Wentao Wu

Wentao Wu, Aleksei Timofeev, Chen Chen et al. · microsoft-research

We present MOFI, Manifold OF Images, a new vision foundation model designed to learn image representations from noisy entity annotated images. MOFI differs from previous work in two key aspects: (i) pre-training data, and (ii) training recipe. Regarding data, we introduce a new approach to automatically assign entity labels to images from noisy image-text pairs. Our approach involves employing a named entity recognition model to extract entities from the alt-text, and then using a CLIP model to select the correct entities as labels of the paired image. It's a simple, cost-effective method that can scale to handle billions of web-mined image-text pairs. Through this method, we have created Image-to-Entities (I2E), a new dataset with 1 billion images and 2 million distinct entities, covering rich visual concepts in the wild. Building upon the I2E dataset, we study different training recipes like supervised pre-training, contrastive pre-training, and multi-task learning. For contrastive pre-training, we treat entity names as free-form text, and further enrich them with entity descriptions. Experiments show that supervised pre-training with large-scale fine-grained entity labels is highly effective for image retrieval tasks, and multi-task training further improves the performance. The final MOFI model achieves 86.66% mAP on the challenging GPR1200 dataset, surpassing the previous state-of-the-art performance of 72.19% from OpenAI's CLIP model. Further experiments on zero-shot and linear probe image classification also show that MOFI outperforms a CLIP model trained on the original image-text data, demonstrating the effectiveness of the I2E dataset in learning strong image representations. We release our code and model weights at https://github.com/apple/ml-mofi.

Zhengfeng Lai, Haotian Zhang, Bowen Zhang et al.

Large-scale web-crawled datasets are fundamental for the success of pre-training vision-language models, such as CLIP. However, the inherent noise and potential irrelevance of web-crawled AltTexts pose challenges in achieving precise image-text alignment. Existing methods utilizing large language models (LLMs) for caption rewriting have shown promise on small, curated datasets like CC3M and CC12M. This study introduces a scalable pipeline for noisy caption rewriting. Unlike recent LLM rewriting techniques, we emphasize the incorporation of visual concepts into captions, termed as Visual-enriched Captions (VeCap). To ensure data diversity, we propose a novel mixed training scheme that optimizes the utilization of AltTexts alongside newly generated VeCap. We showcase the adaptation of this method for training CLIP on large-scale web-crawled datasets, termed VeCLIP. Employing this cost-effective pipeline, we effortlessly scale our dataset up to 300 million samples named VeCap dataset. Our results show significant advantages in image-text alignment and overall model performance. For example, VeCLIP achieves up to +25.2% gain in COCO and Flickr30k retrieval tasks under the 12M setting. For data efficiency, VeCLIP achieves +3% gain while only using 14% of the data employed in the vanilla CLIP and 11% in ALIGN. We also note the VeCap data is complementary with other well curated datasets good for zero-shot classification tasks. When combining VeCap and DFN, our model can achieve strong performance on both of image-text retrieval and zero-shot classification tasks, e.g. 83.1% accuracy@1 on ImageNet zero-shot for a H/14 model. We release the pre-trained models at https://github.com/apple/ml-veclip.

Lijie Xu, Shuang Qiu, Binhang Yuan et al. · microsoft-research

Stochastic gradient descent (SGD) is the cornerstone of modern machine learning (ML) systems. Despite its computational efficiency, SGD requires random data access that is inherently inefficient when implemented in systems that rely on block-addressable secondary storage such as HDD and SSD, e.g., TensorFlow/PyTorch and in-DB ML systems over large files. To address this impedance mismatch, various data shuffling strategies have been proposed to balance the convergence rate of SGD (which favors randomness) and its I/O performance (which favors sequential access). In this paper, we first conduct a systematic empirical study on existing data shuffling strategies, which reveals that all existing strategies have room for improvement -- they all suffer in terms of I/O performance or convergence rate. With this in mind, we propose a simple but novel hierarchical data shuffling strategy, CorgiPile. Compared with existing strategies, CorgiPile avoids a full data shuffle while maintaining comparable convergence rate of SGD as if a full shuffle were performed. We provide a non-trivial theoretical analysis of CorgiPile on its convergence behavior. We further integrate CorgiPile into PyTorch by designing new parallel/distributed shuffle operators inside a new CorgiPileDataSet API. We also integrate CorgiPile into PostgreSQL by introducing three new physical operators with optimizations. Our experimental results show that CorgiPile can achieve comparable convergence rate with the full shuffle based SGD for both deep learning and generalized linear models. For deep learning models on ImageNet dataset, CorgiPile is 1.5X faster than PyTorch with full data shuffle. For in-DB ML with linear models, CorgiPile is 1.6X-12.8X faster than two state-of-the-art in-DB ML systems, Apache MADlib and Bismarck, on both HDD and SSD.

