ACML 2023 Papers

Deep Metric Learning (DML) is a prominent subfield of machine learning with extensive practical applications in learning visual similarities. However, DML systems are vulnerable to input distributions during inference that differ from the training data, such as adversarial examples (AXs). In this paper, we introduce MDProp, a framework that enhances the clean data performance and adversarial robustness of DML models by generating novel Multi-Targeted AXs and Unadversarial Examples, in addition to conventional single-targeted AXs, in the feature space. To handle the input distribution shift caused by the generated novel input distributions, MDProp scales the separate batch normalization layer strategy. Our comprehensive experimental analysis demonstrates that MDProp outperforms current state-of-the-art convolutional neural networks by up to 2.95% in terms of R@1 scores for clean data, while simultaneously improving adversarial robustness by up to 2.12 times. Additionally, MDProp achieves state-of-the-art results in data-scarce setting while utilizing only half of the training data. Implementation is available at \url{https://github.com/intherejeet/MDProp}.

Recent advances in manifold optimization for the Gaussian mixture model (GMM) have gained increasing interest. In this work, instead of directly addressing the manifold optimization on covariance matrices of GMM, we consider the GMM fitting as an optimization of the density function over a statistical manifold and seek the natural gradient to speed up the optimization process. We present an upper bound for the Kullback–Leibler (KL) divergence between two GMMs and obtain simple closed-form expressions for the natural gradients. With the natural gradients, we then apply the Riemannian stochastic gradient descent (RSGD) algorithm to optimize covariance matrices on a symmetric and positive definite (SPD) matrix manifold. We further propose a Riemannian Adam (RAdam) algorithm that extends the momentum method and adaptive learning in the Euclidean space to the SPD manifold space. Extensive simulations show that the proposed algorithms scale well to high-dimensional large-scale datasets and outperform expectation maximization (EM) algorithms in fitted log-likelihood.

Attribute-aware face recognition has gained increasing attention in recent years due to its potential to improve the robustness of face recognition systems. However, this may raise concerns about potential biases and privacy issues. To alleviate this, some studies involve complex designs to obtain independent ID and attribute features and fuse them based on the application scenario (for better accuracy or fairness). In this paper, we obviate their complex design and demonstrate that the Nearest neighbours Proxy Triplet (NPT) loss has an intrinsic capability for feature disentanglement. To further enhance the effectiveness of NPT, we propose a novel margin-based loss, namely Adaptive-rank NPT, which naturally separates the identity and attribute features. While a margin-based loss ensures inter-class separability, it imposes no constraints on intra-class compactness. The samples that meet the inter-class margin will not contribute to network training. To mitigate this issue, we propose an adaptive distance measurement to promote the compactness of the learned features, resulting in the final Ada$^2$NPT loss. The experimental results obtained on several benchmarks demonstrate the superiority and merits of the proposed loss function over the state-of-the-art losses in terms of accuracy and fairness.

A common problem encountered in many real-world applications is level set estimation where the goal is to determine the region in the function domain where the function is above or below a given threshold. When the function is black-box and expensive to evaluate, the level sets need to be found in a minimum set of function evaluations. Existing methods often assume a discrete search space with a finite set of data points for function evaluations and estimating the level sets. When applied to a continuous search space, these methods often need to first discretize the space which leads to poor results while needing high computational time. While some methods cater for the continuous setting, they still lack a proper guarantee for theoretical convergence. To address this problem, we propose a novel algorithm that does not need any discretization and can directly work in continuous search spaces. Our method suggests points by constructing an acquisition function that is defined as a measure of confidence of the function being higher or lower than the given threshold. A theoretical analysis for the convergence of the algorithm to an accurate solution is provided. On multiple synthetic and real-world datasets, our algorithm successfully outperforms state-of-the-art methods.

