ACML 2023 论文列表

Through ground-breaking advances in Machine Learning its real-world applications have become commonplace in many areas over the past decade. Deep and complex models are able to solve difficult tasks with super-human precision. But for manufacturing quality control, in theory a ideal match for these methods, the step from proof-of-concept towards live deployment is often not feasible. One major obstacle is the unreliability of Machine Learning predictions when confronted with data diverging from the known characteristics. While overall accuracy is high, wrong results may be returned with no indication of their uncertainty. In manufacturing, where scarce errors mean great damages, additional safety measures are required. In this work, I present Visual Machine Learning Control (VMLC), an approach developed upon a real world visual quality control system that operates in a high throughput manufacturing line. Instead of applying sole classification or anomaly detection, both is done in combination. A scalar metric derived from an Auto-Encoder reconstruction error measures the compliance of captured images with the training data the system is trained on. This metric is integrated into the widely used framework of industrial Statistical Process Control (SPC), significantly increasing robustness through meaningful control limits and enabling active learning. The system is evaluated on a large dataset of real-world industrial welding images.

Amazon ships billions of packages to its customers annually within the United States. Shipping cost of these packages are used on the day of shipping (day 0) to estimate profitability of sales. Downstream systems utilize these days 0 profitability estimates to make financial decisions, such as pricing strategies and delisting loss-making products. However, obtaining accurate shipping cost estimates on day 0 is complex for reasons like delay in carrier invoicing or fixed cost components getting recorded at monthly cadence. Inaccurate shipping cost estimates can lead to bad decision, such as pricing items too low or high, or promoting the wrong product to the customers. Current solutions for estimating shipping costs on day 0 rely on tree-based models that require extensive manual engineering efforts. In this study, we propose a novel architecture called the Rate Card Transformer (RCT) that uses self-attention to encode all package shipping information such as package attributes, carrier information and route plan. Unlike other transformer-based tabular models, RCT has the ability to encode a variable list of one-to-many relations of a shipment, allowing it to capture more information about a shipment. For example, RCT can encode properties of all products in a package. Our results demonstrate that cost predictions made by the RCT have 28.82% less error compared to tree-based GBDT model. Moreover, the RCT outperforms the state-of-the-art transformer-based tabular model, FTTransformer, by 6.08%. We also illustrate that the RCT learns a generalized manifold of the rate card that can improve the performance of tree-based models.

Recent Continuous Graph Neural Networks (CGNNs) have attracted great attention due to its merits of infinite depth without oversmoothing. However, most of the existing CGNNs perform low-pass filtering in nature, as they are derived from discrete Laplacian-smoothing based graph neural networks (GNNs). While prior research has shown the promising results of high-pass filtering for node representation learning, particularly on heterophilous graphs, there remains a need to extend it to continuous domain and explore the synergy between two filtering channels. In this paper, by leveraging low-pass and high-pass filtering, we propose a novel dual-channel continuous graph neural network architecture to address this gap. In particular, we introduce a dimension masking method to coordinate the contribution of all low and high pass filtered feature dimensions to node classification. Our aim is to deepen the understanding of the link between high and low filters, unraveling their distinct roles in learning node representations. To evaluate the effectiveness of our framework, we conduct extensive experiments focusing on the node classification task of heterophilous graphs. Our results demonstrate the competitive performance of our approach, showcasing its robustness to oversmoothing.