Bojan Karlaš, David Dao, Matteo Interlandi et al. · microsoft-research

Developing modern machine learning (ML) applications is data-centric, of which one fundamental challenge is to understand the influence of data quality to ML training -- "Which training examples are 'guilty' in making the trained ML model predictions inaccurate or unfair?" Modeling data influence for ML training has attracted intensive interest over the last decade, and one popular framework is to compute the Shapley value of each training example with respect to utilities such as validation accuracy and fairness of the trained ML model. Unfortunately, despite recent intensive interest and research, existing methods only consider a single ML model "in isolation" and do not consider an end-to-end ML pipeline that consists of data transformations, feature extractors, and ML training. We present DataScope (ease.ml/datascope), the first system that efficiently computes Shapley values of training examples over an end-to-end ML pipeline, and illustrate its applications in data debugging for ML training. To this end, we first develop a novel algorithmic framework that computes Shapley value over a specific family of ML pipelines that we call canonical pipelines: a positive relational algebra query followed by a K-nearest-neighbor (KNN) classifier. We show that, for many subfamilies of canonical pipelines, computing Shapley value is in PTIME, contrasting the exponential complexity of computing Shapley value in general. We then put this to practice -- given an sklearn pipeline, we approximate it with a canonical pipeline to use as a proxy. We conduct extensive experiments illustrating different use cases and utilities. Our results show that DataScope is up to four orders of magnitude faster over state-of-the-art Monte Carlo-based methods, while being comparably, and often even more, effective in data debugging.

Xiao Wang, Wentao Wu, Chenglong Li et al.

Understanding vehicles in images is important for various applications such as intelligent transportation and self-driving system. Existing vehicle-centric works typically pre-train models on large-scale classification datasets and then fine-tune them for specific downstream tasks. However, they neglect the specific characteristics of vehicle perception in different tasks and might thus lead to sub-optimal performance. To address this issue, we propose a novel vehicle-centric pre-training framework called VehicleMAE, which incorporates the structural information including the spatial structure from vehicle profile information and the semantic structure from informative high-level natural language descriptions for effective masked vehicle appearance reconstruction. To be specific, we explicitly extract the sketch lines of vehicles as a form of the spatial structure to guide vehicle reconstruction. The more comprehensive knowledge distilled from the CLIP big model based on the similarity between the paired/unpaired vehicle image-text sample is further taken into consideration to help achieve a better understanding of vehicles. A large-scale dataset is built to pre-train our model, termed Autobot1M, which contains about 1M vehicle images and 12693 text information. Extensive experiments on four vehicle-based downstream tasks fully validated the effectiveness of our VehicleMAE. The source code and pre-trained models will be released at https://github.com/Event-AHU/VehicleMAE.

Yu Liu, Hantian Zhang, Luyuan Zeng et al.

We conduct an empirical study of machine learning functionalities provided by major cloud service providers, which we call machine learning clouds. Machine learning clouds hold the promise of hiding all the sophistication of running large-scale machine learning: Instead of specifying how to run a machine learning task, users only specify what machine learning task to run and the cloud figures out the rest. Raising the level of abstraction, however, rarely comes free - a performance penalty is possible. How good, then, are current machine learning clouds on real-world machine learning workloads? We study this question with a focus on binary classication problems. We present mlbench, a novel benchmark constructed by harvesting datasets from Kaggle competitions. We then compare the performance of the top winning code available from Kaggle with that of running machine learning clouds from both Azure and Amazon on mlbench. Our comparative study reveals the strength and weakness of existing machine learning clouds and points out potential future directions for improvement.

Lijie Xu, Chulin Xie, Yiran Guo et al. · microsoft-research

Current federated learning (FL) approaches view decentralized training data as a single table, divided among participants either horizontally (by rows) or vertically (by columns). However, these approaches are inadequate for handling distributed relational tables across databases. This scenario requires intricate SQL operations like joins and unions to obtain the training data, which is either costly or restricted by privacy concerns. This raises the question: can we directly run FL on distributed relational tables? In this paper, we formalize this problem as relational federated learning (RFL). We propose TablePuppet, a generic framework for RFL that decomposes the learning process into two steps: (1) learning over join (LoJ) followed by (2) learning over union (LoU). In a nutshell, LoJ pushes learning down onto the vertical tables being joined, and LoU further pushes learning down onto the horizontal partitions of each vertical table. TablePuppet incorporates computation/communication optimizations to deal with the duplicate tuples introduced by joins, as well as differential privacy (DP) to protect against both feature and label leakages. We demonstrate the efficiency of TablePuppet in combination with two widely-used ML training algorithms, stochastic gradient descent (SGD) and alternating direction method of multipliers (ADMM), and compare their computation/communication complexity. We evaluate the SGD/ADMM algorithms developed atop TablePuppet by training diverse ML models. Our experimental results show that TablePuppet achieves model accuracy comparable to the centralized baselines running directly atop the SQL results. Moreover, ADMM takes less communication time than SGD to converge to similar model accuracy.