Deep reinforcement learning (RL) has shown remarkable performance, but end-users do not understand how the system solves tasks due to the black-box nature of neural networks. Many methods from explainable machine learning have been adapted to RL. However, they do not focus on the unique challenges of explaining actions’ short-term and long-term consequences. This work introduces a new perspective to understanding RL policies by clustering states into abstract states utilizing attention maps, giving a bird’s-eye view of the policy’s behavior. We learn the attention maps iteratively together with the clustering of states by masking the input features to estimate their importance. In contrast to previous works that have uninterpretable abstract states and/or clustering objectives using state values that are non-human intuitive, we only leverage attention maps in the clustering. The policy only indirectly affects the clustering via attention maps. This allows us to give global explanations from the view of feature attention, a quantity a human can relate to given interpretable features. The experiments demonstrate that our method provides faithful abstractions by capturing state semantics, policy behavior, and feature attention. Furthermore, we show that our attention maps can mask state features without affecting policy performance.

The introduction of the generative adversarial imitation learning (GAIL) algorithm has spurred the development of scalable imitation learning approaches using deep neural networks. Many of the algorithms that followed used a similar procedure, combining on-policy actor-critic algorithms with inverse reinforcement learning. More recently there have been an even larger breadth of approaches, most of which use off-policy algorithms. However, with the breadth of algorithms, everything from datasets to base reinforcement learning algorithms to evaluation settings can vary, making it difficult to fairly compare them. In this work we re-implement 6 different IL algorithms, updating 3 of them to be off-policy, base them on a common off-policy algorithm (SAC), and evaluate them on a widely-used expert trajectory dataset (D4RL) for the most common benchmark (MuJoCo). After giving all algorithms the same hyperparameter optimisation budget, we compare their results for a range of expert trajectories. In summary, GAIL, with all of its improvements, consistently performs well across a range of sample sizes, AdRIL is a simple contender that performs well with one important hyperparameter to tune, and behavioural cloning remains a strong baseline when data is more plentiful.

Balancing exploration and exploitation has long been recognized as an important theme in the online planning algorithms for POMDP problems. Explorative actions on one hand prevent the planning from falling into the suboptimal dilemma, while hindering the convergence of the planning procedure on the other hand. Therefore, it is meaningful to maintain the exploration as well as taking a step forward towards exploitation. Note that there is a deviation between the action selection criteria in the planning procedure and in the execution procedure, which inspires us to build a bridge between these two criteria to accelerate the convergence. A Partially Observable Monte Carlo Planning algorithm based on Path Modification (POMCP-PM) is presented in the paper, which modifies the backtracing paths by considering the two criteria simultaneously when updating the values of parent nodes. The algorithm is general as the Upper Confidence Bound Apply to Tree (UCT) algorithm used to select actions can be easily replaced by other criteria. Experimental results demonstrate that POMCP-PM outperforms POMCP with varying numbers of simulations on several scenarios with different scales.

Aspect-based-sentiment-analysis (ABSA) is a fine-grained sentiment evaluation task, which analyzes the emotional polarity of the evaluation aspects. Generally, the emotional polarity of an aspect exists in the corresponding opinion expression, whose diversity has great impact on model’s performance. To mitigate this problem, we propose a novel and simple counterfactual data augmentation method to generate opinion expressions with reversed sentiment polarity. In particular, the integrated gradients are calculated to locate and mask the opinion expression. Then, a prompt combined with the reverse expression polarity is added to the original text, and a Pre-trained language model (PLM), T5, is finally was employed to predict the masks. The experimental results shows the proposed counterfactual data augmentation method performs better than current augmentation methods on three ABSA datasets, i.e. Laptop, Restaurant, and MAMS.

