Evaluation of keyword spotting (KWS) systems that detect keywords in speech is a challenging task under realistic privacy constraints. The KWS is designed to only collect data when the keyword is present, limiting the availability of hard samples that may contain false negatives, and preventing direct estimation of model recall from production data. Alternatively, complementary data collected from other sources may not be fully representative of the real application. In this work, we propose an evaluation technique which we call AB/BA analysis. Our framework evaluates a candidate KWS model B against a baseline model A, using cross-dataset offline decoding for relative recall estimation, without requiring negative examples. Moreover, we propose a formulation with assumptions that allow estimation of relative false positive rate between models with low variance even when the number of false positives is small. Finally, we propose to leverage machine-generated soft labels, in a technique we call Semi-Supervised AB/BA analysis, that improves the analysis time, privacy, and cost. Experiments with both simulation and real data show that AB/BA analysis is successful at measuring recall improvement in conjunction with the trade-off in relative false positive rate.
相關內容
We present a data-driven approach to characterizing nonidentifiability of a model's parameters and illustrate it through dynamic as well as steady kinetic models. By employing Diffusion Maps and their extensions, we discover the minimal combinations of parameters required to characterize the output behavior of a chemical system: a set of effective parameters for the model. Furthermore, we introduce and use a Conformal Autoencoder Neural Network technique, as well as a kernel-based Jointly Smooth Function technique, to disentangle the redundant parameter combinations that do not affect the output behavior from the ones that do. We discuss the interpretability of our data-driven effective parameters, and demonstrate the utility of the approach both for behavior prediction and parameter estimation. In the latter task, it becomes important to describe level sets in parameter space that are consistent with a particular output behavior. We validate our approach on a model of multisite phosphorylation, where a reduced set of effective parameters (nonlinear combinations of the physical ones) has previously been established analytically.
Age of information (AoI) is an effective performance metric measuring the freshness of information and is popular for applications involving status update. Most of the existing works have adopted average AoI as the metric, which cannot provide strict performance guarantees. In this work, the outage probability of the peak AoI exceeding a given threshold is analyzed in a multi-source system under round robin scheduling. Two queueing disciplines are considered, namely the first-come-first-serve (FCFS) queue and the single packet queue. For FCFS, upper and lower bounds on the outage probability are derived which coincides asymptotically, characterizing its true scaling. For the single packet queue, an upper bound is derived whose effectiveness is validated by the simulation results. The analysis concretizes the common belief that single packet queueing has a better AoI performance than FCFS. Moreover, it also reveals that the two disciplines would have similar asymptotic performance when the inter-arrival time is much larger than the total transmission time.
Aleatoric uncertainty quantification seeks for distributional knowledge of random responses, which is important for reliability analysis and robustness improvement in machine learning applications. Previous research on aleatoric uncertainty estimation mainly targets closed-formed conditional densities or variances, which requires strong restrictions on the data distribution or dimensionality. To overcome these restrictions, we study conditional generative models for aleatoric uncertainty estimation. We introduce two metrics to measure the discrepancy between two conditional distributions that suit these models. Both metrics can be easily and unbiasedly computed via Monte Carlo simulation of the conditional generative models, thus facilitating their evaluation and training. We demonstrate numerically how our metrics provide correct measurements of conditional distributional discrepancies and can be used to train conditional models competitive against existing benchmarks.
Recent techniques for approximating Nash equilibria in very large games leverage neural networks to learn approximately optimal policies (strategies). One promising line of research uses neural networks to approximate counterfactual regret minimization (CFR) or its modern variants. DREAM, the only current CFR-based neural method that is model free and therefore scalable to very large games, trains a neural network on an estimated regret target that can have extremely high variance due to an importance sampling term inherited from Monte Carlo CFR (MCCFR). In this paper we propose an unbiased model-free method that does not require any importance sampling. Our method, ESCHER, is principled and is guaranteed to converge to an approximate Nash equilibrium with high probability in the tabular case. We show that the variance of the estimated regret of a tabular version of ESCHER with an oracle value function is significantly lower than that of outcome sampling MCCFR and tabular DREAM with an oracle value function. We then show that a deep learning version of ESCHER outperforms the prior state of the art -- DREAM and neural fictitious self play (NFSP) -- and the difference becomes dramatic as game size increases.
Domain generalization (DG) aims at learning generalizable models under distribution shifts to avoid redundantly overfitting massive training data. Previous works with complex loss design and gradient constraint have not yet led to empirical success on large-scale benchmarks. In this work, we reveal the mixture-of-experts (MoE) model's generalizability on DG by leveraging to distributively handle multiple aspects of the predictive features across domains. To this end, we propose Sparse Fusion Mixture-of-Experts (SF-MoE), which incorporates sparsity and fusion mechanisms into the MoE framework to keep the model both sparse and predictive. SF-MoE has two dedicated modules: 1) sparse block and 2) fusion block, which disentangle and aggregate the diverse learned signals of an object, respectively. Extensive experiments demonstrate that SF-MoE is a domain-generalizable learner on large-scale benchmarks. It outperforms state-of-the-art counterparts by more than 2% across 5 large-scale DG datasets (e.g., DomainNet), with the same or even lower computational costs. We further reveal the internal mechanism of SF-MoE from distributed representation perspective (e.g., visual attributes). We hope this framework could facilitate future research to push generalizable object recognition to the real world. Code and models are released at //github.com/Luodian/SF-MoE-DG.
