Machine learning malware detectors are vulnerable to adversarial EXEmples, i.e., carefully-crafted Windows programs tailored to evade detection. Unlike other adversarial problems, attacks in this context must be functionality-preserving, a constraint that is challenging to address. As a consequence, heuristic algorithms are typically used, which inject new content, either randomly-picked or harvested from legitimate programs. In this paper, we show how learning malware detectors can be cast within a zeroth-order optimization framework, which allows incorporating functionality-preserving manipulations. This permits the deployment of sound and efficient gradient-free optimization algorithms, which come with theoretical guarantees and allow for minimal hyper-parameters tuning. As a by-product, we propose and study ZEXE, a novel zeroth-order attack against Windows malware detection. Compared to state-of-the-art techniques, ZEXE provides improvement in the evasion rate, reducing to less than one third the size of the injected content.

It has recently been argued that AI models' representations are becoming aligned as their scale and performance increase. Empirical analyses have been designed to support this idea and conjecture the possible alignment of different representations toward a shared statistical model of reality. In this paper, we propose a learning-theoretic perspective to representation alignment. First, we review and connect different notions of alignment based on metric, probabilistic, and spectral ideas. Then, we focus on stitching, a particular approach to understanding the interplay between different representations in the context of a task. Our main contribution here is to relate the properties of stitching to the kernel alignment of the underlying representation. Our results can be seen as a first step toward casting representation alignment as a learning-theoretic problem.

This paper addresses the covariate shift problem in the context of nonparametric regression within reproducing kernel Hilbert spaces (RKHSs). Covariate shift arises in supervised learning when the input distributions of the training and test data differ, presenting additional challenges for learning. Although kernel methods have optimal statistical properties, their high computational demands in terms of time and, particularly, memory, limit their scalability to large datasets. To address this limitation, the main focus of this paper is to explore the trade-off between computational efficiency and statistical accuracy under covariate shift. We investigate the use of random projections where the hypothesis space consists of a random subspace within a given RKHS. Our results show that, even in the presence of covariate shift, significant computational savings can be achieved without compromising learning performance.

In this work we consider the problem of numerical integration, i.e., approximating integrals with respect to a target probability measure using only pointwise evaluations of the integrand. We focus on the setting in which the target distribution is only accessible through a set of $n$ i.i.d. observations, and the integrand belongs to a reproducing kernel Hilbert space. We propose an efficient procedure which exploits a small i.i.d. random subset of $m \lt n$ samples drawn either uniformly or using approximate leverage scores from the initial observations. Our main result is an upper bound on the approximation error of this procedure for both sampling strategies. It yields sufficient conditions on the subsample size to recover the standard (optimal) $n^{-1/2}$ rate while reducing drastically the number of functions evaluations---and thus the overall computational cost. Moreover, we obtain rates with respect to the number $m$ of evaluations of the integrand which adapt to its smoothness, and match known optimal rates for instance for Sobolev spaces. We illustrate our theoretical findings with numerical experiments on real datasets, which highlight the attractive efficiency-accuracy tradeoff of our method compared to existing randomized and greedy quadrature methods. We note that, the problem of numerical integration in RKHS amounts to designing a discrete approximation of the kernel mean embedding of the target distribution. As a consequence, direct applications of our results also include the efficient computation of maximum mean discrepancies between distributions and the design of efficient kernel-based tests.

The increasing deployment of robots has significantly enhanced the automation levels across a wide and diverse range of industries. This article investigates the automation challenges of laser-based dermatology procedures in the beauty industry. This group of related manipulation tasks involves delivering energy from a cosmetic laser onto the skin with repetitive patterns. To automate this procedure, we propose to use a robotic manipulator and endow it with the dexterity of a skilled dermatology practitioner through a learning-from-demonstration framework. To ensure that the cosmetic laser can properly deliver the energy onto the skin surface of an individual, we develop a novel structured prediction-based imitation learning algorithm with the merit of handling geometric constraints. Notably, our proposed algorithm effectively tackles the imitation challenges associated with quasi-periodic motions, a common feature of many laser-based cosmetic tasks. The conducted real-world experiments illustrate the performance of our robotic beautician in mimicking realistic dermatological procedures. Our new method is shown to not only replicate the rhythmic movements from the provided demonstrations but also to adapt the acquired skills to previously unseen scenarios and subjects.

In this paper, we address the Sim2Real gap in the field of vision-based tactile sensors for classifying object surfaces. We train a Diffusion Model to bridge this gap using a relatively small dataset of real-world images randomly collected from unlabeled everyday objects via the DIGIT sensor. Subsequently, we employ a simulator to generate images by uniformly sampling the surface of objects from the YCB Model Set. These simulated images are then translated into the real domain using the Diffusion Model and automatically labeled to train a classifier. During this training, we further align features of the two domains using an adversarial procedure. Our evaluation is conducted on a dataset of tactile images obtained from a set of ten 3D-printed YCB objects. The results reveal a total accuracy of 81.9%, a significant improvement compared to the 34.7% achieved by the classifier trained solely on simulated images. This demonstrates the effectiveness of our approach. We further validate our approach using the classifier on a 6D object pose estimation task from tactile data.

The impressive performance of deep learning architectures is associated with a massive increase in model complexity. Millions of parameters need to be tuned, with training and inference time scaling accordingly, together with energy consumption. But is massive fine-tuning always necessary? In this paper, focusing on image classification, we consider a simple transfer learning approach exploiting pre-trained convolutional features as input for a fast-to-train kernel method. We refer to this approach as top-tuning since only the kernel classifier is trained on the target dataset. In our study, we perform more than 3000 training processes focusing on 32 small to medium-sized target datasets, a typical situation where transfer learning is necessary. We show that the top-tuning approach provides comparable accuracy with respect to fine-tuning, with a training time between one and two orders of magnitude smaller. These results suggest that top-tuning is an effective alternative to fine-tuning in small/medium datasets, being especially useful when training time efficiency and computational resources saving are crucial.

We investigate an extension of classical empirical risk minimization, where the hypothesis space consists of a random subspace within a given Hilbert space. Specifically, we examine the Nyström method where the subspaces are defined by a random subset of the data. This approach recovers Nyström approximations used in kernel methods as a specific case. Using random subspaces naturally leads to computational advantages, but a key question is whether it compromises the learning accuracy. Recently, the tradeoffs between statistics and computation have been explored for the square loss and self-concordant losses, such as the logistic loss. In this paper, we extend these analyses to general convex Lipschitz losses, which may lack smoothness, such as the hinge loss used in support vector machines. Our main results show the existence of various scenarios where computational gains can be achieved without sacrificing learning performance. When specialized to smooth loss functions, our analysis recovers most previous results. Moreover, it allows to consider classification problems and translate the surrogate risk bounds into classification error bounds. Indeed, this gives the opportunity to compare the effect of Nyström approximations when combined with different loss functions such as the hinge or the square loss.

Multi-fingered robotic hands have potential to enable robots to perform sophisticated manipulation tasks. However, teaching a robot to grasp objects with an anthropomorphic hand is an arduous problem due to the high dimensionality of state and action spaces. Deep Reinforcement Learning (DRL) offers techniques to design control policies for this kind of problems without explicit environment or hand modeling. However, state-of-the-art model-free algorithms have proven inefficient for learning such policies. The main problem is that the exploration of the environment is unfeasible for such high-dimensional problems, thus hampering the initial phases of policy optimization. One possibility to address this is to rely on off-line task demonstrations, but, oftentimes, this is too demanding in terms of time and computational resources. To address these problems, we propose the A Grasp Pose is All You Need (G-PAYN) method for the anthropomorphic hand of the iCub humanoid. We develop an approach to automatically collect task demonstrations to initialize the training of the policy. The proposed grasping pipeline starts from a grasp pose generated by an external algorithm, used to initiate the movement. Then a control policy (previously trained with the proposed G-PAYN) is used to reach and grab the object. We deployed the iCub into the MuJoCo simulator and use it to test our approach with objects from the YCB-Video dataset. Results show that G-PAYN outperforms current DRL techniques in the considered setting in terms of success rate and execution time with respect to the baselines. The code to reproduce the experiments is released together with the paper with an open source license11https://github.com/hsp-iit/rl-icub-dexterous-manipulation.

The theory of Koopman operators allows to deploy non-parametric machine learning algorithms to predict and analyze complex dynamical systems. Estimators such as principal component regression (PCR) or reduced rank regression (RRR) in kernel spaces can be shown to provably learn Koopman operators from finite empirical observations of the system's time evolution. Scaling these approaches to very long trajectories is a challenge and requires introducing suitable approximations to make computations feasible. In this paper, we boost the efficiency of different kernel-based Koopman operator estimators using random projections (sketching). We derive, implement and test the new ``sketched'' estimators with extensive experiments on synthetic and large-scale molecular dynamics datasets. Further, we establish non asymptotic error bounds giving a sharp characterization of the trade-offs between statistical learning rates and computational efficiency. Our empirical and theoretical analysis shows that the proposed estimators provide a sound and efficient way to learn large scale dynamical systems. In particular our experiments indicate that the proposed estimators retain the same accuracy of PCR or RRR, while being much faster.

