can produce strong multiple-choice biomedical question-answering results competitive with much larger models, such as achieving a score of 57.3% on MedMCQA (dev) and 69.0% on the MMLU Medical Genetics exam. BioMedLM can also be fine-tuned to produce useful answers to patient questions on medical topics. This demonstrates that smaller models can potentially serve as transparent, privacy-preserving, economical and environmentally friendly foundations for particular NLP applications, such as in biomedicine. The model is available on the Hugging Face Hub: https://huggingface.co/stanford-crfm/BioMedLM.

Quinn Leng, Jacob Portes, Sam Havens et al.

Retrieval Augmented Generation (RAG) has emerged as a crucial technique for enhancing the accuracy of Large Language Models (LLMs) by incorporating external information. With the advent of LLMs that support increasingly longer context lengths, there is a growing interest in understanding how these models perform in RAG scenarios. Can these new long context models improve RAG performance? This paper presents a comprehensive study of the impact of increased context length on RAG performance across 20 popular open source and commercial LLMs. We ran RAG workflows while varying the total context length from 2,000 to 128,000 tokens (and 2 million tokens when possible) on three domain-specific datasets, and report key insights on the benefits and limitations of long context in RAG applications. Our findings reveal that while retrieving more documents can improve performance, only a handful of the most recent state of the art LLMs can maintain consistent accuracy at long context above 64k tokens. We also identify distinct failure modes in long context scenarios, suggesting areas for future research.

Logan Weber, Jesse Michel, Alex Renda et al.

A $\textit{neural surrogate of a program}$ is a neural network that mimics the behavior of a program. Researchers have used these neural surrogates to automatically tune program inputs, adapt programs to new settings, and accelerate computations. Researchers traditionally develop neural surrogates by training on input-output examples from a single program. Alternatively, language models trained on a large dataset including many programs can consume program text, to act as a neural surrogate. Using a language model to both generate a surrogate and act as a surrogate, however, leading to a trade-off between resource consumption and accuracy. We present $\textit{neural surrogate compilation}$, 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 that they produce neural surrogates that are $1.9$-$9.5\times$ as data-efficient, produce visual results that are $1.0$-$1.3\times$ more similar to ground truth, and train in $4.3$-$7.3\times$ fewer epochs than neural surrogates trained from scratch.

Tianlong Chen, Jonathan Frankle, Shiyu Chang et al.

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 and MoCo. 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. 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). 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.

Tianlong Chen, Jonathan Frankle, Shiyu Chang et al.

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.

Ellie Y. Cheng, Logan Weber, Tian Jin et al.

The rise of large language models (LLMs) has introduced a new type of programming: natural language programming. By writing prompts that direct LLMs to perform natural language processing, code generation, reasoning, etc., users are writing code in natural language -- natural language code -- for the LLM to execute. An emerging area of research enables interoperability between natural language code and formal languages such as Python. We present a novel programming abstraction, shared program state, that removes the manual work required to enable interoperability between natural language code and program state. With shared program state, programmers can write natural code that directly writes program variables, computes with program objects, and implements control flow in the program. We present a schema for specifying natural function interfaces that extend programming systems to support natural code and leverage this schema to specify shared program state as a natural function interface. We implement shared program state in the Nightjar programming system. Nightjar enables programmers to write Python programs that contain natural code that shares the Python program state. We show that Nightjar programs achieve comparable or higher task accuracy than manually written implementations (+4-19%), while decreasing the lines of code by 39.6% on average. The tradeoff to using Nightjar is that it may incur runtime overhead (0.4-4.3x runtime of manual implementations).

Cambridge Yang, Michael Littman, Michael Carbin · cambridge

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.

Tian Jin, Ellie Y. Cheng, Zack Ankner et al.