Relation extraction plays a vital role in knowledge graph construction. In contrast with the traditional relation extraction on a single sentence, extracting relations from multiple sentences as a whole will harvest more valuable and richer knowledge. Recently, the Transformer-based pre-trained language models (TPLMs) are widely adopted to tackle document-level relation extraction (DocRE). Graph-based methods, aiming to acquire knowledge between entities to form entity-level relation graphs, have facilitated the rapid development of DocRE by infusing their proposed models with the knowledge. However, beyond entity-level knowledge, we discover many other kinds of knowledge that can aid humans to extract relations. It remains unclear whether and in which way they can be adopted to improve the performance of the Transformer, which affects the maximum performance gain of Transformer-based methods. In this paper, we propose a novel weighted multi-channel Transformer (WMCT) to infuse unlimited kinds of knowledge into the vanilla Transformer. Based on WMCT, we also explore five kinds of knowledge to enhance both its reasoning ability and expressive power. Our extensive experimental results demonstrate that: (1) more knowledge makes the performance of the Transformer better and (2) more informative knowledge leads to more performance gain. We appeal to future Transformer-based work to consider exploring more informative knowledge to improve the performance of the Transformer.

Mathematical models of observational phenomena are at the core of experimental sciences. By learning the parameters of such models from typically noisy observations, we can interpret and predict the phenomena under investigation. This process, however, assumes that the model itself is correct and that we are only uncertain of its parameters. In practice, this is rarely true, but rather the model is a simplification of the actual generative process. One proposed remedy is a post hoc investigation of how the model differs from reality, by explicitly modeling the discrepancy between the two. In this paper, we use transformed Gaussian processes as flexible models for this. Our formulation relaxes the assumption on the correctness of the model by assuming it is only correct in expectation, and it directly supports both additive and multiplicative corrections, treated separately in the literature, using suitable transformations. We demonstrate the approach in two example cases: modeling human growth (relation age-height) and modeling the risk attitude (relation reward-utility). The former provides a simple example, while the second case highlights the importance of the transformations in obtaining meaningful information about the discrepancy.

Spiking Neural Networks (SNNs) that operate in an event-driven manner and employ binary spike representation have recently emerged as promising candidates for energy-efficient computing. However, a cost bottleneck arises in obtaining high-performance SNNs: training a SNN model requires a large number of time steps in addition to the usual learning iterations, hence this limits their energy efficiency. This paper proposes a general training framework that enhances feature learning and activation efficiency within a limited time step, providing a new solution for more energy-efficient SNNs. Our framework allows SNN neurons to learn robust spike feature from different receptive fields and update neuron states by utilizing both current stimuli and recurrence information transmitted from other neurons. This setting continuously complements information within a single time step. Additionally, we propose a projection function to merge these two stimuli to smoothly optimize neuron weights (spike firing threshold and activation). We evaluate the proposal for both convolution and recurrent models. Our experimental results indicate state-of-the-art visual classification tasks, including CIFAR10, CIFAR100, and TinyImageNet. Our framework achieves 72.41% and 72.31% top-1 accuracy with only 1 time step on CIFAR100 for CNNs and RNNs, respectively. Our method reduces 10X and 3X joule energy than a standard ANN and SNN, respectively, on CIFAR10, without additional time steps.

In this paper, we propose a new approach called Adaptive Behavioral Costs in Reinforcement Learning (ABC-RL) for training a human-like agent with competitive strength. While deep reinforcement learning agents have recently achieved superhuman performance in various video games, some of these unconstrained agents may exhibit actions, such as shaking and spinning, that are not typically observed in human behavior, resulting in peculiar gameplay experiences. To behave like humans and retain similar performance, ABC-RL augments behavioral limitations as cost signals in reinforcement learning with dynamically adjusted weights. Unlike traditional constrained policy optimization, we propose a new formulation that minimizes the behavioral costs subject to a constraint of the value function. By leveraging the augmented Lagrangian, our approach is an approximation of the Lagrangian adjustment, which handles the trade-off between the performance and the human-like behavior. Through experiments conducted on 3D games in DMLab-30 and Unity ML-Agents Toolkit, we demonstrate that ABC-RL achieves the same performance level while significantly reducing instances of shaking and spinning. These findings underscore the effectiveness of our proposed approach in promoting more natural and human-like behavior during gameplay.