When domain knowledge is limited and experimentation is restricted by ethical, financial, or time constraints, practitioners turn to observational causal discovery methods to recover the causal structure, exploiting the statistical properties of their data. Because causal discovery without further assumptions is an ill-posed problem, each algorithm comes with its own set of usually untestable assumptions, some of which are hard to meet in real datasets. Motivated by these considerations, this paper extensively benchmarks the empirical performance of recent causal discovery methods on observational _iid_ data generated under different background conditions, allowing for violations of the critical assumptions required by each selected approach. Our experimental findings show that score matching-based methods demonstrate surprising performance in the false positive and false negative rate of the inferred graph in these challenging scenarios, and we provide theoretical insights into their performance. This work is also the first effort to benchmark the stability of causal discovery algorithms with respect to the values of their hyperparameters. Finally, we hope this paper will set a new standard for the evaluation of causal discovery methods and can serve as an accessible entry point for practitioners interested in the field, highlighting the empirical implications of different algorithm choices.

Finite-difference methods are a class of algorithms designed to solve black-box optimization problems by approximating a gradient of the target function on a set of directions. In black-box optimization, the non-smooth setting is particularly relevant since, in practice, differentiability and smoothness assumptions cannot be verified. To cope with nonsmoothness, several authors use a smooth approximation of the target function and show that finite difference methods approximate its gradient. Recently, it has been proved that imposing a structure in the directions allows improving performance. However, only the smooth setting was considered. To close this gap, we introduce and analyze O-ZD, the first structured finite-difference algorithm for non-smooth black-box optimization. Our method exploits a smooth approximation of the target function and we prove that it approximates its gradient on a subset of random {\em orthogonal} directions. We analyze the convergence of O-ZD under different assumptions. For non-smooth convex functions, we obtain the optimal complexity. In the non-smooth non-convex setting, we characterize the number of iterations needed to bound the expected norm of the smoothed gradient. For smooth functions, our analysis recovers existing results for structured zeroth-order methods for the convex case and extends them to the non-convex setting. We conclude with numerical simulations where assumptions are satisfied, observing that our algorithm has very good practical performances.

Compressive learning is an approach to efficient large scale learning based on sketching an entire dataset to a single mean embedding (the sketch), i.e. a vector of generalized moments. The learning task is then approximately solved as an inverse problem using an adapted parametric model. Previous works in this context have focused on sketches obtained by averaging random features, that while universal can be poorly adapted to the problem at hand. In this paper, we propose and study the idea of performing sketching based on data-dependent Nyström approximation. From a theoretical perspective we prove that the excess risk can be controlled under a geometric assumption relating the parametric model used to learn from the sketch and the covariance operator associated to the task at hand. Empirically, we show for k-means clustering and Gaussian modeling that for a fixed sketch size, Nyström sketches indeed outperform those built with random features.

Kernel methods provide a principled approach to nonparametric learning. While their basic implementations scale poorly to large problems, recent advances showed that approximate solvers can efficiently handle massive datasets. A shortcoming of these solutions is that hyperparameter tuning is not taken care of, and left for the user to perform. Hyperparameters are crucial in practice and the lack of automated tuning greatly hinders efficiency and usability. In this paper, we work to fill in this gap focusing on kernel ridge regression based on the Nyström approximation. After reviewing and contrasting a number of hyperparameter tuning strategies, we propose a complexity regularization criterion based on a data dependent penalty, and discuss its efficient optimization. Then, we proceed to a careful and extensive empirical evaluation highlighting strengths and weaknesses of the different tuning strategies. Our analysis shows the benefit of the proposed approach, that we hence incorporate in a library for large scale kernel methods to derive adaptively tuned solutions.

Gaussian process optimization is a successful class of algorithms(e.g. GP-UCB) to optimize a black-box function through sequential evaluations. However, for functions with continuous domains, Gaussian process optimization has to rely on either a fixed discretization of the space, or the solution of a non-convex ptimization subproblem at each evaluation. The first approach can negatively affect performance, while the second approach requires a heavy computational burden. A third option, only recently theoretically studied, is to adaptively discretize the function domain. Even though this approach avoids the extra non-convex optimization costs, the overall computational complexity is still prohibitive. An algorithm such as GP-UCB has a runtime of $O(T^4)$, where $T$ is the number of iterations. In this paper, we introduce Ada-BKB (Adaptive Budgeted Kernelized Bandit), a no-regret Gaussian process optimization algorithm for functions on continuous domains, that provably runs in $O(T^2 d_\text{eff}^2)$, where $d_\text{eff}$ is the effective dimension of the explored space, and which is typically much smaller than $T$. We corroborate our theoretical findings with experiments on synthetic non-convex functions and on the real-world problem of hyper-parameter optimization, confirming the good practical performances of the proposed approach.

Computing a Gaussian process (GP) posterior has a computational cost cubical in the number of historical points. A reformulation of the same GP posterior highlights that this complexity mainly depends on how many unique historical points are considered. This can have important implication in active learning settings, where the set of historical points is constructed sequentially by the learner. We show that sequential black-box optimization based on GPs (GP-Opt) can be made efficient by sticking to a candidate solution for multiple evaluation steps and switch only when necessary. Limiting the number of switches also limits the number of unique points in the history of the GP. Thus, the efficient GP reformulation can be used to exactly and cheaply compute the posteriors required to run the GP-Opt algorithms. This approach is especially useful in real-world applications of GP-Opt with high switch costs (e.g. switching chemicals in wet labs, data/model loading in hyperparameter optimization). As examples of this meta-approach, we modify two well-established GP-Opt algorithms, GP-UCB and GP-EI, to switch candidates as infrequently as possible adapting rules from batched GP-Opt. These versions preserve all the theoretical no-regret guarantees while improving practical aspects of the algorithms such as runtime, memory complexity, and the ability of batching candidates and evaluating them in parallel.

Kernel mean embeddings are a powerful tool to represent probability distributions over arbitrary spaces as single points in a Hilbert space. Yet, the cost of computing and storing such embeddings prohibits their direct use in large-scale settings. We propose an efficient approximation procedure based on the Nystr{ö}m method, which exploits a small random subset of the dataset. Our main result is an upper bound on the approximation error of this procedure. It yields sufficient conditions on the subsample size to obtain the standard (1/sqrt(n)) rate while reducing computational costs. We discuss applications of this result for the approximation of the maximum mean discrepancy and quadrature rules, and we illustrate our theoretical findings with numerical experiments.

We study the behavior of error bounds for multiclass classification under suitable margin conditions. For a wide variety of methods we prove that the classification error under a hard-margin condition decreases exponentially fast without any bias-variance trade-off. Different convergence rates can be obtained in correspondence of different margin assumptions. With a self-contained and instructive analysis we are able to generalize known results from the binary to the multiclass setting.

Monitoring plankton populations in situ is fundamental to preserve the aquatic ecosystem. Plankton microorganisms are in fact susceptible of minor environmental perturbations, that can reflect into consequent morphological and dynamical modifications. Nowadays, the availability of advanced automatic or semi-automatic acquisition systems has been allowing the production of an increasingly large amount of plankton image data. The adoption of machine learning algorithms to classify such data may be affected by the significant cost of manual annotation, due to both the huge quantity of acquired data and the numerosity of plankton species. To address these challenges, we propose an efficient unsupervised learning pipeline to provide accurate classification of plankton microorganisms. We build a set of image descriptors exploiting a two-step procedure. First, a Variational Autoencoder (VAE) is trained on features extracted by a pre-trained neural network. We then use the learnt latent space as image descriptor for clustering. We compare our method with state-of-the-art unsupervised approaches, where a set of pre-defined hand-crafted features is used for clustering of plankton images. The proposed pipeline outperforms the benchmark algorithms for all the plankton datasets included in our analysis, providing better image embedding properties.

We study a class of dynamical systems modelled as stationary Markov chains that admit an invariant distribution via the corresponding transfer or Koopman operator. While data-driven algorithms to reconstruct such operators are well known, their relationship with statistical learning is largely unexplored. We formalize a framework to learn the Koopman operator from finite data trajectories of the dynamical system. We consider the restriction of this operator to a reproducing kernel Hilbert space and introduce a notion of risk, from which different estimators naturally arise. We link the risk with the estimation of the spectral decomposition of the Koopman operator. These observations motivate a reduced-rank operator regression (RRR) estimator. We derive learning bounds for the proposed estimator, holding both in i.i.d and non i.i.d. settings, the latter in terms of mixing coefficients. Our results suggest RRR might be beneficial over other widely used estimators as confirmed in numerical experiments both for forecasting and mode decomposition.