Decoding with autoregressive large language models (LLMs) traditionally occurs sequentially, generating one token after another. An emerging line of work explored parallel decoding by identifying and simultaneously generating semantically independent chunks of LLM responses. However, these techniques rely on hand-crafted heuristics tied to syntactic structures like lists and paragraphs, making them rigid and imprecise. We present PASTA, a learning-based system that teaches LLMs to identify semantic independence and express parallel decoding opportunities in their own responses. At its core are PASTA-LANG and its interpreter: PASTA-LANG is an annotation language that enables LLMs to express semantic independence in their own responses; the language interpreter acts on these annotations to orchestrate parallel decoding on-the-fly at inference time. Through a two-stage finetuning process, we train LLMs to generate PASTA-LANG annotations that optimize both response quality and decoding speed. Evaluation on AlpacaEval, an instruction following benchmark, shows that our approach Pareto-dominates existing methods in terms of decoding speed and response quality; our results demonstrate geometric mean speedups ranging from 1.21x to 1.93x with corresponding quality changes of +2.2% to -7.1%, measured by length-controlled win rates against sequential decoding baseline.

Mathew Jacob, Erik Lindgren, Matei Zaharia et al.

Rerankers, typically cross-encoders, are computationally intensive but are frequently used because they are widely assumed to outperform cheaper initial IR systems. We challenge this assumption by measuring reranker performance for full retrieval, not just re-scoring first-stage retrieval. To provide a more robust evaluation, we prioritize strong first-stage retrieval using modern dense embeddings and test rerankers on a variety of carefully chosen, challenging tasks, including internally curated datasets to avoid contamination, and out-of-domain ones. Our empirical results reveal a surprising trend: the best existing rerankers provide initial improvements when scoring progressively more documents, but their effectiveness gradually declines and can even degrade quality beyond a certain limit. We hope that our findings will spur future research to improve reranking.

Nandan Thakur, Jimmy Lin, Sam Havens et al.

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 topics 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 topics, 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 topics) 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.

Jacob Portes, Connor Jennings, Erica Ji Yuen et al.

How does retrieval performance scale with pretraining FLOPs? We benchmark retrieval performance across LLM model sizes from 125 million parameters to 7 billion parameters pretrained on datasets ranging from 1 billion tokens to more than 2 trillion tokens. We find that retrieval performance on zero-shot BEIR tasks predictably scales with LLM size, training duration, and estimated FLOPs. We also show that In-Context Learning scores are strongly correlated with retrieval scores across retrieval tasks. Finally, we highlight the implications this has for the development of LLM-based retrievers.

Daniel Israel, Tian Jin, Ellie Cheng et al.

A central challenge in large language model inference is the trade-off between generation speed and output quality. Autoregressive models produce high-quality text but generate tokens sequentially. Diffusion models can generate tokens in parallel but often need many iterations to match the same quality. We propose planned diffusion, a hybrid method that combines the strengths of both paradigms. Planned diffusion works in two stages: first, the model creates a short autoregressive plan that breaks the output into smaller, independent spans. Second, the model generates these spans simultaneously using diffusion. This approach expands the speed-quality Pareto frontier and provides a practical path to faster, high-quality text generation. On AlpacaEval, a suite of 805 instruction-following prompts, planned diffusion achieves Pareto-optimal trade-off between quality and latency, achieving 1.27x to 1.81x speedup over autoregressive generation with only 0.87\% to 5.4\% drop in win rate, respectively. Our sensitivity analysis shows that the planning mechanism of planned diffusion is minimal and reliable, and simple runtime knobs exist to provide flexible control of the quality-latency trade-off.

Alex Renda, Yi Ding, Michael Carbin

Surrogates, models that mimic the behavior of programs, form the basis of a variety of development workflows. We study three surrogate-based design patterns, evaluating each in case studies on a large-scale CPU simulator. With surrogate compilation, programmers develop a surrogate that mimics the behavior of a program to deploy to end-users in place of the original program. Surrogate compilation accelerates the CPU simulator under study by $1.6\times$. With surrogate adaptation, programmers develop a surrogate of a program then retrain that surrogate on a different task. Surrogate adaptation decreases the simulator's error by up to $50\%$. With surrogate optimization, programmers develop a surrogate of a program, optimize input parameters of the surrogate, then plug the optimized input parameters back into the original program. Surrogate optimization finds simulation parameters that decrease the simulator's error by $5\%$ compared to the error induced by expert-set parameters. In this paper we formalize this taxonomy of surrogate-based design patterns. We further describe the programming methodology common to all three design patterns. Our work builds a foundation for the emerging class of workflows based on programming with surrogates of programs.