Recent advances in Artificial Intelligence (AI) technology have promoted their use in almost every field. The growing complexity of deep neural networks (DNNs) makes it increasingly difficult and important to explain the inner workings and decisions of the network. However, most current techniques for explaining DNNs focus mainly on interpreting classification tasks. This paper proposes a method to explain the decision for any object detection model called D-CLOSE. To closely track the model’s behavior, we used multiple levels of segmentation on the image and a process to combine them. We performed tests on the MS-COCO dataset with the YOLOX model, which shows that our method outperforms D-RISE and can give a better quality and less noise explanation.

This paper studies the fixed-confidence best arm identification (BAI) problem in the bandit framework in the canonical single-parameter exponential models. For this problem, many policies have been proposed, but most of them require solving an optimization problem at every round and/or are forced to explore an arm at least a certain number of times except those restricted to the Gaussian model. To address these limitations, we propose a novel policy that combines Thompson sampling with a computationally efficient approach known as the best challenger rule. While Thompson sampling was originally considered for maximizing the cumulative reward, we demonstrate that it can be used to naturally explore arms in BAI without forcing it. We show that our policy is asymptotically optimal for any two-armed bandit problems and achieves near optimality for general $K$-armed bandit problems for $K \geq 3$. Nevertheless, in numerical experiments, our policy shows competitive performance compared to asymptotically optimal policies in terms of sample complexity while requiring less computation cost. In addition, we highlight the advantages of our policy by comparing it to the concept of $\beta$-optimality, a relaxed notion of asymptotic optimality commonly considered in the analysis of a class of policies including the proposed one.

Tiny object detection has gained considerable attention in the research community owing to the frequent occurrence of tiny objects in numerous critical real-world scenarios. However, convolutional neural networks (CNNs) used as the backbone for object detection architectures typically neglect Nyquist’s sampling theorem during down-sampling operations, resulting in aliasing and degraded performance. This is likely to be a particular issue for tiny objects that occupy very few pixels and therefore have high spatial frequency features. This paper applied an existing approach WaveCNet for anti-aliasing to tiny object detection. WaveCNet addresses aliasing by replacing standard down-sampling processes in CNNs with Wavelet Pooling (WaveletPool) layers, effectively suppressing aliasing. We modify the original WaveCNet to apply WaveletPool in a consistent way in both pathways of the residual blocks in ResNets. Additionally, we also propose a bottom-heavy version of the backbone, which further improves the performance of tiny object detection while also reducing the required number of parameters by almost half. Experimental results on the TinyPerson, WiderFace, and DOTA datasets demonstrate the importance of anti-aliasing in tiny object detection and the effectiveness of the proposed method which achieves new state-of-the-art results on all three datasets. Codes and experiment results are released at \url{https://github.com/freshn/Anti-aliasing-Tiny-Object-Detection.git}.

We conduct a detailed investigation of tempered posteriors and uncover a number of crucial and previously undiscussed points. Contrary to previous results, we first show that for realistic models and datasets and the tightly controlled case of the Laplace approximation to the posterior, stochasticity does not in general improve test accuracy. The coldest temperature is often optimal. One might think that Bayesian models with some stochasticity can at least obtain improvements in terms of calibration. However, we show empirically that when gains are obtained this comes at the cost of degradation in test accuracy. We then discuss how targeting Frequentist metrics using Bayesian models provides a simple explanation of the need for a temperature parameter $\lambda$ in the optimization objective. Contrary to prior works, we finally show through a PAC-Bayesian analysis that the temperature $\lambda$ cannot be seen as simply fixing a misspecified prior or likelihood.