Deep Learning (DL) based methods for object detection achieve remarkable performance at the cost of computationally expensive training and extensive data labeling. Robots embodiment can be exploited to mitigate this burden by acquiring automatically annotated training data via a natural interaction with a human showing the object of interest, hand-held. However, learning solely from this data may introduce biases (the so-called domain shift), and prevents adaptation to novel tasks. While Weakly-supervised Learning offers a well-established set of techniques to cope with these problems in general-purpose Computer Vision, its adoption in challenging robotic domains is still at a preliminary stage. In this work, we target the scenario of a robot trained in a teacher-learner setting to detect handheld objects. The aim is to improve detection performance in different settings by letting the robot explore the environment with a limited human labeling budget. We compare several techniques for WSL in detection pipelines to reduce model re-training costs without compromising accuracy, proposing solutions which target the considered robotic scenario. We show that the robot can improve adaptation to novel domains, either by interacting with a human teacher (Active Learning) or with an autonomous supervision (Semi-supervised Learning). We integrate our strategies into an on-line detection method, achieving efficient model update capabilities with few labels. We experimentally benchmark our method on challenging robotic object detection tasks under domain shift1.

The visual system of a robot has different requirements depending on the application: it may require high accuracy or reliability, be constrained by limited resources, or need fast adaptation to dynamically changing environments. In this article, we focus on the instance segmentation task and provide a comprehensive study of different techniques that allow adapting an object segmentation model in the presence of novel objects or different domains. We propose a pipeline for fast instance segmentation learning designed for robotic applications where data come in stream. It is based on an hybrid method leveraging on a pre-trained convolutional neural network for feature extraction and fast-to-train Kernel-based classifiers. We also propose a training protocol that allows to shorten the training time by performing feature extraction during the data acquisition. We benchmark the proposed pipeline on two robotics datasets and we deploy it on a real robot, i.e., the iCub humanoid. To this aim, we adapt our method to an incremental setting in which novel objects are learned online by the robot. The code to reproduce the experiments is publicly available on GitHub.11[Online]. Available:https://github.com/hsp-iit/online-detection

We study iterative regularization for linear models, when the bias is convex but not necessarily strongly convex. We characterize the stability properties of a primal-dual gradient based approach, analyzing its convergence in the presence of worst case deterministic noise. As a main example, we specialize and illustrate the results for the problem of robust sparse recovery. Key to our analysis is a combination of ideas from regularization theory and optimization in the presence of errors. Theoretical results are complemented by experiments showing that state-of-the-art performances are achieved with considerable computational speed-ups.

We analyze the prediction error of ridge regression in an asymptotic regime where the sample size and dimension go to infinity at a proportional rate. In particular, we consider the role played by the structure of the true regression parameter. We observe that the case of a general deterministic parameter can be reduced to the case of a random parameter from a structured prior. The latter assumption is a natural adaptation of classic smoothness assumptions in nonparametric regression, which are known as source conditions in the the context of regularization theory for inverse problems. Roughly speaking, we assume the large coefficients of the parameter are in correspondence to the principal components. In this setting a precise characterisation of the test error is obtained, depending on the inputs covariance and regression parameter structure. We illustrate this characterisation in a simplified setting to investigate the influence of the true parameter on optimal regularisation for overparameterized models. We show that interpolation (no regularisation) can be optimal even with bounded signal-to-noise ratio (SNR), provided that the parameter coefficients are larger on high-variance directions of the data, corresponding to a more regular function than posited by the regularization term. This contrasts with previous work considering ridge regression with isotropic prior, in which case interpolation is only optimal in the limit of infinite SNR.

We study a natural extension of classical empirical risk minimization, where the hypothesis space is a random subspace of a given space. In particular, we consider possibly data dependent subspaces spanned by a random subset of the data, recovering as a special case Nyström approaches for kernel methods. Considering random subspaces naturally leads to computational savings, but the question is whether the corresponding learning accuracy is degraded. These statistical-computational tradeoffs have been recently explored for the least squares loss and self-concordant loss functions, such as the logistic loss. Here, we work to ex- tend these results to convex Lipschitz loss functions, that might not be smooth, such as the hinge loss used in support vector ma- chines. This extension requires developing new proofs, that use different technical tools. Our main results show the existence of different settings, depending on how hard the learning problem is, for which computational efficiency can be improved with no loss in performance. Theoretical results are illustrated with simple numerical experiments.

Object segmentation is a key component in the visual system of a robot that performs tasks like grasping and object manipulation, especially in presence of occlusions. Like many other computer vision tasks, the adoption of deep architectures has made available algorithms that perform this task with remarkable performance. However, adoption of such algorithms in robotics is hampered by the fact that training requires large amount of computing time and it cannot be performed on-line.In this work, we propose a novel architecture for object segmentation, that overcomes this problem and provides comparable performance in a fraction of the time required by the state-of-the-art methods. Our approach is based on a pre-trained Mask R-CNN, in which various layers have been replaced with a set of classifiers and regressors that are retrained for a new task. We employ an efficient Kernel-based method that allows for fast training on large scale problems. Our approach is validated on the YCB-Video dataset which is widely adopted in the computer vision and robotics community, demonstrating that we can achieve and even surpass performance of the state-of-the-art, with a significant reduction (~6×) of the training time.The code to reproduce the experiments is publicly available on GitHub1.

We introduce ParK, a new large-scale solver for kernel ridge regression. Our approach combines partitioning with random projections and iterative optimization to reduce space and time complexity while provably maintaining the same statistical accuracy. In particular, constructing suitable partitions directly in the feature space rather than in the input space, we promote orthogonality between the local estimators, thus ensuring that key quantities such as local effective dimension and bias remain under control. We characterize the statistical-computational tradeoff of our model, and demonstrate the effectiveness of our method by numerical experiments on large-scale datasets.

Geometric representation learning has shown great promise for important tasks inartificial intelligence and machine learning. However, an open problem is yethow to integrate non-Euclidean representations with standard machine learningmethods.In this work, we consider the task of regression onto hyperbolic space for whichwe propose two approaches: a non-parametric kernel-method for which we also proveexcess risk bounds and a parametric deep learning model that is informed bythe geodesics of the target space.By recasting predictions on trees as manifold regression problems we demonstrate the applications of our approach on two challenging tasks: 1)hierarchical classification via label embeddings and 2) inventing new conceptsby predicting their embedding in a continuous representation of a base taxonomy.In our experiments, we find that the proposed estimators outperform their naivecounterparts that perform regression in the ambient Euclidean space.

In this paper, we analyze a Nyström based approach to efficient large scale kernel principal component analysis (PCA). The latter is a natural nonlinear extension of classical PCA based on considering a nonlinear feature map or the corresponding kernel. Like other kernel approaches, kernel PCA enjoys good mathematical and statistical properties but, numerically, it scales poorly with the sample size. Our analysis shows that Nyström sampling greatly improves computational efficiency without incurring any loss of statistical accuracy. While similar effects have been observed in supervised learning, this is the first such result for PCA. Our theoretical findings are based on a combination of analytic and concentration of measure techniques. Our study is more broadly motivated by the question of understanding the interplay between statistical and computational requirements for learning.

Gaussian processes (GP) are one of the most successful frameworks to model uncertainty. However, GP optimization (e.g., GP-UCB) suffers from major scalability issues. Experimental time grows linearly with the number of evaluations, unless candidates are selected in batches (e.g., using GP-BUCB) and evaluated in parallel. Furthermore, computational cost is often prohibitive since algorithms such as GP-BUCB require a time at least quadratic in the number of dimensions and iterations to select each batch. In this paper, we introduce BBKB (Batch Budgeted Kernel Bandits), the first no-regret GP optimization algorithm that provably runs in near-linear time and selects candidates in batches. This is obtained with a new guarantee for the tracking of the posterior variances that allows BBKB to choose increasingly larger batches, improving over GP-BUCB. Moreover, we show that the same bound can be used to adaptively delay costly updates to the sparse GP approximation used by BBKB, achieving a near-constant per-step amortized cost. These findings are then confirmed in several experiments, where BBKB is much faster than state-of-the-art methods.

