Pattern Recognition Letters最新文献

We explore a data-driven approach to generating neural network parameters to determine whether generative models can capture the underlying distribution of a collection of neural network checkpoints. We compile a dataset of checkpoints from neural networks trained within the multi-agent reinforcement learning framework, thus potentially producing previously unseen combinations of neural network parameters. In particular, our generative model is a conditional transformer-based variational autoencoder that, when provided with random noise and a specified performance metric – in our context, returns – predicts the appropriate distribution over the parameter space to achieve the desired performance metric. Our method successfully generates parameters for a specified optimal return without further fine-tuning. We also show that the parameters generated using this approach are more constrained and less variable and, most importantly, perform on par with those trained directly under the multi-agent reinforcement learning framework. We test our method on the neural network architectures commonly employed in the most advanced state-of-the-art algorithms.

The purpose of this paper is to present an unsupervised video anomaly detection method using Optical Flow decomposition and Spatio-Temporal feature learning (OFST). This method employs a combination of optical flow reconstruction and video frame prediction to achieve satisfactory results. The proposed OFST framework is composed of two modules: the Multi-Granularity Memory-augmented Autoencoder with Optical Flow Decomposition (MG-MemAE-OFD) and a Two-Stream Network based on Spatio-Temporal feature learning (TSN-ST). The MG-MemAE-OFD module is composed of three functional blocks: optical flow decomposition, autoencoder, and multi-granularity memory networks. The optical flow decomposition block is used to extract the main motion information of objects in optical flow, and the granularity memory network is utilized to memorize normal patterns and improve the quality of the reconstructions. To predict video frames, we introduce a two-stream network based on spatiotemporal feature learning (TSN-ST), which adopts parallel standard Transformer blocks and a temporal block to learn spatiotemporal features from video frames and optical flows. The OFST combines these two modules so that the prediction error of abnormal samples is further increased due to the larger reconstruction error. In contrast, the normal samples obtain a lower reconstruction error and prediction error. Therefore, the anomaly detection capability of the method is greatly enhanced. Our proposed model was evaluated on public datasets. Specifically, in terms of the area under the curve (AUC), our model achieved an accuracy of 85.74% on the Ped1 dataset, 99.62% on the Ped2 dataset, 93.89% on the Avenue dataset, and 76.0% on the ShanghaiTech Dataset. Our experimental results show an average improvement of 1.2% compared to the current state-of-the-art.

Handwritten mathematical expression recognition (HMER) is an appealing task due to its wide applications and research challenges. Previous deep learning-based methods used string decoder to emphasize on expression symbol awareness and achieved considerable recognition performance. However, these methods still meet an obstacle in recognizing handwritten symbols with varying appearance, in which huge appearance variations significantly lead to the ambiguity of symbol representation. To this end, our intuition is to employ printed expressions with unified appearance to serve as the template of handwritten expressions, alleviating the effects brought by varying symbol appearance. In this paper, we propose a contrastive learning method, where handwritten symbols with identical semantic are clustered together through the guidance of printed symbols, leading model to enhance the robustness of symbol semantic representations. Specifically, we propose an anchor generation scheme to obtain printed expression images corresponding with handwritten expressions. We propose a contrastive learning objective, termed Semantic-NCE Loss, to pull together printed and handwritten symbols with identical semantic. Moreover, we employ a string decoder to parse the calibrated semantic representations, outputting satisfactory expression symbols. The experiment results on benchmark datasets CROHME 14/16/19 demonstrate that our method noticeably improves recognition accuracy of handwritten expressions and outperforms the standard string decoder methods.

Since all training data is interpolated, interpolating classifiers have zero training error. However, recent work provides compelling reasons to investigate these classifiers, including their significance for ensemble methods. Interpolation kernel machines, which belong to the class of interpolating classifiers, are capable of good generalization and have proven to be an effective substitute for support vector machines, particularly for graph classification. In this work, we further enhance their performance by studying multiple kernel learning. To this end, we propose a general scheme of polynomial combined kernel functions, employing both quadratic and cubic kernel combinations in our experimental work. Our findings demonstrate that this approach improves performance compared to individual graph kernels. Our work supports the use of interpolation kernel machines as an alternative to support vector machines, thereby contributing to greater methodological diversity.

