Abstract
Recent research has shown that seemingly fair machine learning models, when used to inform decisions that have an impact on peoples' lives or well-being (e.g., applications involving education, employment, and lending), can inadvertently increase social inequality in the long term. This is because prior fairness-aware algorithms only consider static fairness constraints, such as equal opportunity or demographic parity. However, enforcing constraints of this type may result in models that have negative long-term impact on disadvantaged individuals and communities. We introduce ELF (Enforcing Long-term Fairness), the first classification algorithm that provides high-confidence fairness guarantees in terms of long-term, or delayed, impact. We prove that the probability that ELF returns an unfair solution is less than a user-specified tolerance and that (under mild assumptions), given sufficient training data, ELF is able to find and return a fair solution if one exists. We show experimentally that our algorithm can successfully mitigate long-term unfairness.
Calibrated Selective Classification
Authors: Adam Fisch, Tommi Jaakkola, Regina Barzilay
Abstract
Selective classification allows models to abstain from making predictions (e.g., say "I don't know") when in doubt in order to obtain better effective accuracy. While typical selective models can be effective at producing more accurate predictions on average, they may still allow for wrong predictions that have high confidence, or skip correct predictions that have low confidence. Providing calibrated uncertainty estimates alongside predictions -- probabilities that correspond to true frequencies -- can be as important as having predictions that are simply accurate on average. However, uncertainty estimates can be unreliable for certain inputs. In this paper, we develop a new approach to selective classification in which we propose a method for rejecting examples with "uncertain" uncertainties. By doing so, we aim to make predictions with {well-calibrated} uncertainty estimates over the distribution of accepted examples, a property we call selective calibration. We present a framework for learning selectively calibrated models, where a separate selector network is trained to improve the selective calibration error of a given base model. In particular, our work focuses on achieving robust calibration, where the model is intentionally designed to be tested on out-of-domain data. We achieve this through a training strategy inspired by distributionally robust optimization, in which we apply simulated input perturbations to the known, in-domain training data. We demonstrate the empirical effectiveness of our approach on multiple image classification and lung cancer risk assessment tasks.
Quo Vadis: Hybrid Machine Learning Meta-Model based on Contextual and Behavioral Malware Representations
Authors: Dmitrijs Trizna
Subjects: Cryptography and Security (cs.CR); Machine Learning (cs.LG)
Abstract
We propose a hybrid machine learning architecture that simultaneously employs multiple deep learning models analyzing contextual and behavioral characteristics of Windows portable executable, producing a final prediction based on a decision from the meta-model. The detection heuristic in contemporary machine learning Windows malware classifiers is typically based on the static properties of the sample since dynamic analysis through virtualization is challenging for vast quantities of samples. To surpass this limitation, we employ a Windows kernel emulation that allows the acquisition of behavioral patterns across large corpora with minimal temporal and computational costs. We partner with a security vendor for a collection of more than 100k int-the-wild samples that resemble the contemporary threat landscape, containing raw PE files and filepaths of applications at the moment of execution. The acquired dataset is at least ten folds larger than reported in related works on behavioral malware analysis. Files in the training dataset are labeled by a professional threat intelligence team, utilizing manual and automated reverse engineering tools. We estimate the hybrid classifier's operational utility by collecting an out-of-sample test set three months later from the acquisition of the training set. We report an improved detection rate, above the capabilities of the current state-of-the-art model, especially under low false-positive requirements. Additionally, we uncover a meta-model's ability to identify malicious activity in validation and test sets even if none of the individual models express enough confidence to mark the sample as malevolent. We conclude that the meta-model can learn patterns typical to malicious samples from representation combinations produced by different analysis techniques. We publicly release pre-trained models and anonymized dataset of emulation reports.
Keyword: scaling
Wasserstein Task Embedding for Measuring Task Similarities
Abstract
Measuring similarities between different tasks is critical in a broad spectrum of machine learning problems, including transfer, multi-task, continual, and meta-learning. Most current approaches to measuring task similarities are architecture-dependent: 1) relying on pre-trained models, or 2) training networks on tasks and using forward transfer as a proxy for task similarity. In this paper, we leverage the optimal transport theory and define a novel task embedding for supervised classification that is model-agnostic, training-free, and capable of handling (partially) disjoint label sets. In short, given a dataset with ground-truth labels, we perform a label embedding through multi-dimensional scaling and concatenate dataset samples with their corresponding label embeddings. Then, we define the distance between two datasets as the 2-Wasserstein distance between their updated samples. Lastly, we leverage the 2-Wasserstein embedding framework to embed tasks into a vector space in which the Euclidean distance between the embedded points approximates the proposed 2-Wasserstein distance between tasks. We show that the proposed embedding leads to a significantly faster comparison of tasks compared to related approaches like the Optimal Transport Dataset Distance (OTDD). Furthermore, we demonstrate the effectiveness of our proposed embedding through various numerical experiments and show statistically significant correlations between our proposed distance and the forward and backward transfer between tasks.