We investigate the generalisation performance of Distributed Gradient Descent with implicit regularisation and random features in the homogenous setting where a network of agents are given data sampled independently from the same unknown distribution. Along with reducing the memory footprint, random features are particularly convenient in this setting as they provide a common parameterisation across agents that allows to overcome previous difficulties in implementing decentralised kernel regression. Under standard source and capacity assumptions, we establish high probability bounds on the predictive performance for each agent as a function of the step size, number of iterations, inverse spectral gap of the communication matrix and number of random features. By tuning these parameters, we obtain statistical rates that are minimax optimal with respect to the total number of samples in the network. The algorithm provides a linear improvement over single-machine gradient descent in memory cost and, when agents hold enough data with respect to the network size and inverse spectral gap, a linear speed up in computational run-time for any network topology. We present simulations that show how the number of random features, iterations and samples impact predictive performance.

Kernel methods provide an elegant and principled approach to nonparametric learning, but so far could hardly be used in large scale problems, since naïve implementations scale poorly with data size. Recent advances have shown the benefits of a number of algorithmic ideas, for example combining optimization, numerical linear algebra and random projections. Here, we push these efforts further to develop and test a solver that takes full advantage of GPU hardware. Towards this end, we designed a preconditioned gradient solver for kernel methods exploiting both GPU acceleration and parallelization with multiple GPUs, implementing out-of-core variants of common linear algebra operations to guarantee optimal hardware utilization. Further, we optimize the numerical precision of different operations and maximize efficiency of matrix-vector multiplications. As a result we can experimentally show dramatic speedups on datasets with billions of points, while still guaranteeing state of the art performance. Additionally, we make our software available as an easy to use library.

We propose and analyze a novel theoretical and algorithmic framework for structured prediction. While so far the term has referred to discrete output spaces, here we consider more general settings, such as manifolds or spaces of probability measures. We define structured prediction as a problem where the output space lacks a vectorial structure. We identify and study a large class of loss functions that implicitly defines a suitable geometry on the problem. The latter is the key to develop an algorithmic framework amenable to a sharp statistical analysis and yielding efficient computations. When dealing with output spaces with infinite cardinality, a suitable implicit formulation of the estimator is shown to be crucial.

Gaussian processes (GP) are a stochastic processes, used as Bayesian approach for the optimization of black-box functions. Despite their effectiveness in simple problems, GP-based algorithms hardly scale to high-dimensional functions, as their per-iteration time and space cost is at least \emph{quadratic} in the number of dimensions $d$ and iterations $t$. Given a set of $A$ alternatives to choose from, the overall runtime $O(t^3 A)$ is prohibitive. In this paper, we introduce BKB (\textit{budgeted kernelized bandit}), a new approximate GP algorithm for optimization under bandit feedback that achieves near-optimal regret (and hence near-optimal convergence rate) with near-constant per-iteration complexity and remarkably no assumption on the input space or covariance of the GP. We combine a kernelized linear bandit algorithm (GP-UCB) leverage score sampling as a randomized matrix sketching and prove that selecting inducing points based on their posterior variance gives an accurate low-rank approximation of the GP, preserving variance estimates and confidence intervals. As a consequence, BKB does not suffer from \emph{variance starvation}, an important problem faced by many previous sparse GP approximations. Moreover, we show that our procedure selects at most $\widetilde{O}(d_{eff})$ points, where $d_{eff}$ is the \emph{effective} dimension of the explored space, which is typically much smaller than both $d$ and $t$. This greatly reduces the dimensionality of the problem, thus leading to a $O(T A d_{eff}^2)$ runtime and $O(A d_{eff})$ space complexity.

Imitation learning offers a general framework where robots can efficiently acquire novel motor skills from demonstrations of a human teacher. While many promising achievements have been shown, the majority of them are only focused on single-stroke movements, without taking into account the problem of multi-tasks sequencing. Conceivably, sequencing different atomic tasks can further augment the robot's capabilities as well as avoid repetitive demonstrations. In this paper, we propose to address the issue of multi-tasks sequencing with emphasis on handling the so-called competing constraints, which emerge due to the existence of the concurrent constraints from Cartesian and joint trajectories. Specifically, we explore the null space of the robot from an information-theoretic perspective in order to maintain imitation fidelity during transition between consecutive tasks. The effectiveness of the proposed method is validated through simulated and real experiments on the iCub humanoid robot.

While stochastic gradient descent (SGD) is one of the major workhorses in machine learning, the learning properties of many practically used variants are still poorly understood. In this paper, we consider least squares learning in a nonparametric setting and contribute to filling this gap by focusing on the effect and interplay of multiple passes, mini-batching and averaging, in particular tail averaging. Our results show how these different variants of SGD can be combined to achieve optimal learning rates, also providing practical insights. A novel key result is that tail averaging allows faster convergence rates than uniform averaging in the nonparametric setting. Further, we show that a combination of tail-averaging and minibatching allows more aggressive step-size choices than using any one of said components.

We study learning properties of accelerated gradient descent methods for linear least-squares in Hilbert spaces. We analyze the implicit regularization properties of Nesterov acceleration and a variant of heavy-ball in terms of corresponding learning error bounds. Our results show that acceleration can provides faster bias decay than gradient descent, but also suffers of a more unstable behavior. As a result acceleration cannot be in general expected to improve learning accuracy with respect to gradient descent, but rather to achieve the same accuracy with reduced computations. Our theoretical results are validated by numerical simulations. Our analysis is based on studying suitable polynomials induced by the accelerated dynamics and combining spectral techniques with concentration inequalities.

In this paper, we discuss how a suitable family of tensor kernels can be used to efficiently solve nonparametric extensions of lp regularized learning methods. Our main contribution is proposing a fast dual algorithm, and showing that it allows to solve the problem efficiently. Our results contrast recent findings suggesting kernel methods cannot be extended beyond Hilbert setting. Numerical experiments confirm the effectiveness of the method.

We propose and analyze a variant of the classic Polyak–Ruppert averaging scheme, broadly used in stochastic gradient methods. Rather than a uniform average of the iterates, we consider a weighted average, with weights decaying in a geometric fashion. In the context of linear least-squares regression, we show that this averaging scheme has the same regularizing effect, and indeed is asymptotically equivalent, to ridge regression. In particular, we derive finite-sample bounds for the proposed approach that match the best known results for regularized stochastic gradient methods.

Latest deep learning methods for object detection provide remarkable performance, but have limits when used in robotic applications. One of the most relevant issues is the long training time, which is due to the large size and imbalance of the associated training sets, characterized by few positive and a large number of negative examples (i.e. background). Proposed approaches are based on end-to-end learning by back-propagation [22] or kernel methods trained with Hard Negatives Mining on top of deep features [8]. These solutions are effective, but prohibitively slow for on-line applications. In this paper we propose a novel pipeline for object detection that overcomes this problem and provides comparable performance, with a 60x training speedup. Our pipeline combines (i) the Region Proposal Network and the deep feature extractor from [22] to efficiently select candidate RoIs and encode them into powerful representations, with (ii) the FALKON [23] algorithm, a novel kernel-based method that allows fast training on large scale problems (millions of points). We address the size and imbalance of training data by exploiting the stochastic subsampling intrinsic into the method and a novel, fast, bootstrapping approach. We assess the effectiveness of the approach on a standard Computer Vision dataset (PASCAL VOC 2007 [5]) and demonstrate its applicability to a real robotic scenario with the iCubWorld Transformations [18] dataset.

We investigate the efficiency of k-means in terms of both statistical and computational requirements. More precisely, we study a Nystr\"om approach to kernel k-means. We analyze the statistical properties of the proposed method and show that it achieves the same accuracy of exact kernel k-means with only a fraction of computations. Indeed, we prove under basic assumptions that sampling $\sqrt{n}$ Nystr\"om landmarks allows to greatly reduce computational costs without incurring in any loss of accuracy. To the best of our knowledge this is the first result showing in this kind for unsupervised learning.

Sketching and stochastic gradient methods are arguably the most common techniques to derive efficient large scale learning algorithms. In this paper, we investigate their application in the context of nonparametric statistical learning. More precisely, we study the estimator defined by stochastic gradient with mini batches and random features. The latter can be seen as form of nonlinear sketching and used to define approximate kernel methods. The considered estimator is not explicitly penalized/constrained and regularization is implicit. Indeed, our study highlights how different parameters, such as number of features, iterations, step-size and mini-batch size control the learning properties of the solutions. We do this by deriving optimal finite sample bounds, under standard assumptions. The obtained results are corroborated and illustrated by numerical experiments.

This paper studies the problem of deriving fast and accurate classification algorithms with uncertainty quantification. Gaussian process classification provides a principled approach, but the corresponding computational burden is hardly sustainable in large-scale problems and devising efficient alternatives is a challenge. In this work, we investigate if and how Gaussian process regression directly applied to classification labels can be used to tackle this question. While in this case training is remarkably faster, predictions need to be calibrated for classification and uncertainty estimation. To this aim, we propose a novel regression approach where the labels are obtained through the interpretation of classification labels as the coefficients of a degenerate Dirichlet distribution. Extensive experimental results show that the proposed approach provides essentially the same accuracy and uncertainty quantification as Gaussian process classification while requiring only a fraction of computational resources.

