Papers

Preserving the performance of a trained model while removing unique characteristics of marked training data points is challenging. Recent research usually suggests retraining a model from scratch with remaining training data or refining the model by reverting the model optimization on the marked data points. Unfortunately, aside from their computational inefficiency, those approaches inevitably hurt the resulting model's generalization ability since they remove not only unique characteristics but also discard shared (and possibly contributive) information. To address the performance degradation problem, this paper presents a novel approach called Performance Unchanged Model Augmentation (PUMA). The proposed PUMA framework explicitly models the influence of each training data point on the model's generalization ability with respect to various performance criteria. It then complements the negative impact of removing marked data by reweighting the remaining data optimally. To demonstrate the effectiveness of the PUMA framework, we compared it with multiple state-of-the-art data removal techniques in the experiments, where we show the PUMA can effectively and efficiently remove the unique characteristics of marked training data without retraining the model that can 1) fool a membership attack, and 2) resist performance degradation. In addition, as PUMA estimates the data importance during its operation, we show it could serve to debug mislabelled data points more efficiently than existing approaches.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

Inspired by two basic mechanisms in animal visual systems, we introduce a feature transform technique that imposes invariance properties in the training of deep neural networks. The resulting algorithm requires less parameter tuning, trains well with an initial learning rate 1.0, and easily generalizes to different tasks. We enforce scale invariance with local statistics in the data to align similar samples at diverse scales. To accelerate convergence, we enforce a GL(n)-invariance property with global statistics extracted from a batch such that the gradient descent solution should remain invariant under basis change. Profiling analysis shows our proposed modifications takes 5% of the computations of the underlying convolution layer. Tested on convolutional networks and transformer networks, our proposed technique requires fewer iterations to train, surpasses all baselines by a large margin, seamlessly works on both small and large batch size training, and applies to different computer vision and language tasks.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

With ever-increasing dataset sizes, subset selection techniques are becoming increasingly important for a plethora of tasks. It is often necessary to guide the subset selection to achieve certain desiderata, which includes focusing or targeting certain data points, while avoiding others. Examples of such problems include: i)targeted learning, where the goal is to find subsets with rare classes or rare attributes on which the model is under performing, and ii)guided summarization, where data (e.g.,image collection, text, document or video) is summarized for quicker human consumption with specific additional user in-tent. Motivated by such applications, we present PRISM, a rich class of PaRameterIzed Submodular information Measures. Through novel functions and their parameterizations, PRISM offers a variety of modeling capabilities that enable a trade-off between desired qualities of a subset like diversity or representation and similarity/dissimilarity with a set of data points. We demonstrate how PRISM can be applied to the two real-world problems mentioned above, which require guided subset selection. In doing so, we show that PRISM interestingly generalizes some past work, therein reinforcing its broad utility. Through extensive experiments on diverse datasets, we demonstrate the superiority of PRISM over the state-of-the-art in targeted learning and in guided image-collection summarization.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

Training neural networks with discrete stochastic variables presents a unique challenge. Backpropagation is not directly applicable, nor are the reparameterization tricks used in networks with continuous stochastic variables. To address this challenge, we present Hindsight Network Credit Assignment (HNCA), a novel gradient estimation algorithm for networks of discrete stochastic units. HNCA works by assigning credit to each unit based on the degree to which its output influences its immediate children in the network. We prove that HNCA produces unbiased gradient estimates with reduced variance compared to the REINFORCE estimator, while the computational cost is similar to that of backpropagation. We first apply HNCA in a contextual bandit setting to optimize a reward function that is unknown to the agent. In this setting, we empirically demonstrate that HNCA significantly outperforms REINFORCE, indicating that the variance reduction implied by our theoretical analysis is significant and impactful. We then show how HNCA can be extended to optimize a more general function of the outputs of a network of stochastic units, where the function is known to the agent. We apply this extended version of HNCA to train a discrete variational auto-encoder and empirically show it compares favourably to other strong methods. We believe that the ideas underlying HNCA can help stimulate new ways of thinking about efficient credit assignment in stochastic compute graphs.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