Practical Real Video Denoising with Realistic Degradation Model
Authors: Jiezhang Cao, Qin Wang, Jingyun Liang, Yulun Zhang, Kai Zhang, Luc Van Gool
Subjects: Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV)
Abstract
Existing video denoising methods typically assume noisy videos are degraded from clean videos by adding Gaussian noise. However, deep models trained on such a degradation assumption will inevitably give rise to poor performance for real videos due to degradation mismatch. Although some studies attempt to train deep models on noisy and noise-free video pairs captured by cameras, such models can only work well for specific cameras and do not generalize well for other videos. In this paper, we propose to lift this limitation and focus on the problem of general real video denoising with the aim to generalize well on unseen real-world videos. We tackle this problem by firstly investigating the common behaviors of video noises and observing two important characteristics: 1) downscaling helps to reduce the noise level in spatial space and 2) the information from the adjacent frames help to remove the noise of current frame in temporal space. Motivated by these two observations, we propose a multi-scale recurrent architecture by making full use of the above two characteristics. Secondly, we propose a synthetic real noise degradation model by randomly shuffling different noise types to train the denoising model. With a synthesized and enriched degradation space, our degradation model can help to bridge the distribution gap between training data and real-world data. Extensive experiments demonstrate that our proposed method achieves the state-of-the-art performance and better generalization ability than existing methods on both synthetic Gaussian denoising and practical real video denoising.
Coherent photonic crossbar as a universal linear operator
Authors: George Giamougiannis, Apostolos Tsakyridis, Yangjin Ma, Angelina Totovic, David Lazovsky, Nikos Pleros
Abstract
Linear optics aim at realizing any real- and/or complex-valued matrix operator via optical elements, addressing a broad field of applications in the areas of quantum photonics, microwave photonics and optical neural networks. The transfer of linear operators into photonic experimental layouts typically relies on Singular Value Decomposition (SVD) techniques combining meshes of cascaded 2x2 Mach Zehnder Interferometers (MZIs), with the main challenges being the precision in the experimental representation of the targeted matrix, referred to as fidelity, and the overall insertion loss. We demonstrate a novel interferometric coherent photonic crossbar architecture (Xbar) that demarcates from state-of-the-art SVD implementations and can realize any linear operator, supporting full restoration of the loss-induced fidelity. Its novel interferometric design allows for the direct mapping of each matrix element to a single, designated Xbar node, bringing down the number of programming steps to only one. We present the theoretical foundations of the Xbar, proving that its insertion losses scale linearly with the node losses as opposed to the exponential scaling witnessed by the SVD counterparts. This leads to a matrix design with significantly lower overall insertion losses compared to SVD-based schemes when utilizing state-of-the-art silicon photonic fabrication metrics, allowing for alternative node technologies with lower energy consumption and higher operational speed credentials to be employed. Finally, we validate that our Xbar architecture is the first linear operator that supports fidelity restoration, outperforming SVD schemes in loss- and phase-error fidelity performance and forming a significantly more robust layout to loss and phase deviations.
Seamless Tracking of Group Targets and Ungrouped Targets Using Belief Propagation
Abstract
This paper considers the problem of tracking a large-scale number of group targets. Usually, multi-target in most tracking scenarios are assumed to have independent motion and are well-separated. However, for group target tracking (GTT), the targets within groups are closely spaced and move in a coordinated manner, the groups can split or merge, and the numbers of targets in groups may be large, which lead to more challenging data association, filtering and computation problems. Within the belief propagation (BP) framework, we propose a scalable group target belief propagation (GTBP) method by jointly inferring target existence variables, group structure, data association and target states. The method can efficiently calculate the approximations of the marginal posterior distributions of these variables by performing belief propagation on the devised factor graph. As a consequence, GTBP is capable of capturing the changes in group structure, e.g., group splitting and merging. Furthermore, we model the evolution of targets as the co-action of the group or single-target motions specified by the possible group structures and corresponding probabilities. This flexible modeling enables seamless and simultaneous tracking of multiple group targets and ungrouped targets. Particularly, GTBP has excellent scalability and low computational complexity. It not only maintains the same scalability as BP, i.e., scaling linearly in the number of sensor measurements and quadratically in the number of targets, but also only scales linearly in the number of preserved group partitions. Finally, numerical experiments are presented to demonstrate the effectiveness and scalability of the proposed GTBP method.