Leverage score sampling provides an appealing way to perform approximate com- putations for large matrices. Indeed, it allows to derive faithful approximations with a complexity adapted to the problem at hand. Yet, performing leverage scores sampling is a challenge in its own right requiring further approximations. In this paper, we study the problem of leverage score sampling for positive definite ma- trices defined by a kernel. Our contribution is twofold. First we provide a novel algorithm for leverage score sampling and second, we exploit the proposed method in statistical learning by deriving a novel solver for kernel ridge regression. Our main technical contribution is showing that the proposed algorithms are currently the most efficient and accurate for these problems.

Structured prediction provides a general framework to deal with supervised problems where the outputs have semantically rich structure. While classical approaches consider finite, albeit potentially huge, output spaces, in this paper we discuss how structured prediction can be extended to a continuous scenario. Specifically, we study a structured prediction approach to manifold-valued regression. We characterize a class of problems for which the considered approach is statistically consistent and study how geometric optimization can be used to compute the corresponding estimator. Promising experimental results on both simulated and real data complete our study.

We consider object recognition in the context of lifelong learning, where a robotic agent learns to discriminate between a growing number of object classes as it accumulates experience about the environment. We propose an incremental variant of the Regularized Least Squares for Classification (RLSC) algorithm, and exploit its structure to seamlessly add new classes to the learned model. The presented algorithm addresses the problem of having an unbalanced proportion of training examples per class, which occurs when new objects are presented to the system for the first time. We evaluate our algorithm on both a machine learning benchmark dataset and two challenging object recognition tasks in a robotic setting. Empirical evidence shows that our approach achieves comparable or higher classification performance than its batch counterpart when classes are unbalanced, while being significantly faster.

Key to multitask learning is exploiting the relationships between different tasks to improve prediction performance. Most previous methods have focused on the case where tasks relations can be modeled as linear operators and regularization approaches can be used successfully. However, in practice assuming the tasks to be linearly related is often restrictive, and allowing for nonlinear structures is a challenge. In this paper, we tackle this issue by casting the problem within the framework of structured prediction. Our main contribution is a novel algorithm for learning multiple tasks which are related by a system of nonlinear equations that their joint outputs need to satisfy. We show that our algorithm can be efficiently implemented and study its generalization properties, proving universal consistency and learning rates. Our theoretical analysis highlights the benefits of non-linear multitask learning over learning the tasks independently. Encouraging experimental results show the benefits of the proposed method in practice.

Kernel methods provide a principled way to perform non linear, nonparametric learning. They rely on solid functional analytic foundations and enjoy optimal statistical properties. However, at least in their basic form, they have limited applicability in large scale scenarios because of stringent computational requirements in terms of time and especially memory. In this paper, we take a substantial step in scaling up kernel methods, proposing FALKON, a novel algorithm that allows to efficiently process millions of points. FALKON is derived combining several algorithmic principles, namely stochastic subsampling, iterative solvers and preconditioning. Our theoretical analysis shows that optimal statistical accuracy is achieved requiring essentially $O(n)$ memory and $O(n\sqrt{n})$ time. An extensive experimental analysis on large scale datasets shows that, even with a single machine, FALKON outperforms previous state of the art solutions, which exploit parallel/distributed architectures.

We study the generalization properties of ridge regression with random features in the statistical learning framework. We show for the first time that $O(1/\sqrt{n})$ learning bounds can be achieved with only $O(\sqrt{n}\log n)$ random features rather than $O({n})$ as suggested by previous results. Further, we prove faster learning rates and show that they might require more random features, unless they are sampled according to a possibly problem dependent distribution. Our results shed light on the statistical computational trade-offs in large scale kernelized learning, showing the potential effectiveness of random features in reducing the computational complexity while keeping optimal generalization properties.

We analyze the learning properties of the stochastic gradient method when multiple passes over the data and mini-batches are allowed. We study how regularization properties are controlled by the step-size, the number of passes and the mini-batch size. In particular, we consider the square loss and show that for a universal step-size choice, the number of passes acts as a regularization parameter, and optimal finite sample bounds can be achieved by early-stopping. Moreover, we show that larger step-sizes are allowed when considering mini-batches. Our analysis is based on a unifying approach, encompassing both batch and stochastic gradient methods as special cases. As a byproduct, we derive optimal convergence results for batch gradient methods (even in the non-attainable cases).

Learning embeddings of entities and relations is an efficient and versatile method to perform machine learning on relational data such as knowledge graphs. In this work, we propose holographic embeddings (HolE) to learn compositional vector space representations of entire knowledge graphs. The proposed method is related to holographic models of associative memory in that it employs circular correlation to create compositional representations. By using correlation as the compositional operator, HolE can capture rich interactions but simultaneously remains efficient to compute, easy to train, and scalable to very large datasets. Experimentally, we show that holographic embeddings are able to outperform state-of-the-art methods for link prediction on knowledge graphs and relational learning benchmark datasets.

Early stopping is a well known approach to reduce the time complexity for performing training and model selection of large scale learning machines. On the other hand, memory/space (rather than time) complexity is the main constraint in many applications, and randomized subsampling techniques have been proposed to tackle this issue. In this paper we ask whether early stopping and subsampling ideas can be combined in a fruitful way. We consider the question in a least squares regression setting and propose a form of randomized iterative regularization based on early stopping and subsampling. In this context, we analyze the statistical and computational properties of the proposed method. Theoretical results are complemented and validated by a thorough experimental analysis.

We study the generalization properties of stochastic gradient methods for learning with convex loss functions and linearly parameterized functions. We show that, in the absence of penalizations or constraints, the stability and approximation properties of the algorithm can be controlled by tuning either the step-size or the number of passes over the data. In this view, these parameters can be seen to control a form of implicit regularization. Numerical results complement the theoretical findings.

This paper presents a novel approach for incremental semiparametric inverse dynamics learning. In particular, we consider the mixture of two approaches: Parametric modeling based on rigid body dynamics equations and nonparametric modeling based on incremental kernel methods, with no prior information on the mechanical properties of the system. The result is an incremental semiparametric approach, leveraging the advantages of both the parametric and nonparametric models. We validate the proposed technique learning the dynamics of one arm of the iCub humanoid robot.

The development of reliable and robust visual recognition systems is a main challenge towards the deployment of autonomous robotic agents in unconstrained environments. Learning to recognize objects requires image representations that are discriminative to relevant information while being invariant to nuisances, such as scaling, rotations, light and background changes, and so forth. Deep Convolutional Neural Networks can learn such representations from large web-collected image datasets and a natural question is how these systems can be best adapted to the robotics context where little supervision is often available. In this work, we investigate different training strategies for deep architectures on a new dataset collected in a real-world robotic setting. In particular we show how deep networks can be tuned to improve invariance and discriminability properties and perform object identification tasks with minimal supervision.

We propose and analyze a regularization approach for structured prediction problems. We characterize a large class of loss functions that allows to naturally embed structured outputs in a linear space. We exploit this fact to design learning algorithms using a surrogate loss approach and regularization techniques. We prove universal consistency and finite sample bounds characterizing the generalization properties of the proposed method. Experimental results are provided to demonstrate the practical usefulness of the proposed approach.

We analyze the learning properties of the stochastic gradient method when multiple passes over the data and mini-batches are allowed. In particular, we consider the square loss and show that for a universal step-size choice, the number of passes acts as a regularization parameter, and optimal finite sample bounds can be achieved by early-stopping. Moreover, we show that larger step-sizes are allowed when considering mini-batches. Our analysis is based on a unifying approach, encompassing both batch and stochastic gradient methods as special cases.

We consider the problem of supervised learning with convex loss functions and propose a new form of iterative regularization based on the subgradient method. Unlike other regularization approaches, in iterative regularization no constraint or penalization is considered, and generalization is achieved by (early) stopping an empirical iteration. We consider a nonparametric setting, in the framework of reproducing kernel Hilbert spaces, and prove consistency and finite sample bounds on the excess risk under general regularity conditions. Our study provides a new class of efficient regularized learning algorithms and gives insights on the interplay between statistics and optimization in machine learning.

Multi-task learning is a natural approach for computer vision applications that require the simultaneous solution of several distinct but related problems, e.g. object detection, classification, tracking of multiple agents, or denoising, to name a few. The key idea is that exploring task relatedness (structure) can lead to improved performances. In this paper, we propose and study a novel sparse, nonparametric approach exploiting the theory of Reproducing Kernel Hilbert Spaces for vector-valued functions. We develop a suitable regularization framework which can be formulated as a convex optimization problem, and is provably solvable using an alternating minimization approach. Empirical tests show that the proposed method compares favorably to state of the art techniques and further allows to recover interpretable structures, a problem of interest in its own right.