With the abundance of available data, many enterprises seek to implement data-driven prescriptive analytics to help them make informed decisions. These prescriptive policies need to satisfy operational constraints, and proactively eliminate rule conflicts, both of which are ubiquitous in practice. It is also desirable for them to be simple and interpretable, so they can be easily verified and implemented. Existing approaches from the literature center around constructing variants of prescriptive decision trees to generate interpretable policies. However, none of the existing methods is able to handle constraints. In this paper, we propose a scalable method that solves the constrained prescriptive policy generation problem. We introduce a novel path-based mixed-integer program (MIP) formulation which identifies a (near) optimal policy efficiently via column generation. The policy generated can be represented as a multiway-split tree which is more interpretable and informative than binary-split trees due to its shorter rules. We demonstrate the efficacy of our method with extensive computational experiments on both synthetic and real datasets.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

It is an essential product requirement of Yahoo Mail to distinguish between personal and machine-generated emails. The old production classifier in Yahoo Mail was based on a simple logistic regression model. That model was trained by aggregating features at the SMTP address level. We propose building deep models at the message level. We built and trained four individual CNN models: (1) a content model with subject and content as input; (2) a sender model with sender email address and name as input; (3) an action model by analyzing email recipients’ action patterns and correspondingly generating target labels based on senders’ opening/deleting behaviors; (4) a salutation model by utilizing senders’ “explicit salutation”signal as positive labels. Next, we built a final full model after exploring different combinations of the above four models. Experimental results on editorial data show that our full model improves the adjusted recall from 70.5% to 78.8% compared to the old production model, while at the same time lifts the precision from 94.7% to 96.0%. Our full model also significantly beats the state-of-the-art Bert model at this task. This full model has been deployed into the current production system (Yahoo Mail 6).

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

We develop new adaptive algorithms for variational inequalities with monotone operators, which capture many problems of interest, notably convex optimization and convex-concave saddle point problems. Our algorithms automatically adapt to unknown problem parameters such as the smoothness and the norm of the operator, and the variance of the stochastic evaluation oracle. We show that our algorithms are universal and simultaneously achieve the optimal convergence rates in the non-smooth, smooth, and stochastic settings. The convergence guarantees of our algorithms improve over existing adaptive methods by a $\Omega(\sqrt{\ln T})$ factor, matching the optimal non-adaptive algorithms. Additionally, prior works require that the optimization domain is bounded. In this work, we remove this restriction and give algorithms for unbounded domains that are adaptive and universal. Our general proof techniques can be used for many variants of the algorithm using one or two operator evaluations per iteration. The classical methods based on the ExtraGradient/MirrorProx algorithm require two operator evaluations per iteration, which is the dominant factor in the running time in many settings.

poster session 5 @ blue 2, poster session 10 @ blue 2, oral session 10 @ blue 2, poster session 5, poster session 10, oral session 10

From an engineering perspective, a design should not only perform well in an ideal condition, but should also resist noises. Such a design methodology, namely robust design, has been widely implemented in the industry for product quality control. However, classic robust design requires a lot of evaluations for a single design target, while the results of these evaluations could not be reused for a new target. To achieve data-efficient robust design, we propose Robust Inverse Design under Noise (RID-Noise), which can utilize existing data to train a conditional invertible neural network. Specifically, we estimate the robustness of a design parameter by its predictability, measured by the prediction error of a forward neural network. We also define a sample-wise weight, which can be used in the maximum weighted likelihood estimation of an inverse model based on a conditional invertible neural network. With the visual results from experiments, we clearly justify how RID-Noise works by learning the distribution and robustness from data. Further experiments on several real-world benchmark tasks with noises confirm that our method is more effective than other state-of-the-art inverse design methods.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

The goal of Approximate Query Processing (AQP) techniques is to provide very fast but “accurate enough” results for costly aggregate queries thereby improving user experience in interactive exploration of large datasets. Recently proposed ML-based AQP techniques can provide very low latency as query execution only involves model inference as compared to traditional query processing on database clusters. However, with increase in the number of filtering predicates (WHERE clauses), the approximation error significantly increases for these methods. Analysts often use queries with a large number of predicates for insights discovery. Thus, maintaining low approximation error is important to prevent analysts from drawing misleading conclusions. In this paper, we propose ELECTRA, a predicate-aware AQP system that can answer analytics-style queries with a large number of predicates with much smaller approximation errors. ELECTRA uses a conditional generative model that learns the conditional distribution of the data and at run-time generates a small (≈1000) but representative sample, on which the query is executed to compute the approximate result. Our evaluations with four different baselines on three real-world datasets show that ELECTRA provides not only lower AQP error but also can efficiently run at the client-side with low latency.