Falls are a major cause of injuries and deaths among older adults worldwide. Accurate fall detection can help reduce potential injuries and additional health complications. Different types of video modalities can be used in a home setting to detect falls, including RGB, Infrared, and Thermal cameras. Anomaly detection frameworks using autoencoders and their variants can be used for fall detection due to the data imbalance that arises from the rarity and diversity of falls. However, the use of reconstruction error in autoencoders can limit the application of networks’ structures that propagate information. In this paper, we propose a new multi-objective loss function called Temporal Shift, which aims to predict both future and reconstructed frames within a window of sequential frames. The proposed loss function is evaluated on a semi-naturalistic fall detection dataset containing multiple camera modalities. The autoencoders were trained on normal activities of daily living (ADL) performed by older adults and tested on ADLs and falls performed by young adults. Temporal shift shows significant improvement to a baseline 3D Convolutional autoencoder, an attention U-Net CAE, and a multi-modal neural network. The greatest improvement was observed in an attention U-Net model improving by $0.20$ AUC ROC for a single camera when compared to reconstruction alone. With significant improvement across different models, this approach has the potential to be widely adopted and improve anomaly detection capabilities in other settings besides fall detection.

Weakly-Supervised Temporal Action Localization (WSTAL) aims to train an action instance localization model from untrimmed videos with only video-level labels, similar to the Object Detection (OD) task. Existing Top-k MIL-based WSTAL methods cannot flexibly define the learning space, which limits the model’s learning efficiency and performance. Faster R-CNN is a classic two-stage object detection architecture with an efficient Region Proposal Network. This paper successfully migrates the Faster R-CNN liked two-stage architecture to the WSTAL task: first to build a T-RPN and integrate it with the traditional WSTAL framework; and then to propose a pseudo label generation mechanism to enable the T-RPN learning without temporal annotations. Our new framework has achieved breakthrough performances on THUMOS-14 and ActivityNet-v1.2 datasets, and comprehensive ablation experiments have verified the effectiveness of the innovations. Code will be available at: \href{https://github.com/ZJUHJ/TRPN}{https://github.com/ZJUHJ/TRPN}.

Trees play a very important role in maintaining the ecosystem. To automate the process of extracting trees from Remote Sensing data and counting thereby, we have used Cartosat-2S merged products @ 0.6m as input. Vegetation indices such as NDVI, and EVI can retrieve the Vegetation class which contains trees and their look-alikes (such as grass, fields, and shrubs) having the same spectral signatures \em{i.e.}, high reflectance in NIR band, and high absorption in Red band. Extraction of only Trees is not possible with these indices. In this paper we present a novel method for tree delineation from its look-alikes using AIML, concentrating more on preparation of input datasets efficiently. Based on the Spectral Separability analysis, only Red and NIR bands from the satellite imagery are utilized in the proposed method to separate Vegetation from the background classes (such as water, bare soil, terrain, and built-up). In addition to the two bands from satellite imagery entropy layer is computed from the NIR band and utilized as the third band to delineate trees from their look-alikes. Deep Neural Networks have the capability of learning complex patterns that can separate trees and their look-alikes. However, the performance is boosted when the entropy layer is added to the input image. The proposed method showed better performance when utilizing 3 bands Red, NIR, and Entropy bands compared to 4 bands i.e. Red, Green, Blue, and NIR bands. The proposed method obtains a precision of 96%, a recall of 90%, and an F1-score of 93% even with a relatively smaller training dataset. The study is performed on the data collected from various locations of the Indian state Rajasthan.

Air pollution raises a detrimental impact on human health and natural environment. Accurate prediction of air quality is crucial for effective pollution control and mitigation strategies. Numerous existing methods for analyzing the variation tendency of a specific air component primarily focus on its temporal and spatial information, neglecting the potential interactions between different attributes within the same time interval. In this paper, we propose a Temporal-Feature correlations Attention-based deep learning Network (TFAN), which incorporates data fusion technology. TFAN focuses on capturing temporal dependencies, feature correlations, and the potential relationship between temporal-feature through the Attention mechanism, and the data fusion method allows for a comprehensive consideration of multiple factors on prediction. Experimental results conducted using real-world data from Beijing City demonstrate that TFAN outperforms various baseline models in prediction accuracy for multiple pollutants by 10+%.

