Cancer is the second leading cause of death in the world. Diagnosing cancer early on can save many lives. Pathologists have to look at tissue microarray (TMA) images manually to identify tumors, which can be time-consuming, inconsistent and subjective. Existing automatic algorithms either have not achieved the accuracy level of a pathologist or require substantial human involvements. A major challenge is that TMA images with different shapes, sizes, and locations can have the same score. Learning staining patterns in TMA images requires a huge number of images, which are severely limited due to privacy and regulation concerns in medical organizations. TMA images from different cancer types may share certain common characteristics, but combining them directly harms the accuracy due to heterogeneity in their staining patterns. Transfer learning is an emerging learning paradigm that allows borrowing strength from similar problems. However, existing approaches typically require a large sample from similar learning problems, while TMA images of different cancer types are often available in small sample size and further existing algorithms are limited to transfer learning from one similar problem. We propose a new transfer learning algorithm that could learn from multiple related problems, where each problem has a small sample and can have a substantially different distribution from the original one. The proposed algorithm has made it possible to break the critical accuracy barrier (the 75% accuracy level of pathologists), with a reported accuracy of 75.9% on breast cancer TMA images from the Stanford Tissue Microarray Database. It is supported by recent developments in transfer learning theory and empirical evidence in clustering technology. This will allow pathologists to confidently adopt automatic algorithms in recognizing tumors consistently with a higher accuracy in real time.
This tutorial survey provides an overview of recent non-asymptotic advances in statistical learning theory as relevant to control and system identification. While there has been substantial progress across all areas of control, the theory is most well-developed when it comes to linear system identification and learning for the linear quadratic regulator, which are the focus of this manuscript. From a theoretical perspective, much of the labor underlying these advances has been in adapting tools from modern high-dimensional statistics and learning theory. While highly relevant to control theorists interested in integrating tools from machine learning, the foundational material has not always been easily accessible. To remedy this, we provide a self-contained presentation of the relevant material, outlining all the key ideas and the technical machinery that underpin recent results. We also present a number of open problems and future directions.
We consider the tasks of learning quantum states, measurements and channels generated by continuous-variable (CV) quantum circuits. This family of circuits is suited to describe optical quantum technologies and in particular it includes state-of-the-art photonic processors capable of showing quantum advantage. We define classes of functions that map classical variables, encoded into the CV circuit parameters, to outcome probabilities evaluated on those circuits. We then establish efficient learnability guarantees for such classes, by computing bounds on their pseudo-dimension or covering numbers, showing that CV quantum circuits can be learned with a sample complexity that scales polynomially with the circuit's size, i.e., the number of modes. Our results show that CV circuits can be trained efficiently using a number of training samples that, unlike their finite-dimensional counterpart, does not scale with the circuit depth.
Estimating the average treatment effect (ATE) from observational data is challenging due to selection bias. Existing works mainly tackle this challenge in two ways. Some researchers propose constructing a score function that satisfies the orthogonal condition, which guarantees that the established ATE estimator is "orthogonal" to be more robust. The others explore representation learning models to achieve a balanced representation between the treated and the controlled groups. However, existing studies fail to 1) discriminate treated units from controlled ones in the representation space to avoid the over-balanced issue; 2) fully utilize the "orthogonality information". In this paper, we propose a moderately-balanced representation learning (MBRL) framework based on recent covariates balanced representation learning methods and orthogonal machine learning theory. This framework protects the representation from being over-balanced via multi-task learning. Simultaneously, MBRL incorporates the noise orthogonality information in the training and validation stages to achieve a better ATE estimation. The comprehensive experiments on benchmark and simulated datasets show the superiority and robustness of our method on treatment effect estimations compared with existing state-of-the-art methods.
