United States

Knowledge distillation is considered as a training and compression strategy in which two neural networks, namely a teacher and a student, are coupled together during training. The teacher network is supposed to be a trustworthy predictor and the student tries to mimic its predictions. Usually, a student with a lighter architecture is selected so we can achieve compression and yet deliver high-quality results. In such a setting, distillation only happens for final predictions whereas the student could also benefit from teacherâ€™s supervision for internal components.

Motivated by this, we studied the problem of distillation for intermediate layers. Since there might not be a one-to-one alignment between student and teacher layers, existing techniques skip some teacher layers and only distill from a subset of them. This shortcoming directly impacts quality, so we instead propose a combinatorial technique which relies on attention. Our model fuses teacher-side information and takes each layerâ€™s significance into consideration, then it performs distillation between combined teacher layers and those of the student. Using our technique, we distilled a 12-layer BERT (Devlin et al. 2019) into 6-, 4-, and 2-layer counterparts and evaluated them on GLUE tasks (Wang et al. 2018). Experimental results show that our combinatorial approach is able to outperform other existing techniques.

AAAI 2021

ALP-KD: Attention-Based Layer Projection for Knowledge Distillation

The purpose of the AAAI conference is to promote research in artificial intelligence (AI) and scientific exchange among AI researchers, practitioners, scientists, and engineers in affiliated disciplines. 

technical paper

Large capacity machine learning (ML) models are prone to membership inference attacks (MIAs), which aim to infer whether the target sample is a member of the target model's training dataset. The serious privacy concerns due to the membership inference have motivated multiple defenses against MIAs, e.g., differential privacy and adversarial regularization. Unfortunately, these defenses produce ML models with unacceptably low classification performances.

Our work proposes a new defense, called distillation for membership privacy (DMP), against MIAs that preserves the utility of the resulting models significantly better than prior defenses.  DMP leverages knowledge distillation to train ML models with membership privacy. We provide a novel criterion to tune the data used for knowledge transfer in order to amplify the membership privacy of DMP.

Our extensive evaluation shows that DMP provides significantly better tradeoffs between membership privacy and classification accuracies compared to state-of-the-art MIA defenses. For instance, DMP achieves ~100% accuracy improvement over adversarial regularization for DenseNet trained on CIFAR100, for similar membership privacy (measured using MIA risk): when the MIA risk is 53.7%, adversarially regularized DenseNet is 33.6% accurate, while DMP-trained DenseNet is 65.3% accurate. We have released our code at github.com/vrt1shjwlkr/AAAI21-MIA-Defense.

Membership Privacy for Machine Learning Models through Knowledge Transfer

Unsupervised domain adaptation for semantic segmentation has been intensively studied due to the low cost of the pixel-level annotation for synthetic data. The most common approaches try to generate images or features mimicking the distribution in the target domain while preserving the semantic contents in the source domain so that a model can be trained with annotations from the latter. However, such methods highly rely on an image translator or feature extractor trained in an elaborated mechanism including adversarial training, which brings in extra complexity and instability in the adaptation process. Furthermore, these methods mainly focus on taking advantage of the labeled source dataset, leaving the unlabeled target dataset not fully utilized. In this paper, we propose a bidirectional style-induced domain adaptation method, called BiSIDA, that employs consistency regularization to efficiently exploit information from the unlabeled target domain dataset, requiring only a simple neural style transfer model. BiSIDA aligns domains by not only transferring source images into the style of target images but also transferring target images into the style of source images to perform high-dimensional perturbation on the unlabeled target images, which is crucial to the success in applying consistency regularization in segmentation tasks. Extensive experiments show that our BiSIDA achieves new state-of-the-art on two commonly-used synthetic-to-real domain adaptation benchmarks: GTA5-to-CityScapes and SYNTHIA-to-CityScapes. Code and pretrained style transfer model are available at: https://github.com/wangkaihong/BiSIDA.