Abstract
This paper presents Lidar-based Simultaneous Localization and Mapping (SLAM) for autonomous driving vehicles. Fusing data from landmark sensors and a strap-down Inertial Measurement Unit (IMU) in an adaptive Kalman filter (KF) plus the observability of the system are investigated. In addition to the vehicle's states and landmark positions, a self-tuning filter estimates the IMU calibration parameters as well as the covariance of the measurement noise. The discrete-time covariance matrix of the process noise, the state transition matrix, and the observation sensitivity matrix are derived in closed-form making them suitable for real-time implementation. Examining the observability of the 3D SLAM system leads to the conclusion that the system remains observable upon a geometrical condition on the alignment of the landmarks.
FusionPortable: A Multi-Sensor Campus-Scene Dataset for Evaluation of Localization and Mapping Accuracy on Diverse Platforms
Authors: Jianhao Jiao, Hexiang Wei, Tianshuai Hu, Xiangcheng Hu, Yilong Zhu, Zhijian He, Jin Wu, Jingwen Yu, Xupeng Xie, Huaiyang Huang, Ruoyu Geng, Lujia Wang, Ming Liu
Abstract
Combining multiple sensors enables a robot to maximize its perceptual awareness of environments and enhance its robustness to external disturbance, crucial to robotic navigation. This paper proposes the FusionPortable benchmark, a complete multi-sensor dataset with a diverse set of sequences for mobile robots. This paper presents three contributions. We first advance a portable and versatile multi-sensor suite that offers rich sensory measurements: 10Hz LiDAR point clouds, 20Hz stereo frame images, high-rate and asynchronous events from stereo event cameras, 200Hz inertial readings from an IMU, and 10Hz GPS signal. Sensors are already temporally synchronized in hardware. This device is lightweight, self-contained, and has plug-and-play support for mobile robots. Second, we construct a dataset by collecting 17 sequences that cover a variety of environments on the campus by exploiting multiple robot platforms for data collection. Some sequences are challenging to existing SLAM algorithms. Third, we provide ground truth for the decouple localization and mapping performance evaluation. We additionally evaluate state-of-the-art SLAM approaches and identify their limitations. The dataset, consisting of raw sensor easurements, ground truth, calibration data, and evaluated algorithms, will be released: https://ram-lab.com/file/site/multi-sensor-dataset.
Calibrated Selective Classification
Authors: Adam Fisch, Tommi Jaakkola, Regina Barzilay
Abstract
Selective classification allows models to abstain from making predictions (e.g., say "I don't know") when in doubt in order to obtain better effective accuracy. While typical selective models can be effective at producing more accurate predictions on average, they may still allow for wrong predictions that have high confidence, or skip correct predictions that have low confidence. Providing calibrated uncertainty estimates alongside predictions -- probabilities that correspond to true frequencies -- can be as important as having predictions that are simply accurate on average. However, uncertainty estimates can be unreliable for certain inputs. In this paper, we develop a new approach to selective classification in which we propose a method for rejecting examples with "uncertain" uncertainties. By doing so, we aim to make predictions with {well-calibrated} uncertainty estimates over the distribution of accepted examples, a property we call selective calibration. We present a framework for learning selectively calibrated models, where a separate selector network is trained to improve the selective calibration error of a given base model. In particular, our work focuses on achieving robust calibration, where the model is intentionally designed to be tested on out-of-domain data. We achieve this through a training strategy inspired by distributionally robust optimization, in which we apply simulated input perturbations to the known, in-domain training data. We demonstrate the empirical effectiveness of our approach on multiple image classification and lung cancer risk assessment tasks.
Keyword: out of distribution detection
There is no result
Keyword: out-of-distribution detection
There is no result
Keyword: expected calibration error
There is no result
Keyword: overconfident
There is no result
Keyword: overconfidence
There is no result
Keyword: confidence
Enforcing Delayed-Impact Fairness Guarantees
Calibrated Selective Classification
Quo Vadis: Hybrid Machine Learning Meta-Model based on Contextual and Behavioral Malware Representations
Keyword: scaling
Wasserstein Task Embedding for Measuring Task Similarities
Practical Real Video Denoising with Realistic Degradation Model
Coherent photonic crossbar as a universal linear operator
Seamless Tracking of Group Targets and Ungrouped Targets Using Belief Propagation
Keyword: calibration
Lidar SLAM for Autonomous Driving Vehicles
FusionPortable: A Multi-Sensor Campus-Scene Dataset for Evaluation of Localization and Mapping Accuracy on Diverse Platforms
Calibrated Selective Classification