Cedric Renggli, Luka Rimanic, Nezihe Merve Gürel et al.

Developing machine learning models can be seen as a process similar to the one established for traditional software development. A key difference between the two lies in the strong dependency between the quality of a machine learning model and the quality of the data used to train or perform evaluations. In this work, we demonstrate how different aspects of data quality propagate through various stages of machine learning development. By performing a joint analysis of the impact of well-known data quality dimensions and the downstream machine learning process, we show that different components of a typical MLOps pipeline can be efficiently designed, providing both a technical and theoretical perspective.

Cedric Renggli, Luka Rimanic, Luka Kolar et al.

In our experience of working with domain experts who are using today's AutoML systems, a common problem we encountered is what we call "unrealistic expectations" -- when users are facing a very challenging task with a noisy data acquisition process, while being expected to achieve startlingly high accuracy with machine learning (ML). Many of these are predestined to fail from the beginning. In traditional software engineering, this problem is addressed via a feasibility study, an indispensable step before developing any software system. In this paper, we present Snoopy, with the goal of supporting data scientists and machine learning engineers performing a systematic and theoretically founded feasibility study before building ML applications. We approach this problem by estimating the irreducible error of the underlying task, also known as the Bayes error rate (BER), which stems from data quality issues in datasets used to train or evaluate ML model artifacts. We design a practical Bayes error estimator that is compared against baseline feasibility study candidates on 6 datasets (with additional real and synthetic noise of different levels) in computer vision and natural language processing. Furthermore, by including our systematic feasibility study with additional signals into the iterative label cleaning process, we demonstrate in end-to-end experiments how users are able to save substantial labeling time and monetary efforts.

Bojan Karlaš, Peng Li, Renzhi Wu et al.

Machine learning (ML) applications have been thriving recently, largely attributed to the increasing availability of data. However, inconsistency and incomplete information are ubiquitous in real-world datasets, and their impact on ML applications remains elusive. In this paper, we present a formal study of this impact by extending the notion of Certain Answers for Codd tables, which has been explored by the database research community for decades, into the field of machine learning. Specifically, we focus on classification problems and propose the notion of "Certain Predictions" (CP) -- a test data example can be certainly predicted (CP'ed) if all possible classifiers trained on top of all possible worlds induced by the incompleteness of data would yield the same prediction. We study two fundamental CP queries: (Q1) checking query that determines whether a data example can be CP'ed; and (Q2) counting query that computes the number of classifiers that support a particular prediction (i.e., label). Given that general solutions to CP queries are, not surprisingly, hard without assumption over the type of classifier, we further present a case study in the context of nearest neighbor (NN) classifiers, where efficient solutions to CP queries can be developed -- we show that it is possible to answer both queries in linear or polynomial time over exponentially many possible worlds. We demonstrate one example use case of CP in the important application of "data cleaning for machine learning (DC for ML)." We show that our proposed CPClean approach built based on CP can often significantly outperform existing techniques in terms of classification accuracy with mild manual cleaning effort.

Fotis Psallidas, Yiwen Zhu, Bojan Karlas et al.

The recent success of machine learning (ML) has led to an explosive growth both in terms of new systems and algorithms built in industry and academia, and new applications built by an ever-growing community of data science (DS) practitioners. This quickly shifting panorama of technologies and applications is challenging for builders and practitioners alike to follow. In this paper, we set out to capture this panorama through a wide-angle lens, by performing the largest analysis of DS projects to date, focusing on questions that can help determine investments on either side. Specifically, we download and analyze: (a) over 6M Python notebooks publicly available on GITHUB, (b) over 2M enterprise DS pipelines developed within COMPANYX, and (c) the source code and metadata of over 900 releases from 12 important DS libraries. The analysis we perform ranges from coarse-grained statistical characterizations to analysis of library imports, pipelines, and comparative studies across datasets and time. We report a large number of measurements for our readers to interpret, and dare to draw a few (actionable, yet subjective) conclusions on (a) what systems builders should focus on to better serve practitioners, and (b) what technologies should practitioners bet on given current trends. We plan to automate this analysis and release associated tools and results periodically.