Consistency Regularization with High-Dimensional Non-Adversarial Source-Guided Perturbation for Unsupervised Domain Adaptation in Segmentation

Designing an incentive compatible auction that maximizes expected revenue is a central problem in Auction Design. Theoretical approaches to the problem have hit some limits in the past decades and analytical solutions are known for only a few simple settings. Computational approaches to the problem through the use of LPs  have their own set of limitations. Building on the success of deep learning, a new approach was recently proposed by  Duetting et al. (2019) in which the auction is modeled by a feed-forward neural network and the design problem is framed as a learning problem. The neural architectures used in that work are general purpose and do not take advantage of any of the symmetries the problem could present, such as permutation equivariance. In this work, we consider auction design problems that have permutation-equivariant symmetry and construct a neural architecture that is capable of perfectly recovering the permutation-equivariant optimal mechanism, which we show is not possible with the previous architecture. We demonstrate that permutation-equivariant architectures are not only capable of recovering previous results, they also have better generalization properties.

A Permutation-Equivariant Neural Network Architecture for Auction Design

This paper focuses on a core task in computational sustainability and statistical ecology: species distribution modeling (SDM). In SDM, the occurrence pattern of a species on a landscape is predicted by environmental features based on observations at a set of locations. At first, SDM may appear to be a binary classification problem, and one might be inclined to employ classic tools (e.g., logistic regression, support vector machines, neural networks) to tackle it. However, wildlife surveys introduce structured noise (especially under-counting) in the species observations. If unaccounted for, these observation errors systematically bias SDMs. To address the unique challenges of SDM, this paper proposes a framework called StatEcoNet. Specifically, this work employs a graphical generative model in statistical ecology to serve as the skeleton of the proposed computational framework and carefully integrates neural networks under the framework. The advantages of StatEcoNet over related approaches are demonstrated on simulated datasets as well as bird species data. Since SDMs are critical tools for ecological science and natural resource management, StatEcoNet may offer boosted computational and analytical powers to a wide range of applications that have significant social impacts, e.g., the study and conservation of threatened species.

StatEcoNet: Statistical Ecology Neural Networks for Species Distribution Modeling

Current end-to-end neural conversation models inherently lack the flexibility to impose semantic control in the response generation process, often resulting in uninteresting responses. Attempts to boost informativeness alone come at the expense of factual accuracy, as attested by pretrained language models' propensity to "hallucinate" facts. While this may be mitigated by access to background knowledge, there is scant guarantee of relevance and informativeness in generated responses. We propose a framework that we call controllable grounded response generation (CGRG), in which lexical control phrases are either provided by a user or automatically extracted by a control phrase predictor from dialogue context and grounding knowledge. Quantitative and qualitative results show that, using this framework, a transformer based model with a novel inductive attention mechanism, trained on a conversation-like Reddit dataset, outperforms strong generation baselines.

A Controllable Model of Grounded Response Generation

In survival analysis, deep learning approaches have been proposed for estimating an individual's probability of survival over some time horizon. Such approaches can capture complex non-linear relationships, without relying on restrictive assumptions regarding the relationship between an individual's characteristics and their underlying survival process. To date, however, these methods have focused primarily on optimizing discriminative performance and have ignored model calibration. Well-calibrated survival curves present realistic and meaningful probabilistic estimates of the true underlying survival process for an individual. However, due to the lack of ground-truth regarding the underlying stochastic process of survival for an individual, optimizing and measuring calibration in survival analysis is an inherently difficult task. In this work, we i) highlight the shortcomings of existing approaches in terms of calibration and ii) propose a new training scheme for optimizing deep survival analysis models that maximizes discriminative performance, subject to good calibration. Compared to state-of-the-art approaches across two publicly available datasets, our proposed training scheme leads to significant improvements in calibration, while maintaining good discriminative performance.