Recently, low-resource dialogue state tracking (DST) has received increasing attention. First obtaining state values then based on values to generate slot types has made great progress in this task. However, obtaining state values is still an under-studied problem. Existing extraction-based approaches cannot capture values that require the understanding of context and are not generalizable either. To address these issues, we propose a novel State VAlue Generation based framework (SVAG), decomposing DST into state value generation and domain slot generation. Specifically, we propose to generate state values and use self-training to further improve state value generation. Moreover, we design an estimator aiming at detecting incomplete generation and incorrect generation for pseudo-labeled data selection during self-training. Experimental results on the MultiWOZ 2.1 dataset show that our method which has only less than 1 billion parameters achieves state-of-the-art performance under the data ratio settings of 5%, 10%, and 25% when limited to models under 100 billion parameters. Compared to models with more than 100 billion parameters, SVAG still reaches competitive results.

With the development of deep learning, breakthroughs in single image super-resolution have been achieved. However, most existing methods are limited to using only spatial domain information or only frequency domain information, and the rich information of the image in the frequency domain space is not fully utilized, so it is still difficult to recover satisfactory texture details. In this paper, we propose a method to fuse the frequency domain and spatial domain information. Our method uses a two-branch network to extract the spatial domain information and the frequency domain information separately and uses a fusion module to fuse the different information in the two domains. We also use the Non-Subsampled Shearlet Transform (NSST) to preserve the texture directionality well, and design two NSST-based directional texture enhancement modules, which are embedded in different parts of the network, to enhance the recovery of texture details in the image reconstruction process. Quantitative and qualitative experimental results show that the method outperforms advanced single-image super-resolution methods in recovering images.

Utilizing a pretrained vision backbone to finetune a model for semantic segmentation is common practice in computer vision. However, there are few works intending to enlarge the semantic context learning capacity by incorporating a segmentation adapter into the backbone. Thus, in this paper, we present a simple but efficient segmentation adapter, termed as SegAdapter, which can be plugged into the pretrained vision backbone to improve the performance of existing models for image semantic segmentation. We summarize SegAdapter with three attractive advantages: 1) SegAdapter is a plug-and-play module demonstrating strong adaptability in CNN and Transformer based models such as ConvNext and Segformer, 2) SegAdapter applies a light-weight High-order Spatial Attention (HSA) to make use of intermediate features from the pretrained backbone which extends the model depth and produces auxiliary segmentation maps for model enhancement, 3) SegAdapter builds a powerful vision backbone by incorporating the semantic context into each stage which takes on some of the functions of the segmentation head. So, SegAdapter augmented model can be used in simple designed decode head to avoid heavy computational cost. By plugging multiple SegAdapter layers into different vision backbones, we construct a series of SegAdapter-based segmentation models. We show through the extensive experiments that SegAdapter can be used with mainstream backbones like CNN and Transformer to improve mIoU performance in a large margin while introducing minimal additional parameters and FLOPs.

Our research demonstrates the significant benefits of using fine-tuning with explanations to enhance the performance of language models. Unlike prompting, which maintains the model’s parameters, fine-tuning allows the model to learn and update its parameters during a training phase. In this study, we applied fine-tuning to various sized language models using data that contained explanations of the output rather than merely presenting the answers. We found that even smaller language models with as few as 60 million parameters benefited substantially from this approach. Interestingly, our results indicated that the detailed explanations were more beneficial to smaller models than larger ones, with the latter gaining nearly the same advantage from any form of explanation, irrespective of its length. Additionally, we demonstrate that the inclusion of explanations enables the models to solve tasks that they were not able to solve without explanations. Lastly, we argue that despite the challenging nature of adding explanations, samples that contain explanations not only reduce the volume of data required for training but also promote a more effective generalization by the model. In essence, our findings suggest that fine-tuning with explanations significantly bolsters the performance of large language models.