Rajiv Movva, Jonathan Frankle, Michael Carbin

Magnitude pruning is a common, effective technique to identify sparse subnetworks at little cost to accuracy. In this work, we ask whether a particular architecture's accuracy-sparsity tradeoff can be improved by combining pruning information across multiple runs of training. From a shared ResNet-20 initialization, we train several network copies (\emph{siblings}) to completion using different SGD data orders on CIFAR-10. While the siblings' pruning masks are naively not much more similar than chance, starting sibling training after a few epochs of shared pretraining significantly increases pruning overlap. We then choose a subnetwork by either (1) taking all weights that survive pruning in any sibling (mask union), or (2) taking only the weights that survive pruning across all siblings (mask intersection). The resulting subnetwork is retrained. Strikingly, we find that union and intersection masks perform very similarly. Both methods match the accuracy-sparsity tradeoffs of the one-shot magnitude pruning baseline, even when we combine masks from up to $k = 10$ siblings.

Alex Renda, Yishen Chen, Charith Mendis et al.

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.

Jonathan Frankle, Gintare Karolina Dziugaite, Daniel M. Roy et al.

Recent work has explored the possibility of pruning neural networks at initialization. We assess proposals for doing so: SNIP (Lee et al., 2019), GraSP (Wang et al., 2020), SynFlow (Tanaka et al., 2020), and magnitude pruning. Although these methods surpass the trivial baseline of random pruning, they remain below the accuracy of magnitude pruning after training, and we endeavor to understand why. We show that, unlike pruning after training, randomly shuffling the weights these methods prune within each layer or sampling new initial values preserves or improves accuracy. As such, the per-weight pruning decisions made by these methods can be replaced by a per-layer choice of the fraction of weights to prune. This property suggests broader challenges with the underlying pruning heuristics, the desire to prune at initialization, or both.

Jonathan S. Rosenfeld, Jonathan Frankle, Michael Carbin et al.

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.

Riyadh Baghdadi, Abdelkader Nadir Debbagh, Kamel Abdous et al.

In this paper, we demonstrate a compiler that can optimize sparse and recurrent neural networks, both of which are currently outside of the scope of existing neural network compilers (sparse neural networks here stand for networks that can be accelerated with sparse tensor algebra techniques). Our demonstration includes a mapping of sparse and recurrent neural networks to the polyhedral model along with an implementation of our approach in TIRAMISU, our state-of-the-art polyhedral compiler. We evaluate our approach on a set of deep learning benchmarks and compare our results with hand-optimized industrial libraries. Our results show that our approach at least matches Intel MKL-DNN and in some cases outperforms it by 5x (on multicore-CPUs).

Alex Renda, Jonathan Frankle, Michael Carbin

Many neural network pruning algorithms proceed in three steps: train the network to completion, remove unwanted structure to compress the network, and retrain the remaining structure to recover lost accuracy. The standard retraining technique, fine-tuning, trains the unpruned weights from their final trained values using a small fixed learning rate. In this paper, we compare fine-tuning to alternative retraining techniques. Weight rewinding (as proposed by Frankle et al., (2019)), rewinds unpruned weights to their values from earlier in training and retrains them from there using the original training schedule. Learning rate rewinding (which we propose) trains the unpruned weights from their final values using the same learning rate schedule as weight rewinding. Both rewinding techniques outperform fine-tuning, forming the basis of a network-agnostic pruning algorithm that matches the accuracy and compression ratios of several more network-specific state-of-the-art techniques.

Jonathan Frankle, Gintare Karolina Dziugaite, Daniel M. Roy et al.

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).

Jonathan Frankle, Gintare Karolina Dziugaite, Daniel M. Roy et al.