Reducing the amount of human supervision is a key problem in machine learning and a natural approach is that of exploiting the relations (structure) among different tasks. This is the idea at the core of multi-task learning. In this context a fundamental question is how to incorporate the tasks structure in the learning problem. We tackle this question by studying a general computational framework that allows to encode a-priori knowledge of the tasks structure in the form of a convex penalty; in this setting a variety of previously proposed methods can be recovered as special cases, including linear and non-linear approaches. Within this framework, we show that tasks and their structure can be efficiently learned considering a convex optimization problem that can be approached by means of block coordinate methods such as alternating minimization and for which we prove convergence to the global minimum.

In this paper we address the problem of unsupervised learning of discrete subword units. Our approach is based on Deep Autoencoders (AEs), whose encoding node values are thresholded to subsequently generate a symbolic, i.e., 1-of-K (with K = No. of subwords), representation of each speech frame. We experiment with two variants of the standard AE which we have named Binarized Autoencoder and Hidden-Markov-Model Encoder. The first forces the binary encoding nodes to have a Ushaped distribution (with peaks at 0 and 1) while minimizing the reconstruction error. The latter jointly learns the symbolic encoding representation (i.e., subwords) and the prior and transition distribution probabilities of the learned subwords. The ABX evaluation of the Zero Resource Challenge - Track 1 shows that a deep AE with only 6 encoding nodes, which assigns to each frame a 1-of-K binary vector with K = 26, can outperform real-valued MFCC representations in the acrossspeaker setting. Binarized AEs can outperform standard AEs when using a larger number of encoding nodes, while HMM Encoders may allow more compact subword transcriptions without worsening the ABX performance.

In the framework of a theory for invariant sensory signal representations, a signature which is invariant and selective for speech sounds can be obtained through projections in template signals and pooling over their transformations under a group. For locally compact groups, e.g., translations, the theory explains the resilience of convolutional neural networks with filter weight sharing and max pooling across their local translations in frequency or time. In this paper we propose a discriminative approach for learning an optimum set of templates, under a family of transformations, namely frequency transpositions and perturbations of the vocal tract length, which are among the primary sources of speech variability. Implicitly, we generalize convolutional networks to transformations other than translations, and derive data-specific templates by training a deep network with convolution-pooling layers and densely connected layers. We demonstrate that such a representation, combining group-generalized convolutions, theoretical invariance guarantees and discriminative template selection, improves frame classification performance over standard translation-CNNs and DNNs on TIMIT and Wall Street Journal datasets.

Within a statistical learning setting, we propose and study an iterative regularization algorithm for least squares defined by an incremental gradient method. In particular, we show that, if all other parameters are fixed a priori, the number of passes over the data (epochs) acts as a regularization parameter, and prove strong universal consistency, i.e. almost sure convergence of the risk, as well as sharp finite sample bounds for the iterates. Our results are a step towards understanding the effect of multiple epochs in stochastic gradient techniques in machine learning and rely on integrating statistical and optimizationresults.

We study Nyström type subsampling approaches to large scale kernel methods, and prove learning bounds in the statistical learning setting, where random sampling and high probability estimates are considered. In particular, we prove that these approaches can achieve optimal learning bounds, provided the subsampling level is suitably chosen. These results suggest a simple incremental variant of Nyström kernel ridge regression, where the subsampling level controls at the same time regularization and computations. Extensive experimental analysis shows that the considered approach achieves state of the art performances on benchmark large scale datasets.

Representations in the auditory cortex might be based on mechanisms similar to the visual ventral stream; modules for building invariance to transformations and multiple layers for compositionality and selectivity. In this paper we propose the use of such computational modules for extracting invariant and discriminative audio representations. Building on a theory of invariance in hierarchical architectures, we propose a novel, mid-level representation for acoustical signals, using the empirical distributions of projections on a set of templates and their transformations. Under the assumption that, by construction, this dictionary of templates is composed from similar classes, and samples the orbit of variance-inducing signal transformations (such as shift and scale), the resulting signature is theoretically guaranteed to be unique, invariant to transformations and stable to deformations. Modules of projection and pooling can then constitute layers of deep networks, for learning composite representations. We present the main theoretical and computational aspects of a framework for unsupervised learning of invariant audio representations, empirically evaluated on music genre classification.

Extracting discriminant, transformation-invariant features from raw audio signals remains a serious challenge for speech recognition. The issue of speaker variability is central to this problem, as changes in accent, dialect, gender, and age alter the sound waveform of speech units at multiple levels (phonemes, words, or phrases). Approaches for dealing with this variability have typically focused on analyzing the spectral properties of speech at the level of frames, on par with frame-level acoustic modeling usually applied to speech recognition systems. In this paper, we propose a framework for representing speech at the word level and extracting features from the acoustic, temporal domain, without the need for spectral encoding or preprocessing. Leveraging recent work on unsupervised learning of invariant sensory representations, we extract a signature for a word by first projecting its raw waveform onto a set of templates and their transformations, and then forming empirical estimates of the resulting one-dimensional distributions via histograms. The representation and relevant parameters are evaluated for word classification on a series of datasets with increasing speakermismatch difficulty, and the results are compared to those of an MFCC-based representation. Index Terms: invariance, acoustic features, speech representation, word classification

We propose a multi-layer feature extraction framework for speech, capable of providing invariant representations. A set of templates is generated by sampling the result of applying smooth, identity-preserving transformations (such as vocal tract length and tempo variations) to arbitrarily-selected speech signals. Templates are then stored as the weights of “neurons”. We use a cascade of such computational modules to factor out different types of transformation variability in a hierarchy, and show that it improves phone classification over baseline features. In addition, we describe empirical comparisons of a) different transformations which may be responsible for the variability in speech signals and of b) different ways of assembling template sets for training. The proposed layered system is an effort towards explaining the performance of recent deep learning networks and the principles by which the human auditory cortex might reduce the sample complexity of learning in speech recognition. Our theory and experiments suggest that invariant representations are crucial in learning from complex, real-world data like natural speech. Our model is built on basic computational primitives of cortical neurons, thus making an argument about how representations might be learned in the human auditory cortex. Index Terms: Invariance, Auditory Cortex, Phonetic Classification, Convolutional Network

In this paper we tackle the problem of estimating the local compliance of tactile arrays exploiting global measurements from a single force and torque sensor. The proposed procedure exploits a transformation matrix (describing the relative position between the local tactile elements and the global force/torque measurements) to define a linear regression problem on the unknown local stiffness. Experiments have been conducted on the foot of the iCub robot, sensorized with a single force/torque sensor and a tactile array of 250 tactile elements (taxels) on the foot sole. Results show that a simple calibration procedure can be employed to estimate the stiffness parameters of virtual springs over a tactile array and to use these model to predict normal forces exerted on the array based only on the tactile feedback. Leveraging on previous works [1] the proposed procedure does not necessarily need a-priori information on the transformation matrix of the taxels which can be directly estimated from available measurements.

In this paper we show that the problem of phase retrieval can be efficiently and provably solved via an alternating minimization algorithm suitably initialized. Our initialization is based on One Bit Phase Retrieval that we introduced in [1], where we showed that O(n log(n)) Gaussian phase-less measurements ensure robust recovery of the phase. In this paper we improve the sample complexity bound to O(n) measurements for sufficiently large n, using a variant of Matrix Bernstein concentration inequality that exploits the intrinsic dimension, together with properties of one bit phase retrieval.

Recent developments in learning sophisticated, hierarchical image representations have led to remarkable progress in the context of visual recognition. While these methods are becoming standard in modern computer vision systems, they are rarely adopted in robotics. The question arises of whether solutions, which have been primarily developed for image retrieval, can perform well in more dynamic and unstructured scenarios. In this paper we tackle this question performing an extensive evaluation of state of the art methods for visual recognition on a iCub robot. We consider the problem of classifying 15 different objects shown by a human demonstrator in a challenging Human-Robot Interaction scenario. The classification performance of hierarchical learning approaches are shown to outperform benchmark solutions based on local descriptors and template matching. Our results show that hierarchical learning systems are computationally efficient and can be used for real-time training and recognition of objects.

A large number of algorithms in machine learning, from principal component analysis (PCA), and its non-linear (kernel) extensions, to more recent spectral embedding and support estimation methods, rely on estimating a linear subspace from samples. In this paper we introduce a general formulation of this problem and derive novel learning error estimates. Our results rely on natural assumptions on the spectral properties of the covariance operator associated to the data distribution, and hold for a wide class of metrics between subspaces. As special cases, we discuss sharp error estimates for the reconstruction properties of PCA and spectral support estimation. Key to our analysis is an operator theoretic approach that has broad applicability to spectral learning methods.