poster session 5 @ blue 2, poster session 10 @ blue 2, oral session 10 @ blue 2, poster session 5, poster session 10, oral session 10

Parameter-free stochastic gradient descent (PFSGD) algorithms do not require setting learning rates while achieving optimal theoretical performance. In practical applications, however, there remains an empirical gap between tuned stochastic gradient descent (SGD) and PFSGD. In this paper, we close the empirical gap with a new parameter-free algorithm based on continuous-time Coin-Betting on truncated models. The new update is derived through the solution of an Ordinary Differential Equation (ODE) and solved in a closed form. We show empirically that this new parameter-free algorithm outperforms algorithms with the ``best default'' learning rates and almost matches the performance of finely tuned baselines without anything to tune.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

Various neural network models have been proposed to tackle combinatorial optimization problems such as the travelling salesman problem (TSP). Existing learning-based TSP methods adopt a simple setting that the training and testing data are independent and identically distributed. However, the existing literature fails to solve TSP instances when training and testing data have different distributions. Concretely, we find that different training and testing distribution will result in more difficult TSP instances, i.e., the solution obtained by the model has a large gap from the optimal solution. To tackle this problem, in this work, we study learning-based TSP methods when training and testing data have different distributions using adaptive-hardness, i.e., how difficult a TSP instance can be for a solver. This problem is challenging because it is non-trivial to (1) define hardness measurement quantitatively; (2) efficiently and continuously generate sufficiently hard TSP instances upon model training; (3) fully utilize instances with different levels of hardness to learn a more powerful TSP solver. To solve these challenges, we first propose a principled hardness measurement to quantify the hardness of TSP instances. Then, we propose a hardness-adaptive generator to generate instances with different hardness. We further propose a curriculum learner fully utilizing these instances to train the TSP solver. Experiments show that our hardness-adaptive generator can generate instances ten times harder than the existing methods, and our proposed method achieves significant improvement over state-of-the-art models in terms of the optimality gap.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

Hyperparameter optimization is a ubiquitous challenge in machine learning, and the performance of a trained model depends crucially upon their effective selection. While a rich set of tools exist for this purpose, there are currently no practical hyperparameter selection methods under the constraint of differential privacy (DP). We study honest hyperparameter selection for differentially private machine learning, in which the process of hyperparameter tuning is accounted for in the overall privacy budget. To this end, we i) show that standard composition tools outperform more advanced techniques in many settings, ii) empirically and theoretically demonstrate an intrinsic connection between the learning rate and clipping norm hyperparameters, iii) show that adaptive optimizers like DPAdam enjoy a significant advantage in the process of honest hyperparameter tuning, and iv) draw upon novel limiting behaviour of Adam in the DP setting to design a new and more efficient optimizer.

poster session 5 @ blue 2, poster session 10 @ blue 2, oral session 5 @ blue 2, poster session 5, poster session 10, oral session 5

We propose a computationally-friendly adaptive learning rate schedule, ``AdaLoss", which directly uses the information of the loss function to adjust the stepsize in gradient descent methods. We prove that this schedule enjoys linear convergence in linear regression. Moreover, we extend the to the non-convex regime, in the context of two-layer over-parameterized neural networks. If the width is sufficiently large (polynomially), then AdaLoss converges robustly \emph{to the global minimum} in polynomial time. We numerically verify the theoretical results and extend the scope of the numerical experiments by considering applications in LSTM models for text clarification and policy gradients for control problems.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

Differential Privacy offers strong guarantees such as immutable privacy under any post-processing. In this work, we propose a differentially private mechanism called PrivateMail for performing supervised manifold learning. We then apply it to the use case of private image retrieval to obtain nearest matches to a client’s target image from a server’s database. PrivateMail releases the target image as part of a differentially private manifold embedding. We give bounds on the global sensitivity of the manifold learning map in order to obfuscate and release embeddings with differential privacy inducing noise. We show that PrivateMail obtains a substantially better performance in terms of the privacy-utility trade off in comparison to several baselines on various datasets. We share code for applying PrivateMail at http://tiny.cc/PrivateMail.