Frances Ann Hubis, Wentao Wu, Ce Zhang

Simplifying machine learning (ML) application development, including distributed computation, programming interface, resource management, model selection, etc, has attracted intensive interests recently. These research efforts have significantly improved the efficiency and the degree of automation of developing ML models. In this paper, we take a first step in an orthogonal direction towards automated quality management for human-in-the-loop ML application development. We build ease. ml/meter, a system that can automatically detect and measure the degree of overfitting during the whole lifecycle of ML application development. ease. ml/meter returns overfitting signals with strong probabilistic guarantees, based on which developers can take appropriate actions. In particular, ease. ml/meter provides principled guidelines to simple yet nontrivial questions regarding desired validation and test data sizes, which are among commonest questions raised by developers. The fact that ML application development is typically a continuous procedure further worsens the situation: The validation and test data sets can lose their statistical power quickly due to multiple accesses, especially in the presence of adaptive analysis. ease. ml/meter addresses these challenges by leveraging a collection of novel techniques and optimizations, resulting in practically tractable data sizes without compromising the probabilistic guarantees. We present the design and implementation details of ease. ml/meter, as well as detailed theoretical analysis and empirical evaluation of its effectiveness.

Cedric Renggli, Bojan Karlaš, Bolin Ding et al.

Continuous integration is an indispensable step of modern software engineering practices to systematically manage the life cycles of system development. Developing a machine learning model is no difference - it is an engineering process with a life cycle, including design, implementation, tuning, testing, and deployment. However, most, if not all, existing continuous integration engines do not support machine learning as first-class citizens. In this paper, we present ease.ml/ci, to our best knowledge, the first continuous integration system for machine learning. The challenge of building ease.ml/ci is to provide rigorous guarantees, e.g., single accuracy point error tolerance with 0.999 reliability, with a practical amount of labeling effort, e.g., 2K labels per test. We design a domain specific language that allows users to specify integration conditions with reliability constraints, and develop simple novel optimizations that can lower the number of labels required by up to two orders of magnitude for test conditions popularly used in real production systems.

Tian Li, Jie Zhong, Ji Liu et al.

We present ease.ml, a declarative machine learning service platform we built to support more than ten research groups outside the computer science departments at ETH Zurich for their machine learning needs. With ease.ml, a user defines the high-level schema of a machine learning application and submits the task via a Web interface. The system automatically deals with the rest, such as model selection and data movement. In this paper, we describe the ease.ml architecture and focus on a novel technical problem introduced by ease.ml regarding resource allocation. We ask, as a "service provider" that manages a shared cluster of machines among all our users running machine learning workloads, what is the resource allocation strategy that maximizes the global satisfaction of all our users? Resource allocation is a critical yet subtle issue in this multi-tenant scenario, as we have to balance between efficiency and fairness. We first formalize the problem that we call multi-tenant model selection, aiming for minimizing the total regret of all users running automatic model selection tasks. We then develop a novel algorithm that combines multi-armed bandits with Bayesian optimization and prove a regret bound under the multi-tenant setting. Finally, we report our evaluation of ease.ml on synthetic data and on one service we are providing to our users, namely, image classification with deep neural networks. Our experimental evaluation results show that our proposed solution can be up to 9.8x faster in achieving the same global quality for all users as the two popular heuristics used by our users before ease.ml.

Xi Wu, Matthew Fredrikson, Wentao Wu et al.

Private regression has received attention from both database and security communities. Recent work by Fredrikson et al. (USENIX Security 2014) analyzed the functional mechanism (Zhang et al. VLDB 2012) for training linear regression models over medical data. Unfortunately, they found that model accuracy is already unacceptable with differential privacy when $\varepsilon = 5$. We address this issue, presenting an explicit connection between differential privacy and stable learning theory through which a substantially better privacy/utility tradeoff can be obtained. Perhaps more importantly, our theory reveals that the most basic mechanism in differential privacy, output perturbation, can be used to obtain a better tradeoff for all convex-Lipschitz-bounded learning tasks. Since output perturbation is simple to implement, it means that our approach is potentially widely applicable in practice. We go on to apply it on the same medical data as used by Fredrikson et al. Encouragingly, we achieve accurate models even for $\varepsilon = 0.1$. In the last part of this paper, we study the impact of our improved differentially private mechanisms on model inversion attacks, a privacy attack introduced by Fredrikson et al. We observe that the improved tradeoff makes the resulting differentially private model more susceptible to inversion attacks. We analyze this phenomenon formally.