We study the problem of estimating a manifold from random samples. In particular, we consider piecewise constant and piecewise linear estimators induced by k-means and k-flats, and analyze their performance. We extend previous results for k-means in two separate directions. First, we provide new results for k-means reconstruction on manifolds and, secondly, we prove reconstruction bounds for higher-order approximation (k-flats), for which no known results were previously available. While the results for k-means are novel, some of the technical tools are well-established in the literature. In the case of k-flats, both the results and the mathematical tools are new.

We study the problem of estimating, in the sense of optimal transport metrics, a measure which is assumed supported on a manifold embedded in a Hilbert space. By establishing a precise connection between optimal transport metrics, optimal quantization, and learning theory, we derive new probabilistic bounds for the performance of a classic algorithm in unsupervised learning (k-means), when used to produce a probability measure derived from the data. In the course of the analysis, we arrive at new lower bounds, as well as probabilistic bounds on the convergence rate of the empirical law of large numbers, which, unlike existing bounds, are applicable to a wide class of measures.

In this paper we dicuss a novel framework for multiclass learning, defined by a suitable coding/decoding strategy, namely the simplex coding, that allows to generalize to multiple classes a relaxation approach commonly used in binary classification. In this framework a relaxation error analysis can be developed avoiding constraints on the considered hypotheses class. Moreover, we show that in this setting it is possible to derive the first provably consistent regularized methods with training/tuning complexity which is {\em independent} to the number of classes. Tools from convex analysis are introduced that can be used beyond the scope of this paper.

We deal with the problem of variable selection when variables must be selected group-wise, with possibly overlapping groups defined a priori. In particular we propose a new optimization procedure for solving the regularized algorithm presented in Jacob et al. 09, where the group lasso penalty is generalized to overlapping groups of variables. While in Jacob et al. 09 the proposed implementation requires explicit replication of the variables belonging to more than one group, our iterative procedure is based on a combination of proximal methods in the primal space and constrained Newton method in a reduced dual space, corresponding to the active groups. This procedure provides a scalable alternative with no need for data duplication, and allows to deal with high dimensional problems without pre-processing to reduce the dimensionality of the data. The computational advantages of our scheme with respect to state-of-the-art algorithms using data duplication are shown empirically with numerical simulations.

In this paper we consider the problem of learning from data the support of a probability distribution when the distribution {\em does not} have a density (with respect to some reference measure). We propose a new class of regularized spectral estimators based on a new notion of reproducing kernel Hilbert space, which we call {\em ``completely regular''}. Completely regular kernels allow to capture the relevant geometric and topological properties of an arbitrary probability space. In particular, they are the key ingredient to prove the universal consistency of the spectral estimators and in this respect they are the analogue of universal kernels for supervised problems. Numerical experiments show that spectral estimators compare favorably to state of the art machine learning algorithms for density support estimation.

A goal of central importance in the study of hierarchical models for object recognition -- and indeed the visual cortex -- is that of understanding quantitatively the trade-off between invariance and selectivity, and how invariance and discrimination properties contribute towards providing an improved representation useful for learning from data. In this work we provide a general group-theoretic framework for characterizing and understanding invariance in a family of hierarchical models. We show that by taking an algebraic perspective, one can provide a concise set of conditions which must be met to establish invariance, as well as a constructive prescription for meeting those conditions. Analyses in specific cases of particular relevance to computer vision and text processing are given, yielding insight into how and when invariance can be achieved. We find that the minimal sets of transformations intrinsic to the hierarchical model needed to support a particular invariance can be clearly described, thereby encouraging efficient computational implementations.

We discuss how a large class of regularization methods, collectively known as spectral regularization and originally designed for solving ill-posed inverse problems, gives rise to regularized learning algorithms. All of these algorithms are consistent kernel methods that can be easily implemented. The intuition behind their derivation is that the same principle allowing for the numerical stabilization of a matrix inversion problem is crucial to avoid overfitting. The various methods have a common derivation but different computational and theoretical properties. We describe examples of such algorithms, analyze their classification performance on several data sets and discuss their applicability to real-world problems.

In this paper we investigate the regularization property of Kernel Principal Component Analysis (KPCA), by studying its application as a preprocessing step to supervised learning problems. We show that performing KPCA and then ordinary least squares on the projected data, a procedure known as kernel principal component regression (KPCR), is equivalent to spectral cut-off regularization, the regularization parameter being exactly the number of principal components to keep. Using probabilistic estimates for integral operators we can prove error estimates for KPCR and propose a parameter choice procedure allowing to prove consistency of the algorithm.

Many works related learning from examples to regularization techniques for inverse problems, emphasizing the strong algorithmic and conceptual analogy of certain learning algorithms with regularization algorithms. In particular it is well known that regularization schemes such as Tikhonov regularization can be effectively used in the context of learning and are closely related to algorithms such as support vector machines. Nevertheless the connection with inverse problem was considered only for the discrete (finite sample) problem and the probabilistic aspects of learning from examples were not taken into account. In this paper we provide a natural extension of such analysis to the continuous (population) case and study the interplay between the discrete and continuous problems. From a theoretical point of view, this allows to draw a clear connection between the consistency approach in learning theory and the stability convergence property in ill-posed inverse problems. The main mathematical result of the paper is a new probabilistic bound for the regularized least-squares algorithm. By means of standard results on the approximation term, the consistency of the algorithm easily follows.