Pruning is a well-established technique for removing unnecessary structure from neural networks after training to improve the performance of inference. Several recent results have explored the possibility of pruning at initialization time to provide similar benefits during training. In particular, the "lottery ticket hypothesis" conjectures that typical neural networks contain small subnetworks that can train to similar accuracy in a commensurate number of steps. The evidence for this claim is that a procedure based on iterative magnitude pruning (IMP) reliably finds such subnetworks retroactively on small vision tasks. However, IMP fails on deeper networks, and proposed methods to prune before training or train pruned networks encounter similar scaling limitations. In this paper, we argue that these efforts have struggled on deeper networks because they have focused on pruning precisely at initialization. We modify IMP to search for subnetworks that could have been obtained by pruning early in training (0.1% to 7% through) rather than at iteration 0. With this change, it finds small subnetworks of deeper networks (e.g., 80% sparsity on Resnet-50) that can complete the training process to match the accuracy of the original network on more challenging tasks (e.g., ImageNet). In situations where IMP fails at iteration 0, the accuracy benefits of delaying pruning accrue rapidly over the earliest iterations of training. To explain these behaviors, we study subnetwork "stability," finding that - as accuracy improves in this fashion - IMP subnetworks train to parameters closer to those of the full network and do so with improved consistency in the face of gradient noise. These results offer new insights into the opportunity to prune large-scale networks early in training and the behaviors underlying the lottery ticket hypothesis

Charith Mendis, Alex Renda, Saman Amarasinghe et al.

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.

Justin Gottschlich, Armando Solar-Lezama, Nesime Tatbul et al.

In this position paper, we describe our vision of the future of machine programming through a categorical examination of three pillars of research. Those pillars are: (i) intention, (ii) invention, and(iii) adaptation. Intention emphasizes advancements in the human-to-computer and computer-to-machine-learning interfaces. Invention emphasizes the creation or refinement of algorithms or core hardware and software building blocks through machine learning (ML). Adaptation emphasizes advances in the use of ML-based constructs to autonomously evolve software.

Jonathan Frankle, Michael Carbin

Neural network pruning techniques can reduce the parameter counts of trained networks by over 90%, decreasing storage requirements and improving computational performance of inference without compromising accuracy. However, contemporary experience is that the sparse architectures produced by pruning are difficult to train from the start, which would similarly improve training performance. We find that a standard pruning technique naturally uncovers subnetworks whose initializations made them capable of training effectively. Based on these results, we articulate the "lottery ticket hypothesis:" dense, randomly-initialized, feed-forward networks contain subnetworks ("winning tickets") that - when trained in isolation - reach test accuracy comparable to the original network in a similar number of iterations. The winning tickets we find have won the initialization lottery: their connections have initial weights that make training particularly effective. We present an algorithm to identify winning tickets and a series of experiments that support the lottery ticket hypothesis and the importance of these fortuitous initializations. We consistently find winning tickets that are less than 10-20% of the size of several fully-connected and convolutional feed-forward architectures for MNIST and CIFAR10. Above this size, the winning tickets that we find learn faster than the original network and reach higher test accuracy.

Vijay Ganesh, Michael Carbin, Martin Rinard

We propose a novel approach to improving software security called Cryptographic Path Hardening, which is aimed at hiding security vulnerabilities in software from attackers through the use of provably secure and obfuscated cryptographic devices to harden paths in programs. By "harden" we mean that certain error-checking if-conditionals in a given program P are replaced by equivalent" we mean that adversaries cannot use semi-automatic program analysis techniques to reason about the hardened program paths and thus cannot discover as-yet-unknown errors along those paths, except perhaps through black-box dictionary attacks or random testing (which we can never prevent). Other than these unpreventable attack methods, we can make program analysis aimed at error-finding "provably hard" for a resource-bounded attacker, in the same sense that cryptographic schemes are hard to break. Unlike security-through-obscurity, in Cryptographic Path Hardening we use provably-secure crypto devices to hide errors and our mathematical arguments of security are the same as the standard ones used in cryptography. One application of Cryptographic Path Hardening is that software patches or filters often reveal enough information to an attacker that they can be used to construct error-revealing inputs to exploit an unpatched version of the program. By "hardening" the patch we make it difficult for the attacker to analyze the patched program to construct error-revealing inputs, and thus prevent him from potentially constructing exploits.