Learning curves plot the expected error of a learning algorithm as a function of the number of labeled samples it receives from a target distribution. They are widely used as a measure of an algorithm's performance, but classic PAC learning theory cannot explain their behavior. As observed by Antos and Lugosi (1996 , 1998), the classic `No Free Lunch' lower bounds only trace the upper envelope above all learning curves of specific target distributions. For a concept class with VC dimension $d$ the classic bound decays like $d/n$, yet it is possible that the learning curve for \emph{every} specific distribution decays exponentially. In this case, for each $n$ there exists a different `hard' distribution requiring $d/n$ samples. Antos and Lugosi asked which concept classes admit a `strong minimax lower bound' -- a lower bound of $d'/n$ that holds for a fixed distribution for infinitely many $n$. We solve this problem in a principled manner, by introducing a combinatorial dimension called VCL that characterizes the best $d'$ for which $d'/n$ is a strong minimax lower bound. Our characterization strengthens the lower bounds of Bousquet, Hanneke, Moran, van Handel, and Yehudayoff (2021), and it refines their theory of learning curves, by showing that for classes with finite VCL the learning rate can be decomposed into a linear component that depends only on the hypothesis class and an exponential component that depends also on the target distribution. As a corollary, we recover the lower bound of Antos and Lugosi (1996 , 1998) for half-spaces in $\mathbb{R}^d$. Finally, to provide another viewpoint on our work and how it compares to traditional PAC learning bounds, we also present an alternative formulation of our results in a language that is closer to the PAC setting.
With the mass construction of Gen III nuclear reactors, it is a popular trend to use deep learning (DL) techniques for fast and effective diagnosis of possible accidents. To overcome the common problems of previous work in diagnosing reactor accidents using deep learning theory, this paper proposes a diagnostic process that ensures robustness to noisy and crippled data and is interpretable. First, a novel Denoising Padded Autoencoder (DPAE) is proposed for representation extraction of monitoring data, with representation extractor still effective on disturbed data with signal-to-noise ratios up to 25.0 and monitoring data missing up to 40.0%. Secondly, a diagnostic framework using DPAE encoder for extraction of representations followed by shallow statistical learning algorithms is proposed, and such stepwise diagnostic approach is tested on disturbed datasets with 41.8% and 80.8% higher classification and regression task evaluation metrics, in comparison with the end-to-end diagnostic approaches. Finally, a hierarchical interpretation algorithm using SHAP and feature ablation is presented to analyze the importance of the input monitoring parameters and validate the effectiveness of the high importance parameters. The outcomes of this study provide a referential method for building robust and interpretable intelligent reactor anomaly diagnosis systems in scenarios with high safety requirements.
Modern machine learning tasks often require considering not just one but multiple objectives. For example, besides the prediction quality, this could be the efficiency, robustness or fairness of the learned models, or any of their combinations. Multi-objective learning offers a natural framework for handling such problems without having to commit to early trade-offs. Surprisingly, statistical learning theory so far offers almost no insight into the generalization properties of multi-objective learning. In this work, we make first steps to fill this gap: we establish foundational generalization bounds for the multi-objective setting as well as generalization and excess bounds for learning with scalarizations. We also provide the first theoretical analysis of the relation between the Pareto-optimal sets of the true objectives and the Pareto-optimal sets of their empirical approximations from training data. In particular, we show a surprising asymmetry: all Pareto-optimal solutions can be approximated by empirically Pareto-optimal ones, but not vice versa.
This note serves three purposes: (i) we provide a self-contained exposition of the fact that conjunctive queries are not efficiently learnable in the Probably-Approximately-Correct (PAC) model, paying clear attention to the complicating fact that this concept class lacks the polynomial-size fitting property, a property that is tacitly assumed in much of the computational learning theory literature; (ii) we establish a strong negative PAC learnability result that applies to many restricted classes of conjunctive queries (CQs), including acyclic CQs for a wide range of notions of "acyclicity"; (iii) we show that CQs (and UCQs) are efficiently PAC learnable with membership queries.