poster session 5 @ blue 2, poster session 10 @ blue 2, oral session 5 @ blue 2, poster session 5, poster session 10, oral session 5

Despite deep learning (DL) success in classification problems, DL classifiers do not provide a sound mechanism to decide when to refrain from predicting. Recent works tried to control the overall prediction risk with classification with rejection options. However, existing works overlook the different significance of different classes. We introduce Set-classifier with class-specific RIsk Bounds (SCRIB) to tackle this problem, assigning multiple labels to each example. Given the output of a black-box model on the validation set, SCRIB constructs a set-classifier that controls the class-specific prediction risks. The key idea is to reject when the set classifier returns more than one label. We validated SCRIB on several medical applications, including sleep staging on electroencephalogram(EEG) data, X-ray COVID image classification, and atrial fibrillation detection based on electrocardiogram (ECG) data.SCRIB obtained desirable class-specific risks, which are 35%-88% closer to the target risks than baseline methods.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

The optimization with orthogonality has been shown useful in training deep neural networks (DNNs). To impose orthogonality on DNNs, both computational efficiency and stability are important. However, existing methods utilizing Riemannian optimization or hard constraints can only ensure stability while those using soft constraints can only improve efficiency. In this paper, we propose a novel method, named Feedback Gradient Descent (FGD), to our knowledge, the first work showing high efficiency and stability simultaneously. FGD induces orthogonality based on the simple yet indispensable Euler discretization of a continuous-time dynamical system on the tangent bundle of the Stiefel manifold. In particular, inspired by a numerical integration method on manifolds called Feedback Integrators, we propose to instantiate it on the tangent bundle of the Stiefel manifold for the first time. In our extensive image classification experiments, FGD comprehensively outperforms the existing state-of-the-art methods in terms of accuracy, efficiency, and stability.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

We study the convergence rate of gradient-based local search methods for solving low-rank matrix recovery problems with general objectives in both symmetric and asymmetric cases, under the assumption of the restricted isometry property. First, we develop a new technique to verify the Polyak-Lojasiewicz inequality in a neighborhood of the global minimizers, which leads to a local linear convergence region for the gradient descent method. Second, based on the local convergence result and a sharp strict saddle property proven in this paper, we present two new conditions that guarantee the global linear convergence of the perturbed gradient descent method. The developed local and global convergence results provide much stronger theoretical guarantees than the existing results. As a by-product, this work significantly improves the existing bounds on the RIP constant required to guarantee the non-existence of spurious solutions.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

In many real-world problems, first-order (FO) derivative evaluations are too expensive or even inaccessible. For solving these problems, zeroth-order (ZO) methods that only need function evaluations are often more efficient than FO methods or sometimes the only options. In this paper, we propose a novel zeroth-order inexact augmented Lagrangian method (ZO-iALM) to solve black-box optimization problems, which involve a composite (i.e., smooth+nonsmooth) objective and functional constraints. Under a certain regularity condition (also assumed by several existing works on FO methods), the query complexity of our ZO-iALM is $\tilde{O}(d\vareps^{-3})$ to find an $\vareps$-KKT point for problems with a nonconvex objective and nonconvex constraints, and $\tilde{O}(d\vareps^{-\frac{5}{2}})$ for nonconvex problems with convex constraints, where $d$ is the variable dimension. This appears to be the first work that develops an iALM-based ZO method for functional constrained optimization and meanwhile achieves query complexity results matching the best-known FO complexity results up to a factor of $d$. With an extensive experimental study, we show the effectiveness of our method. The applications of our method span from classical optimization problems to practical machine learning examples such as resource allocation in sensor networks and adversarial example generation.

poster session 5 @ blue 2, poster session 10 @ blue 2, oral session 10 @ blue 2, poster session 5, poster session 10, oral session 10

In social choice theory, (Kemeny) rank aggregation is a well-studied problem where the goal is to combine rankings from multiple voters into a single ranking on the same set of items. Since rankings can reveal preferences of voters (which a voter might like to keep private), it is important to aggregate preferences in such a way to preserve privacy. In this work, we present differentially private algorithms for rank aggregation in the pure and approximate settings along with distribution-independent utility upper and lower bounds. In addition to bounds in the central model, we also present utility bounds for the local model of differential privacy.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