Estimating Calibrated Individualized Survival Curves with Deep Learning

The black-box nature of neural models has motivated a line of research that aims to generate natural language rationales to explain why a model made certain predictions. Such rationale generation models, to date, have been trained on dataset-specific crowdsourced rationales, but this approach is costly and is not generalizable to new tasks and domains. In this paper, we investigate the extent to which neural models can reason about natural language rationales that explain model predictions, relying only on distant supervision with no additional annotation cost for human-written rationales. We investigate multiple ways to automatically generate rationales using pre-trained language models, neural knowledge models, and distant supervision from related tasks, and train generative models capable of composing explanatory rationales for unseen instances. We demonstrate our approach on the defeasible inference task, a nonmonotonic reasoning task in which an inference may be strengthened or weakened when new information (an update) is introduced. Our model shows promises at generating post-hoc rationales explaining why an inference is more or less likely given the additional information, however, it mostly generates trivial rationales reflecting the fundamental limitations of neural language models. Conversely, the more realistic setup of jointly predicting the update or its type and generating rationale is more challenging, suggesting an important future direction.

Learning to Rationalize for Nonmonotonic Reasoning with Distant Supervision

Controlling bias in training datasets is vital for ensuring equal treatment, or parity, between different groups in downstream applications. A naive solution is to transform the data so that it is statistically independent of group membership, but this may throw away too much information when a reasonable compromise between fairness and accuracy is desired. Another common approach is to limit the ability of a particular adversary who seeks to maximize parity. Unfortunately, representations produced by adversarial approaches may still retain biases as their efficacy is tied to the complexity of the adversary used during training. To this end, we theoretically establish that by limiting the mutual information between representations and protected attributes, we can assuredly control the parity of any downstream classifier. We demonstrate an effective method for controlling parity through mutual information based on contrastive information estimators and show that they outperform approaches that rely on variational bounds based on complex generative models. We test our approach on UCI Adult and Heritage Health datasets and demonstrate that our approach provides more informative representations across a range of desired parity thresholds while providing strong theoretical guarantees on the parity of any downstream algorithm.

Controllable Guarantees for Fair Outcomes via Contrastive Information Estimation

Many smartphone applications use inertial measurement units (IMUs) to sense movement, but the use of these sensors for pedestrian localization can be challenging due to their noise characteristics. Recent data-driven inertial odometry approaches have demonstrated the increasing feasibility of inertial navigation. However, they still rely upon conventional smartphone orientation estimates that they assume to be accurate, while in fact these orientation estimates can be a significant source of error. To address the problem of inaccurate orientation estimates, we present a two-stage, data-driven pipeline using a commodity smartphone that first estimates device orientations and then estimates device position. The orientation module relies on a recurrent neural network and Extended Kalman Filter to obtain orientation estimates that are used to then rotate raw IMU measurements into the appropriate reference frame. The position module then passes those measurements through another recurrent network architecture to perform localization. Our proposed method outperforms state-of-the-art methods in both orientation and position error on a large dataset we constructed that contains 20 hours of pedestrian motion across 3 buildings and 15 subjects. Code and data are available at https://github.com/KlabCMU/IDOL.

IDOL: Inertial Deep Orientation-Estimation and Localization

Graph Neural Networks (GNNs) have proven to be useful for many different practical applications. However, many existing GNN models have implicitly assumed homophily among the nodes connected in the graph, and therefore have largely overlooked the important setting of heterophily, where most connected nodes are from different classes. In this work, we propose a novel framework called CPGNN that generalizes GNNs for graphs with either homophily or heterophily. The proposed framework incorporates an interpretable compatibility matrix for modeling the heterophily or homophily level in the graph, which can be learned in an end-to-end fashion, enabling it to go beyond the assumption of strong homophily. Theoretically, we show that replacing the compatibility matrix in our framework with the identity (which represents pure homophily) reduces to GCN. Our extensive experiments demonstrate the effectiveness of our approach in more realistic and challenging experimental settings with significantly less training data compared to previous works: CPGNN variants achieve state-of-the-art results in heterophily settings with or without contextual node features, while maintaining comparable performance in homophily settings.

Downloads

Next from AAAI 2021

Membership Privacy for Machine Learning Models through Knowledge Transfer

Similar lecture

Cross-Layer Distillation with Semantic Calibration

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AAAI 2021

Membership Privacy for Machine Learning Models through Knowledge Transfer

Similar lecture

Cross-Layer Distillation with Semantic Calibration

Downloads