Generalization analyses of deep learning typically assume that the training converges to a fixed point. But, recent results indicate that in practice, the weights of deep neural networks optimized with stochastic gradient descent often oscillate indefinitely. To reduce this discrepancy between theory and practice, this paper focuses on the generalization of neural networks whose training dynamics do not necessarily converge to fixed points. Our main contribution is to propose a notion of statistical algorithmic stability (SAS) that extends classical algorithmic stability to non-convergent algorithms and to study its connection to generalization. This ergodic-theoretic approach leads to new insights when compared to the traditional optimization and learning theory perspectives. We prove that the stability of the time-asymptotic behavior of a learning algorithm relates to its generalization and empirically demonstrate how loss dynamics can provide clues to generalization performance. Our findings provide evidence that networks that "train stably generalize better" even when the training continues indefinitely and the weights do not converge.
Despite years of effort, the quantum machine learning community has only been able to show quantum learning advantages for certain contrived cryptography-inspired datasets in the case of classical data. In this note, we discuss the challenges of finding learning problems that quantum learning algorithms can learn much faster than any classical learning algorithm, and we study how to identify such learning problems. Specifically, we reflect on the main concepts in computational learning theory pertaining to this question, and we discuss how subtle changes in definitions can mean conceptually significantly different tasks, which can either lead to a separation or no separation at all. Moreover, we study existing learning problems with a provable quantum speedup to distill sets of more general and sufficient conditions (i.e., ``checklists'') for a learning problem to exhibit a separation between classical and quantum learners. These checklists are intended to streamline one's approach to proving quantum speedups for learning problems, or to elucidate bottlenecks. Finally, to illustrate its application, we analyze examples of potential separations (i.e., when the learning problem is build from computational separations, or when the data comes from a quantum experiment) through the lens of our approach.
Natural language processing (NLP) technology has shown great commercial value in applications such as sentiment analysis. But NLP models are vulnerable to the threat of pirated redistribution, damaging the economic interests of model owners. Digital watermarking technology is an effective means to protect the intellectual property rights of NLP model. The existing NLP model protection mainly designs watermarking schemes by improving both security and robustness purposes, however, the security and robustness of these schemes have the following problems, respectively: (1) Watermarks are difficult to defend against fraudulent declaration by adversary and are easily detected and blocked from verification by human or anomaly detector during the verification process. (2) The watermarking model cannot meet multiple robustness requirements at the same time. To solve the above problems, this paper proposes a novel watermarking framework for NLP model based on the over-parameterization of depth model and the multi-task learning theory. Specifically, a covert trigger set is established to realize the perception-free verification of the watermarking model, and a novel auxiliary network is designed to improve the robustness and security of the watermarking model. The proposed framework was evaluated on two benchmark datasets and three mainstream NLP models, and the results show that the framework can successfully validate model ownership with 100% validation accuracy and advanced robustness and security without compromising the host model performance.
Proof-of-Learning (PoL) proposes that a model owner logs training checkpoints to establish a proof of having expended the computation necessary for training. The authors of PoL forego cryptographic approaches and trade rigorous security guarantees for scalability to deep learning. They empirically argued the benefit of this approach by showing how spoofing--computing a proof for a stolen model--is as expensive as obtaining the proof honestly by training the model. However, recent work has provided a counter-example and thus has invalidated this observation. In this work we demonstrate, first, that while it is true that current PoL verification is not robust to adversaries, recent work has largely underestimated this lack of robustness. This is because existing spoofing strategies are either unreproducible or target weakened instantiations of PoL--meaning they are easily thwarted by changing hyperparameters of the verification. Instead, we introduce the first spoofing strategies that can be reproduced across different configurations of the PoL verification and can be done for a fraction of the cost of previous spoofing strategies. This is possible because we identify key vulnerabilities of PoL and systematically analyze the underlying assumptions needed for robust verification of a proof. On the theoretical side, we show how realizing these assumptions reduces to open problems in learning theory.We conclude that one cannot develop a provably robust PoL verification mechanism without further understanding of optimization in deep learning.