We study the problem of learning generative adversarial networks (GANs) for a rare class of an unlabeled dataset subject to a labeling budget. This problem is motivated from practical applications in domains including security (e.g., synthesizing packets for DNS amplification attacks), systems and networking (e.g., synthesizing workloads that trigger high resource usage), and machine learning (e.g., generating images from a rare class). Existing approaches are unsuitable, either requiring fully-labeled datasets or sacrificing the fidelity of the rare class for that of the common classes. We propose RareGAN, a novel synthesis of three key ideas: (1) extending conditional GANs to use labelled and unlabelled data for better generalization; (2) an active learning approach that requests the most useful labels; and (3) a weighted loss function to favor learning the rare class. We show that RareGAN achieves a better fidelity-diversity tradeoff on the rare class than prior work across different applications, budgets, rare class fractions, GAN losses, and architectures.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

In this paper, we study a general low-rank matrix recovery problem with linear measurements corrupted by some noise. The objective is to understand under what conditions on the restricted isometry property (RIP) of the problem local search methods can find the ground truth with a small error. By analyzing the landscape of the non-convex problem, we first propose a global guarantee on the maximum distance between an arbitrary local minimizer and the ground truth under the assumption that the RIP constant is smaller than $1/2$. We show that this distance shrinks to zero as the intensity of the noise reduces. Our new guarantee is sharp in terms of the RIP constant and is much stronger than the existing results. We then present a local guarantee for problems with an arbitrary RIP constant, which states that any local minimizer is either considerably close to the ground truth or far away from it. Next, we prove the strict saddle property, which guarantees the global convergence of the perturbed gradient descent method in polynomial time. The developed results demonstrate how the noise intensity and the RIP constant of the problem affect the landscape of the problem.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

Submodular functions are at the core of many machine learning and data mining tasks. The underlying submodular functions for many of these tasks are decomposable, i.e., they are sum of several simple submodular functions. In many data intensive applications, however, the number of underlying submodular functions in the original function is so large that we need prohibitively large amount of time to process it and/or it does not even fit in the main memory. To overcome this issue, we introduce the notion of sparsification for decomposable submodular functions whose objective is to obtain an accurate approximation of the original function that is a (weighted) sum of only a few submodular functions. Our main result is a polynomial-time randomized sparsification algorithm such that the expected number of functions used in the output is independent of the number of underlying submodular functions in the original function. We also study the effectiveness of our algorithm under various constraints such as matroid and cardinality constraints. We complement our theoretical analysis with an empirical study of the performance of our algorithm.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

We study the problem of releasing the set union of users' items subject to differential privacy. Previous approaches consider only the set of items for each user as the input. We propose incorporating the item frequency, which is typically available in set union problems, to boost the utility of private mechanisms. However, using the global item frequency over all users would largely increase privacy loss. We propose to use the local item frequency of each user to approximate the global item frequency without incurring additional privacy loss. Local item frequency allows us to design greedy set union mechanisms that are differentially private, which is impossible for previous greedy proposals. Moreover, while all previous works have to use uniform sampling to limit the number of items each user would contribute to, our construction eliminates the sampling step completely and allows our mechanisms to consider all of the users' items. Finally, we propose to transfer the knowledge of the global item frequency from a public dataset into our mechanism, which further boosts utility even when the public and private datasets are from different domains. We evaluate the proposed methods on multiple real-life datasets.

poster session 5 @ blue 2, poster session 10 @ blue 2, poster session 5, poster session 10

We consider the problem of automatically constructing computer programs from input-output examples. We investigate how to augment probabilistic and neural program synthesis methods with new search algorithms, proposing a framework called distribution-based search. Within this framework, we introduce two new search algorithms: Heap Search, an enumerative method, and SQRT Sampling, a probabilistic method. We prove certain optimality guarantees for both methods, show how they integrate with probabilistic and neural techniques, and demonstrate how they can operate at scale across parallel compute environments. Collectively these findings offer theoretical and applied studies of search algorithms for program synthesis that integrate with recent developments in machine-learned program synthesizers.

poster session 5 @ blue 2, poster session 10 @ blue 2, oral session 5 @ blue 2, poster session 5, poster session 10, oral session 5