Few-shot classification is a challenging task to recognize unseen classes with limited data. Following the success of Vision Transformer in various large-scale datasets image recognition domains, recent few-shot classification methods employ transformer-style. However, most of them focus only on cross-attention between support and query sets, mainly considering channel-similarity. To address this issue, we introduce dual-similarity network (DSN) in which attention maps for the same target within a class are made identical. With the network, a way of effective training through the integration of the channel-similarity and the map-similarity has been sought. Our method, while focused on $N$-way $K$-shot scenarios, also demonstrates strong performance in 1-shot settings through augmentation. The experimental results verify the effectiveness of DSN on widely used benchmark datasets.

Owing to the advancements in deep learning, object detection has made significant progress in estimating the positions and classes of multiple objects within an image. However, detecting objects of various scales within a single image remains a challenging problem. In this study, we suggest a scale-aware token matching to predict the positions and classes of objects for transformer-based object detection. We train a model by matching detection tokens with ground truth considering its size, unlike the previous methods that performed matching without considering the scale during the training process. We divide one detection token set into multiple sets based on scale and match each token set differently with ground truth, thereby, training the model without additional computation costs. The experimental results demonstrate that scale information can be assigned to tokens. Scale-aware tokens can independently learn scale-specific information by using a novel loss function, which improves the detection performance on small objects.

In cross-modal retrieval, most existing hashing-based methods merely considered the relationship between feature representations to reduce the heterogeneous gap for data from various modalities, whereas they neglected the correlation between feature representations and the corresponding labels. This leads to the loss of significant semantic information, and the degradation of the class discriminability of the model. To tackle these issues, this paper presents a novel cross-modal retrieval method called coding self-representative and label-relaxed hashing (CSLRH) for cross-modal retrieval. Specifically, we propose a self-representation learning term to enhance the class-specific feature representations and reduce the noise interference. Additionally, we introduce a label-relaxed regression to establish semantic relations between the hash codes and the label information, aiming to enhance the semantic discriminability. Moreover, we incorporate a non-linear regression to capture the correlation of non-linear features in hash codes for cross-modal retrieval. Experimental results on three widely-used datasets verify the effectiveness of our proposed method, which can generate more discriminative hash codes to improve the precisions of cross-modal retrieval.

The popular softmax loss and its recent extensions have achieved great success in deep learning-based image classification. However, the data for training image classifiers often exhibit a highly skewed distribution in quality, i.e., the number of data with good quality is much more than that with low quality. If this problem is ignored, low-quality data are hard to classify correctly. In this paper, we discover the positive correlation between the quality of an image and its feature norm ($L_2$-norm) learned from softmax loss through careful experiments on various applications with different deep neural networks. Based on this finding, we propose a contraction mapping function to compress the range of feature norms of training images according to their quality and embed this contraction mapping function into softmax loss and its extensions to produce novel learning objectives. Experiments on various applications, including handwritten digit recognition, lung nodule classification, and face recognition, demonstrate that the proposed approach is promising to effectively deal with the problem of learning quality imbalance data and leads to significant and stable improvements in the classification accuracy. The code is available at https://github.com/Huiyu-Li/CM-M-Softmax-Loss.

Training a large set of machine learning algorithms to convergence in order to select the best-performing algorithm for a dataset is computationally wasteful. Moreover, in a budget-limited scenario, it is crucial to carefully select an algorithm candidate and allocate a budget for training it, ensuring that the limited budget is optimally distributed to favor the most promising candidates. Casting this problem as a Markov Decision Process, we propose a novel framework in which an agent must select in the process of learning the most promising algorithm without waiting until it is fully trained. At each time step, given an observation of partial learning curves of algorithms, the agent must decide whether to allocate resources to further train the most promising algorithm (exploitation), to wake up another algorithm previously put to sleep, or to start training a new algorithm (exploration). In addition, our framework allows the agent to meta-learn from learning curves on past datasets along with dataset meta-features and algorithm hyperparameters. By incorporating meta-learning, we aim to avoid myopic decisions based solely on premature learning curves on the dataset at hand. We introduce two benchmarks of learning curves that served in international competitions at WCCI’22 and AutoML-conf’22, of which we analyze the results. Our findings show that both meta-learning and the progression of learning curves enhance the algorithm selection process, as evidenced by methods of winning teams and our DDQN baseline, compared to heuristic baselines or a random search. Interestingly, our cost-effective baseline, which selects the best-performing algorithm w.r.t. a small budget, can perform decently when learning curves do not intersect frequently.