Graph neural networks (GNNs) have demonstrated impressive achievements in diverse graph tasks, and research on their expressive power has experienced significant growth in recent years. The well-known Weisfeiler and Lehman (WL) isomorphism test has been widely used to assess GNNs’ ability to distinguish graph structures. However, despite being considered less expressive than other GNNs in graph-level tasks based on the WL test, two prominent GNN models, namely graph convolution networks (GCN) and attention-based graph networks (GAT), still exhibit strong performance in node-level classification tasks. In this paper, we present a comprehensive analysis of their expressive power using a novel evaluation metric: the number of linear regions. We demonstrate that by enhancing GCN with refined graph Ricci curvature, our proposed high-rank graph convolution network (HRGCN) can match or even surpass the prediction advantage of attention models. Thus, the two models exhibit equivalent node-level expressive powers. This fresh perspective highlights the evaluation of GNNs’ expressive power in node-level classifications rather than solely at the graph level. Experimental results showcase that the proposed HRGCN model outperforms the state-of-the-art in various classification and prediction tasks.

The salp swarm algorithm(SSA) has been successfully used to solve the feature selection problem due to its fast convergence and simple structure. However, existing SSA-based methods still suffer from the issue of low classification accuracy due to the problem of getting trapped in local optima. Therefore, this paper proposes a novel feature selection method for classification based on SSA, which can continuously generate new sub-populations to improve the search environment of the main population. Specifically, a flip-prohibition(F-P) operator is first proposed to help the main population, which may currently fall into a local optimum, find a new and more promising region. A multi-surrogate technique is suggested to evaluate the region to determine the position of sub-populations, which can reduce the high computational cost. In addition, a population initialization method is developed according to the importance of features and the dimensionality of the dataset. Finally, a communication mechanism is presented to enable different sub-populations to learn from each other. By comparing the proposed method with other 6 feature selection methods on 16 datasets, we demonstrate that the proposed method has better classification ability and can select a smaller feature subset in most cases.

Reducing the memory usage and computational complexity of high-performance deep neural networks while minimizing degradation of accuracy is a key issue in implementing these models on edge devices. To address this issue, partial quantization methods have been proposed to partially reduce the weight parameters of neural network models. However, the accuracy of existing methods degrades rapidly with increasing compression ratio. Although retraining can compensate for this issue to some extent, it is computationally very expensive. In this study, we propose a mixed-precision quantization algorithm without retraining or degradation in accuracy. In the proposed method, first, the difference between values after and before quantization losses of each channel in the layers of the pretrained model is calculated for all channels. Next, the layers are divided into two groups called semilayers according to whether the loss difference is positive or negative. The priorities for quantization in the semilayers are determined based on the Kulback-Leibler divergence derived from the probability distribution of the softmax output after and before quantization. The same process is repeated as a mixed-precision quantization while gradually decreasing the bitwidth, for example, with 8-, 6-, and 4-bit quantizations, and so forth. The results of an experimental evaluation show that the proposed method successfully compressed a ResNet-18 model by 81.44%, a ResNet-34 model by 84.25%, and a ResNet-50 model by 80.39% on image classification tasks using the ImageNet dataset, and a ResNet-18 model by 80.56% on image classification tasks using the CIFAR-10 dataset, with no degradation of the inference accuracy of the pretrained models.

Sequential maximization of expected improvement (EI) is one of the most widely used policies in Bayesian optimization because of its simplicity and ability to handle noisy observations. In particular, the improvement function often uses the best posterior mean as the best incumbent in noisy settings. However, the uncertainty associated with the incumbent solution is often neglected in many analytic EI-type methods: a closed-form acquisition function is derived in the noise-free setting, but then applied to the setting with noisy observations. To address this limitation, we propose a modification of EI that corrects its closed-form expression by incorporating the covariance information provided by the Gaussian Process (GP) model. This acquisition function specializes to the classical noise-free result, and we argue should replace that formula in Bayesian optimization software packages, tutorials, and textbooks. This enhanced acquisition provides good generality for noisy and noiseless settings. We show that our method achieves a sublinear convergence rate on the cumulative regret bound under heteroscedastic observation noise. Our empirical results demonstrate that our proposed acquisition function can outperform EI in the presence of noisy observations on benchmark functions for black-box optimization, as well as on parameter search for neural network model compression.