Many works have shown that strong connections relate learning from ex- amples to regularization techniques for ill-posed inverse problems. Nev- ertheless by now there was no formal evidence neither that learning from examples could be seen as an inverse problem nor that theoretical results in learning theory could be independently derived using tools from reg- ularization theory. In this paper we provide a positive answer to both questions. Indeed, considering the square loss, we translate the learning problem in the language of regularization theory and show that consis- tency results and optimal regularization parameter choice can be derived by the discretization of the corresponding inverse problem. 1 Introduction The main goal of learning from examples is to infer an estimator, given a finite sample of data drawn according to a fixed but unknown probabilistic input-output relation. The desired property of the selected estimator is to perform well on new data, i.e. it should gen- eralize. The fundamental works of Vapnik and further developments [16], [8], [5], show that the key to obtain a meaningful solution to the above problem is to control the complex- ity of the solution space. Interestingly, as noted by [12], [8], [2], this is the idea underlying regularization techniques for ill-posed inverse problems [15], [7]. In such a context to avoid undesired oscillating behavior of the solution we have to restrict the solution space. Not surprisingly the form of the algorithms proposed in both theories is strikingly similar. Anyway a careful analysis shows that a rigorous connection between learning and regular- ization for inverse problem is not straightforward. In this paper we consider the square loss and show that the problem of learning can be translated into a convenient inverse problem and consistency results can be derived in a general setting. When a generic loss is consid- ered the analysis becomes immediately more complicated. Some previous works on this subject considered the special case in which the elements of the input space are fixed and not probabilistically drawn [11], [9]. Some weaker results in the same spirit of those presented in this paper can be found in [13] where anyway the connections with inverse problems is not discussed. Finally, our analysis is close to the idea of stochastic inverse problems discussed in [16]. It follows the plan of the paper. Af- ter recalling the main concepts and notation of learning and inverse problems, in section 4 we develop a formal connection between the two theories. In section 5 the main results are stated and discussed. Finally in section 6 we conclude with some remarks and open problems. 2 Learning from examples We briefly recall some basic concepts of learning theory [16], [8]. In the framework of learning, there are two sets of variables: the input space X, compact subset of Rn, and the output space Y , compact subset of R. The relation between the input x X and the output y Y is described by a probability distribution (x, y) = (x)(y|x) on X Y . The distribution is known only through a sample z = (x, y) = ((x1, y1), . . . , (x , y )), called training set, drawn i.i.d. according to . The goal of learning is, given the sample z, to find a function fz : X R such that fz(x) is an estimate of the output y when the new input x is given. The function fz is called estimator and the rule that, given a sample z, provides us with fz is called learning algorithm. Given a measurable function f : X R, the ability of f to describe the distribution is measured by its expected risk defined as I[f ] = (f (x) - y)2 d(x,y). XY The regression function g(x) = y d(y|x), Y is the minimizer of the expected risk over the set of all measurable functions and always exists since Y is compact. Usually, the regression function cannot be reconstructed exactly since we are given only a finite, possibly small, set of examples z. To overcome this problem, in the regularized least squares algorithm an hypothesis space H is fixed, and, given > 0, an estimator f z is defined as the solution of the regularized least squares problem, 1 min{ (f (xi) - yi)2 + f 2H}. (1) f H i=1 The regularization parameter has to be chosen depending on the available data, = ( , z), in such a way that, for every > 0 lim P I[f ( ,z) z ] - inf I[f] = 0. (2) + f H We note that in general inffH I[f ] is larger that I[g] and represents a sort of irreducible error associated with the choice of the space H. The above convergence in probability is usually called consistency of the algorithm [16] [14]. 3 Ill-Posed Inverse Problems and Regularization In this section we give a very brief account of linear inverse problems and regularization theory [15], [7]. Let H and K be two Hilbert spaces and A : H K a linear bounded operator. Consider the equation Af = g (3) where g, g K and g - g K . Here g represents the exact, unknown data and g the available, noisy data. Finding the function f satisfying the above equation, given A and g, is the linear inverse problem associated to Eq. (3). The above problem is, in general, ill- posed, that is, the Uniqueness can be restored introducing the Moore-Penrose generalized inverse f = Ag defined as the minimum norm solution of the problem min Af - g 2 . (4) K f H However the operator A is usually not bounded so, in order to ensure a continuous de- pendence of the solution on the data, the following Tikhonov regularization scheme can be considered1 min{ Af - g 2 + f 2 K H}, (5) f H whose unique minimizer is given by f = (AA + I )-1Ag , (6) where A denotes the adjoint of A. A crucial step in the above algorithm is the choice of the regularization parameter = (, g), as a function of the noise level and the data g, in such a way that lim f (,g) = 0, (7) - f 0 H that is, the regularized solution f (,g) converges to the generalized solution f = Ag (f exists if and only if P g Range(A), where P is the projection on the closure of the range of A and, in that case, Af = P g) when the noise goes to zero. The similarity between regularized least squares algorithm (1) and Tikhonov regulariza- tion (5) is apparent. However, several difficulties emerge. First, to treat the problem of learning in the setting of ill-posed inverse problems we have to define a direct problem by means of a suitable operator A. Second, in the context of learning, it is not clear the nature of the noise . Finally we have to clarify the relation between consistency (2) and the kind of convergence expressed by (7). In the following sections we will show a possible way to tackle these problems. 4 Learning as an Inverse Problem We can now show how the problem of learning can be rephrased in a framework close to the one presented in the previous section. We assume that hypothesis space H is a reproducing kernel Hilbert space [1] with a contin- uous kernel K : X X R. If x X, we let Kx(s) = K(s, x), and, if is the marginal distribution of on X, we define the bounded linear operator A : H L2(X, ) as (Af )(x) = f, Kx = f (x), H 1In the framework of inverse problems, many other regularization procedures are introduced [7]. For simplicity we only treat the Tikhonov regularization. that is, A is the canonical injection of H in L2(X, ). In particular, for all f H, the expected risk becomes, I[f ] = Af - g 2 + I[g], L2(X,) where g is the regression function [2]. The above equation clarifies that if the expected risk admits a minimizer fH on the hypothesis space H, then it is exactly the generalized solution2 f = Ag of the problem Af = g. (8) Moreover, given a training set z = (x, y), we get a discretized version Ax : H E of A, that is (Axf)i = f, Kx = f (x i H i), where E = R is the finite dimensional euclidean space endowed with the scalar product 1 y, y = y E iyi. i=1 It is straightforward to check that 1 (f (xi) - yi)2 = Axf - y 2 , E i=1 so that the estimator f z given by the regularized least squares algorithm is the regularized solution of the discrete problem Axf = y. (9) At this point it is useful to remark the following two facts. First, in learning from examples we are not interested into finding an approximation of the generalized solution of the dis- cretized problem (9), but we want to find a stable approximation of the solution of the exact problem (8) (compare with [9]). Second, we notice that in learning theory the consistency property (2) involves the control of the quantity I[f z ] - inf I[f] = Af - g 2 Af . (10) L2(X,) - inf - g 2L2(X,) f H f H If P is the projection on the closure of the range of A, the definition of P gives 2 I[f z ] - inf I[f] = Afz - P g (11) f H L2(X,) (the above equality stronlgy depends on the fact that the loss function is the square loss). In the inverse problem setting, the square root of the above quantity is called the residue of the solution f z . Hence, consistency is controlled by the residue of the estimator, instead of the reconstruction error f z - f (as in inverse problems). In particular, consistency H is a weaker condition than the one required by (7) and does not require the existence of the generalized solution fH. 5 Regularization, Stochastic Noise and Consistency To apply the framework of ill-posed inverse problems of Section 3 to the formulation of learning proposed above, we note that the operator Ax in the discretized problem (9) differs from the operator A in the exact problem (8) and a measure of the difference between Ax and A is required. Moreover, the noisy data y E and the exact data g L2(X, ) belong to different spaces, so that the notion of noise has to be modified. Given the above premise our derivation of consistency results is developed in two steps: we first study the residue of the solution by means of a measure of the noise due to discretization and then we show a possible way to give a probabilistic evaluation of the noise previously introduced. 2The fact that fH is the minimal norm solution of (4) is ensured by the assumption that the support of the measure is X, since in this case the operator A is injective. 5.1 Bounding the Residue of the Regularized Solution We recall that the regularized solutions of problems (9) and (8) are given by f = (A A y, z x x + I )-1A x f = (AA + I)-1Ag. The above equations show that f and f depend only on A A z x x and AA which are operators from H into H and on Ay and Ag which are elements of x H, so that the space E disappears. This observation suggests that noise levels could be A A x x - AA L(H) and A y , where is the uniform operator norm. To this purpose, for x - Ag H L(H) every = (1, 2) R2+ we define the collection of training sets. U = {z (X Y ) | Ay A x - Ag H 1, Ax x - AA L(H) 2, N} and we let M = sup{|y| | y Y }. The next theorem is the central result of the paper. Theorem 1 If > 0, the following inequalities hold 1. for any training set z U M Af 2 + 1 z - P g L2(X,) - Af - P g L2(X,) 4 2 2. if P g Range(A), for any training set z U, M f 2 + 1 z - f H - f - f H 3 2 2 Moreover if we choose = (, z) in such a way that lim0 sup (,z) = 0 zU 2 lim 1 0 sup = 0 zU (12) (,z) then lim 2 0 sup = 0 zU (,z) lim sup Af (,z) = 0. (13) z - Pg 0 zU L2(X,) We omit the complete proof and refer to [3]. Briefly, the idea is to note that Af z - P g L2(X,) - Af - P g L2(X,) 1 Af = (AA) 2 (f z - Af L2(X,) z - f) H where the last equation follows by polar decomposition of the operator A. Moreover a simple algebraic computation gives f A A y+(AA+I)-1(A y z -f = (AA+I)-1(AA-Ax x)(Ax x+I)-1Ax x -Ag) where the relevant quantities for definition of the noise appear. The first item in the above proposition quantifies the difference between the residues of the regularized solutions of the exact and discretized problems in terms of the noise level = (1, 2). As mentioned before this is exactly the kind of result needed to derive consistency. On the other hand the last part of the proposition gives sufficient conditions on the parameter to ensure convergence of the residue to zero as the level noise decreases. The above results were obtained introducing the collection U of training sets compatible with a certain noise level . It is left to quantify the noise level corresponding to a training set of cardinality . This will be achieved in a probabilistic setting in the next section. 5.2 Stochastic Evaluation of the Noise In this section we estimate the discretization noise = (1, 2). Theorem 2 Let 1, 2 > 0 and = supxX K(x, x), then M 2 P Ag - A x y A H + 1, AA - Ax x L(H) + 2 2 2 1 2 1 - e-22M2 - e-24 (14) The proof is given in [3] and it is based on McDiarmid inequality [10] applied to the random variables F (z) = A x y - Ag G(z) = A A . H x x - AA L(H) Other estimates of the noise can be given using, for example, union bounds and Hoeffd- ing's inequality. Anyway rather then providing a tight analysis our concern was to find an natural, explicit and easy to prove estimate of . 5.3 Consistency and Regularization Parameter Choice Combining Theorems 1 and 2, we easily derive the following corollary. Corollary 1 Given 0 < < 1, with probability greater that 1 - , Af z - Pg - Af - Pg L2(X,) L2(X,) M 1 4 + 1 + log (15) 2 2

In this letter, we investigate the impact of choosing different loss functions from the viewpoint of statistical learning theory. We introduce a convexity assumption, which is met by all loss functions commonly used in the literature, and study how the bound on the estimation error changes with the loss. We also derive a general result on the minimizer of the expected risk for a convex loss function in the case of classification. The main outcome of our analysis is that for classification, the hinge loss appears to be the loss of choice. Other things being equal, the hinge loss leads to a convergence rate practically indistinguishable from the logistic loss rate and much better than the square loss rate. Furthermore, if the hypothesis space is sufficiently rich, the bounds obtained for the hinge loss are not loosened by the thresholding stage.