Semantic Web or Knowledge Graphs (KG) emerged to one of the most important information source for intelligent systems requiring access to structured knowledge. One of the major challenges is the extraction and processing of unambiguous information from textual data. Following the human perception, overlapping semantic linkages between two named entities become clear due to our common-sense about the context a relationship lives in which is not the case when we look at it from an automatically driven process of a machine. In this work, we are interested in the problem of Relational Resolution within the scope of KGs, i.e, we are investigating the inherent semantic of relationships between entities within a network. We propose a new adaptive AutoEncoder, called V-Coder, to identify relations inherently connecting entities from different domains. Those relations can be considered as being ambiguous and are candidates for disentanglement. Likewise to the Adaptive Learning Theory (ART), our model learns new patterns from the KG by increasing units in a competitive layer without discarding the previous observed patterns whilst learning the quality of each relation separately. The evaluation on real-world datasets of Freebase, Yago and NELL shows that the V-Coder is not only able to recover links from corrupted input data, but also shows that the semantic disclosure of relations in a KG show the tendency to improve link prediction. A semantic evaluation wraps the evaluation up.
This paper seeks to answer the following question: \textit{"What can we learn by predicting accuracy?"}. Indeed, classification is one of the most popular tasks in machine learning, and many loss functions have been developed to maximize this non-differentiable objective function. Unlike past work on loss function design, which was guided mainly by intuition and theory before being validated by experimentation, here we propose to approach this problem in the opposite way: we seek to extract knowledge by experimentation. This data-driven approach is similar to that used in physics to discover general laws from data. We used a symbolic regression method to automatically find a mathematical expression highly correlated with a linear classifier's accuracy. The formula discovered on more than 260 datasets of embeddings has a Pearson's correlation of 0.96 and a $r^2$ of 0.93. More interestingly, this formula is highly explainable and confirms insights from various previous papers on loss design. We hope this work will open new perspectives in the search for new heuristics leading to a deeper understanding of machine learning theory.
Deep learning methodology contributes a lot to the development of hyperspectral image (HSI) analysis community. However, it also makes HSI analysis systems vulnerable to adversarial attacks. To this end, we propose a masked spatial-spectral autoencoder (MSSA) in this paper under self-supervised learning theory, for enhancing the robustness of HSI analysis systems. First, a masked sequence attention learning module is conducted to promote the inherent robustness of HSI analysis systems along spectral channel. Then, we develop a graph convolutional network with learnable graph structure to establish global pixel-wise combinations.In this way, the attack effect would be dispersed by all the related pixels among each combination, and a better defense performance is achievable in spatial aspect.Finally, to improve the defense transferability and address the problem of limited labelled samples, MSSA employs spectra reconstruction as a pretext task and fits the datasets in a self-supervised manner.Comprehensive experiments over three benchmarks verify the effectiveness of MSSA in comparison with the state-of-the-art hyperspectral classification methods and representative adversarial defense strategies.
Investigating cooperativity of interlocutors is central in studying pragmatics of dialogue. Models of conversation that only assume cooperative agents fail to explain the dynamics of strategic conversations. Thus, we investigate the ability of agents to identify non-cooperative interlocutors while completing a concurrent visual-dialogue task. Within this novel setting, we study the optimality of communication strategies for achieving this multi-task objective. We use the tools of learning theory to develop a theoretical model for identifying non-cooperative interlocutors and apply this theory to analyze different communication strategies. We also introduce a corpus of non-cooperative conversations about images in the GuessWhat?! dataset proposed by De Vries et al. (2017). We use reinforcement learning to implement multiple communication strategies in this context and find empirical results validate our theory.
