This paper presents the machine learning architecture of the Snips Voice Platform, a software solution to perform Spoken Language Understanding on microprocessors typical of IoT devices. The embedded inference is fast and accurate while enforcing privacy by design, as no personal user data is ever collected. Focusing on Automatic Speech Recognition and Natural Language Understanding, we detail our approach to training high-performance Machine Learning models that are small enough to run in real-time on small devices. Additionally, we describe a data generation procedure that provides sufficient, high-quality training data without compromising user privacy.1 https://www.voicebot.ai/2018/03/07/new-voicebot-report-says-nearly-20-u-sadults-smart-speakers/ 2 In French: https://www.cnil.fr/fr/enceintes-intelligentes-des-assistants-vocauxconnectes-votre-vie-privee 3 https://www.eugdpr.org/

We study the stability of amorphous solids, focussing on the distribution P (x) of the local stress increase x that would lead to an instability. We argue that this distribution behaves as P (x) ∼ x θ , where the exponent θ is larger than zero if the elastic interaction between rearranging regions is non-monotonic, and increases with the interaction range. For a class of finite-dimensional models we show that stability implies a lower bound on θ, which is found to lie near saturation. For quadrupolar interactions these models yield θ ≈ 0.6 for d = 2 and θ ≈ 0.4 in d = 3 where d is the spatial dimension, accurately capturing previously unresolved observations in atomistic models, both in quasi-static flow and after a fast quench. In addition, we compute the Herschel-Buckley exponent in these models and show that it depends on a subtle choice of dynamical rules, whereas the exponent θ does not.

Random projections have proven extremely useful in many signal processing and machine learning applications. However, they often require either to store a very large random matrix, or to use a different, structured matrix to reduce the computational and memory costs. Here, we overcome this difficulty by proposing an analog, optical device, that performs the random projections literally at the speed of light without having to store any matrix in memory. This is achieved using the physical properties of multiple coherent scattering of coherent light in random media. We use this device on a simple task of classification with a kernel machine, and we show that, on the MNIST database, the experimental results closely match the theoretical performance of the corresponding kernel. This framework can help make kernel methods practical for applications that have large training sets and/or require real-time prediction. We discuss possible extensions of the method in terms of a class of kernels, speed, memory consumption and different problems.

Abstract-We consider the problem of partially recovering hidden binary variables from the observation of (few) censored edge weights, a problem with applications in community detection, correlation clustering and synchronization. We describe two spectral algorithms for this task based on the non-backtracking and the Bethe Hessian operators. These algorithms are shown to be asymptotically optimal for the partial recovery problem, in that they detect the hidden assignment as soon as it is information theoretically possible to do so. A. IntroductionIn many inference problems, the available data can be represented on a weighted graph. Given the knowledge of the edge weights, the task is to infer latent variables carried by the nodes. Here, we shall consider the problem of recovering binary node labels from censored edge measurements [1], [2]. Specifically, given an Erdős-Rényi random graph G = (V, E) ∈ G(n, α/n) with n nodes carrying latent variables σ i = ±1, 1 ≤ i ≤ n, we draw the edge labels J ij = ±1, (ij) ∈ E from the following distribution:where is a noise parameter. In the noiseless case = 0, we have σ i σ j = J ij and one can easily recover the communities in each connected component along a spanning tree. When = 1/2, on the other hand, the graph doesn't contain any information about the latent variables σ i , and recovery is impossible. What happens in between? The problem of exactly recovering the latent variables σ i has been studied in [1]. It turns out that, asymptotically in the large n limit, exact recovery is shown to be possible if and only ifwhere α is the average degree of the graph. Note that the variable of an isolated vertex cannot be recovered so that the average degree has to grow at least like log n, as in the Coupon collector's problem, to ensure that the graph is connected.We consider in this paper the case where the average degree α will remain fixed as n tends to infinity. In this setting, we cannot ask for exact recovery and we consider here a different question: is it possible to infer an assignmentσ i of the latent variables that is positively correlated with the planted variables σ i ? We call positively correlated an assignmentσ i such that the following quantity, called overlap, is strictly positive:In the limit n → ∞, this overlap vanishes for a random guessσ i , and is equal to unity if the recovery is exact. We will refer to the task of finding a positively correlated assignment σ i as partial recovery. This task has been shown [3], [4] to be possible only ifTo the best of our knowledge, there is no rigorous proof that this bound is also sufficient. In [3], the same authors also showed that belief propagation (BP) allows to saturate this bound. However, there is no rigorous analysis of BP for this problem and the fact that condition (4) is necessary and sufficient was left as a conjecture in [3] and only the necessary part was proved in [4]. Moreover, from a practical point of view, BP requires the knowledge of the noise parameter .In this contribution, we describe two simple sp...

Abstract-We consider the problem of grouping items into clusters based on few random pairwise comparisons between the items. We introduce three closely related algorithms for this task: a belief propagation algorithm approximating the Bayes optimal solution, and two spectral algorithms based on the non-backtracking and Bethe Hessian operators. For the case of two symmetric clusters, we conjecture that these algorithms are asymptotically optimal in that they detect the clusters as soon as it is information theoretically possible to do so. We substantiate this claim for one of the spectral approaches we introduce.