The accelerated use of digital cameras prompts an increasing concern about privacy and security, particularly in applications such as action recognition. In this paper, we propose an optimizing framework to provide robust visual privacy protection along the human action recognition pipeline. Our framework parameterizes the camera lens to successfully degrade the quality of the videos to inhibit privacy attributes and protect against adversarial attacks while maintaining relevant features for activity recognition. We validate our approach with extensive simulations and hardware experiments.
Each year, deep learning demonstrates new and improved empirical results with deeper and wider neural networks. Meanwhile, with existing theoretical frameworks, it is difficult to analyze networks deeper than two layers without resorting to counting parameters or encountering sample complexity bounds that are exponential in depth. Perhaps it may be fruitful to try to analyze modern machine learning under a different lens. In this paper, we propose a novel information-theoretic framework with its own notions of regret and sample complexity for analyzing the data requirements of machine learning. With our framework, we first work through some classical examples such as scalar estimation and linear regression to build intuition and introduce general techniques. Then, we use the framework to study the sample complexity of learning from data generated by deep sign neural networks, deep ReLU neural networks, and deep networks that are infinitely wide but have a bounded sum of weights. For sign neural networks, we recover sample-complexity bounds that follow from VC-dimension based arguments. For the latter two neural network environments, we establish new results that suggest that the sample complexity of learning under these data generating processes is at most linear and quadratic, respectively, in network depth.
Many of the successes of machine learning are based on minimizing an averaged loss function. However, it is well-known that this paradigm suffers from robustness issues that hinder its applicability in safety-critical domains. These issues are often addressed by training against worst-case perturbations of data, a technique known as adversarial training. Although empirically effective, adversarial training can be overly conservative, leading to unfavorable trade-offs between nominal performance and robustness. To this end, in this paper we propose a framework called probabilistic robustness that bridges the gap between the accurate, yet brittle average case and the robust, yet conservative worst case by enforcing robustness to most rather than to all perturbations. From a theoretical point of view, this framework overcomes the trade-offs between the performance and the sample-complexity of worst-case and average-case learning. From a practical point of view, we propose a novel algorithm based on risk-aware optimization that effectively balances average- and worst-case performance at a considerably lower computational cost relative to adversarial training. Our results on MNIST, CIFAR-10, and SVHN illustrate the advantages of this framework on the spectrum from average- to worst-case robustness.
The crude Monte Carlo approximates the integral $$S(f)=\int_a^b f(x)\,\mathrm dx$$ with expected error (deviation) $\sigma(f)N^{-1/2},$ where $\sigma(f)^2$ is the variance of $f$ and $N$ is the number of random samples. If $f\in C^r$ then special variance reduction techniques can lower this error to the level $N^{-(r+1/2)}.$ In this paper, we consider methods of the form $$\overline M_{N,r}(f)=S(L_{m,r}f)+M_n(f-L_{m,r}f),$$ where $L_{m,r}$ is the piecewise polynomial interpolation of $f$ of degree $r-1$ using a partition of the interval $[a,b]$ into $m$ subintervals, $M_n$ is a Monte Carlo approximation using $n$ samples of $f,$ and $N$ is the total number of function evaluations used. We derive asymptotic error formulas for the methods $\overline M_{N,r}$ that use nonadaptive as well as adaptive partitions. Although the convergence rate $N^{-(r+1/2)}$ cannot be beaten, the asymptotic constants make a huge difference. For example, for $\int_0^1(x+d)^{-1}\mathrm dx$ and $r=4$ the best adaptive methods overcome the nonadaptive ones roughly $10^{12}$ times if $d=10^{-4},$ and $10^{29}$ times if $d=10^{-8}.$ In addition, the proposed adaptive methods are easily implementable and can be well used for automatic integration. We believe that the obtained results can be generalized to multivariate integration.
We study the problem of histogram estimation under user-level differential privacy, where the goal is to preserve the privacy of all entries of any single user. While there is abundant literature on this classical problem under the item-level privacy setup where each user contributes only one data point, little has been known for the user-level counterpart. We consider the heterogeneous scenario where both the quantity and distribution of data can be different for each user. We propose an algorithm based on a clipping strategy that almost achieves a two-approximation with respect to the best clipping threshold in hindsight. This result holds without any distribution assumptions on the data. We also prove that the clipping bias can be significantly reduced when the counts are from non-i.i.d. Poisson distributions and show empirically that our debiasing method provides improvements even without such constraints. Experiments on both real and synthetic datasets verify our theoretical findings and demonstrate the effectiveness of our algorithms.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191