arXiv subjects

Xi Yu

Publications and source records attributed to Xi Yu.

43 records · Page 3Linked to original sources

arXiv2021

Gated Information Bottleneck for Generalization in Sequential Environments

Deep neural networks suffer from poor generalization to unseen environments when the underlying data distribution is different from that in the training set. By learning minimum sufficient representations from training data, the information bottleneck (IB) approach has demonstrated its effectiveness to improve generalization in different AI applications. In this work, we propose a new neural network-based IB approach, termed gated information bottleneck (GIB), that dynamically drops spurious correlations and progressively selects the most task-relevant features across different environments by a trainable soft mask (on raw features). GIB enjoys a simple and tractable objective, without any variational approximation or distributional assumption. We empirically demonstrate the superiority of GIB over other popular neural network-based IB approaches in adversarial robustness and out-of-distribution (OOD) detection. Meanwhile, we also establish the connection between IB theory and invariant causal representation learning, and observed that GIB demonstrates appealing performance when different environments arrive sequentially, a more practical scenario where invariant risk minimization (IRM) fails. Code of GIB is available at https://github.com/falesiani/GIB

cs.LG↗

arXiv2021

Privacy-Preserving Federated Learning on Partitioned Attributes

Real-world data is usually segmented by attributes and distributed across different parties. Federated learning empowers collaborative training without exposing local data or models. As we demonstrate through designed attacks, even with a small proportion of corrupted data, an adversary can accurately infer the input attributes. We introduce an adversarial learning based procedure which tunes a local model to release privacy-preserving intermediate representations. To alleviate the accuracy decline, we propose a defense method based on the forward-backward splitting algorithm, which respectively deals with the accuracy loss and privacy loss in the forward and backward gradient descent steps, achieving the two objectives simultaneously. Extensive experiments on a variety of datasets have shown that our defense significantly mitigates privacy leakage with negligible impact on the federated learning task.

cs.LG↗

arXiv2021

Deep Deterministic Information Bottleneck with Matrix-based Entropy Functional

We introduce the matrix-based Renyi's $α$-order entropy functional to parameterize Tishby et al. information bottleneck (IB) principle with a neural network. We term our methodology Deep Deterministic Information Bottleneck (DIB), as it avoids variational inference and distribution assumption. We show that deep neural networks trained with DIB outperform the variational objective counterpart and those that are trained with other forms of regularization, in terms of generalization performance and robustness to adversarial attack.Code available at https://github.com/yuxi120407/DIB

cs.LG↗

arXiv2021

Measuring Dependence with Matrix-based Entropy Functional

Measuring the dependence of data plays a central role in statistics and machine learning. In this work, we summarize and generalize the main idea of existing information-theoretic dependence measures into a higher-level perspective by the Shearer's inequality. Based on our generalization, we then propose two measures, namely the matrix-based normalized total correlation ($T_α^*$) and the matrix-based normalized dual total correlation ($D_α^*$), to quantify the dependence of multiple variables in arbitrary dimensional space, without explicit estimation of the underlying data distributions. We show that our measures are differentiable and statistically more powerful than prevalent ones. We also show the impact of our measures in four different machine learning problems, namely the gene regulatory network inference, the robust machine learning under covariate shift and non-Gaussian noises, the subspace outlier detection, and the understanding of the learning dynamics of convolutional neural networks (CNNs), to demonstrate their utilities, advantages, as well as implications to those problems. Code of our dependence measure is available at: https://bit.ly/AAAI-dependence

cs.LG↗

arXiv2018

Optimal Threshold-Based Control Policies for Persistent Monitoring on Graphs

We consider the optimal multi-agent persistent monitoring problem defined by a team of cooperating agents visiting a set of nodes (targets) on a graph with the objective of minimizing a measure of overall node state uncertainty. The solution to this problem involves agent trajectories defined both by the sequence of nodes to be visited by each agent and the amount of time spent at each node. Since such optimal trajectories are generally intractable, we propose a class of distributed threshold-based parametric controllers through which agent transitions from one node to the next are controlled by threshold parameters on the node uncertainty states. The resulting behavior of the agent-target system can be described by a hybrid dynamic system. This enables the use of Infinitesimal Perturbation Analysis (IPA) to determine on line (locally) optimal threshold parameters through gradient descent methods and thus obtain optimal controllers within this family of threshold-based policies. We further show that in a single-agent case the IPA gradient is monotonic, which implies a simple structure whereby an agent visiting a node should reduce the uncertainty state to zero before moving to the next node. Simulation examples are included to illustrate our results and compare them to optimal solutions derived through dynamic programming when this is possible.

math.OC↗

arXiv2017

Decentralized Event-Driven Algorithms for Multi-Agent Persistent Monitoring

We address the issue of identifying conditions under which the centralized solution to the optimal multi-agent persistent monitoring problem can be recovered in a decentralized event-driven manner. In this problem, multiple agents interact with a finite number of targets and the objective is to control their movement in order to minimize an uncertainty metric associated with the targets. In a one-dimensional setting, it has been shown that the optimal solution can be reduced to a simpler parametric optimization problem and that the behavior of agents under optimal control is described by a hybrid system. This hybrid system can be analyzed using Infinitesimal Perturbation Analysis (IPA) to obtain a complete on-line solution through an event-driven centralized gradient-based algorithm. We show that the IPA gradient can be recovered in a distributed manner in which each agent optimizes its trajectory based on local information, except for one event requiring communication from a non-neighbor agent. Simulation examples are included to illustrate the effectiveness of this "almost decentralized" algorithm and its fully decentralized counterpart where the aforementioned non-local event is ignored.

math.OC↗

arXiv2016

Optimal Event-Driven Multi-Agent Persistent Monitoring of a Finite Set of Targets

We consider the problem of controlling the movement of multiple cooperating agents so as to minimize an uncertainty metric associated with a finite number of targets. In a one-dimensional mission space, we adopt an optimal control framework and show that the solution is reduced to a simpler parametric optimization problem: determining a sequence of locations where each agent may dwell for a finite amount of time and then switch direction. This amounts to a hybrid system which we analyze using Infinitesimal Perturbation Analysis (IPA) to obtain a complete on-line solution through an event-driven gradient-based algorithm which is also robust with respect to the uncertainty model used. The resulting controller depends on observing the events required to excite the gradient-based algorithm, which cannot be guaranteed. We solve this problem by proposing a new metric for the objective function which creates a potential field guaranteeing that gradient values are non-zero. This approach is compared to an alternative graph-based task scheduling algorithm for determining an optimal sequence of target visits. Simulation examples are included to demonstrate the proposed methods.

math.OC↗