The practical success of overparameterized neural networks has motivated the recent scientific study of interpolating methods, which perfectly fit their training data. Certain interpolating methods, including neural networks, can fit noisy training data without catastrophically bad test performance, in defiance of standard intuitions from statistical learning theory. Aiming to explain this, a body of recent work has studied benign overfitting, a phenomenon where some interpolating methods approach Bayes optimality, even in the presence of noise. In this work we argue that while benign overfitting has been instructive and fruitful to study, many real interpolating methods like neural networks do not fit benignly: modest noise in the training set causes nonzero (but non-infinite) excess risk at test time, implying these models are neither benign nor catastrophic but rather fall in an intermediate regime. We call this intermediate regime tempered overfitting, and we initiate its systematic study. We first explore this phenomenon in the context of kernel (ridge) regression (KR) by obtaining conditions on the ridge parameter and kernel eigenspectrum under which KR exhibits each of the three behaviors. We find that kernels with powerlaw spectra, including Laplace kernels and ReLU neural tangent kernels, exhibit tempered overfitting. We then empirically study deep neural networks through the lens of our taxonomy, and find that those trained to interpolation are tempered, while those stopped early are benign. We hope our work leads to a more refined understanding of overfitting in modern learning.
Compressed sensing has been a very successful high-dimensional signal acquisition and recovery technique that relies on linear operations. However, the actual measurements of signals have to be quantized before storing or processing. 1(One)-bit compressed sensing is a heavily quantized version of compressed sensing, where each linear measurement of a signal is reduced to just one bit: the sign of the measurement. Once enough of such measurements are collected, the recovery problem in 1-bit compressed sensing aims to find the original signal with as much accuracy as possible. The recovery problem is related to the traditional "halfspace-learning" problem in learning theory. For recovery of sparse vectors, a popular reconstruction method from 1-bit measurements is the binary iterative hard thresholding (BIHT) algorithm. The algorithm is a simple projected sub-gradient descent method, and is known to converge well empirically, despite the nonconvexity of the problem. The convergence property of BIHT was not theoretically justified, except with an exorbitantly large number of measurements (i.e., a number of measurement greater than $\max\{k^{10}, 24^{48}, k^{3.5}/\epsilon\}$, where $k$ is the sparsity, $\epsilon$ denotes the approximation error, and even this expression hides other factors). In this paper we show that the BIHT algorithm converges with only $\tilde{O}(\frac{k}{\epsilon})$ measurements. Note that, this dependence on $k$ and $\epsilon$ is optimal for any recovery method in 1-bit compressed sensing. With this result, to the best of our knowledge, BIHT is the only practical and efficient (polynomial time) algorithm that requires the optimal number of measurements in all parameters (both $k$ and $\epsilon$). This is also an example of a gradient descent algorithm converging to the correct solution for a nonconvex problem, under suitable structural conditions.
We study the continuous-time counterpart of Q-learning for reinforcement learning (RL) under the entropy-regularized, exploratory diffusion process formulation introduced by Wang et al. (2020). As the conventional (big) Q-function collapses in continuous time, we consider its first-order approximation and coin the term ``(little) q-function". This function is related to the instantaneous advantage rate function as well as the Hamiltonian. We develop a ``q-learning" theory around the q-function that is independent of time discretization. Given a stochastic policy, we jointly characterize the associated q-function and value function by martingale conditions of certain stochastic processes, in both on-policy and off-policy settings. We then apply the theory to devise different actor-critic algorithms for solving underlying RL problems, depending on whether or not the density function of the Gibbs measure generated from the q-function can be computed explicitly. One of our algorithms interprets the well-known Q-learning algorithm SARSA, and another recovers a policy gradient (PG) based continuous-time algorithm proposed in Jia and Zhou (2022b). Finally, we conduct simulation experiments to compare the performance of our algorithms with those of PG-based algorithms in Jia and Zhou (2022b) and time-discretized conventional Q-learning algorithms.
Recent works show that adversarial examples exist for random neural networks [Daniely and Schacham, 2020] and that these examples can be found using a single step of gradient ascent [Bubeck et al., 2021]. In this work, we extend this line of work to "lazy training" of neural networks -- a dominant model in deep learning theory in which neural networks are provably efficiently learnable. We show that over-parametrized neural networks that are guaranteed to generalize well and enjoy strong computational guarantees remain vulnerable to attacks generated using a single step of gradient ascent.