查看更多>>摘要：Nuclei segmentation and classification of hematoxylin and eosin-stained histology images is a challenging task due to a variety of issues, such as color inconsistency that results from the non-uniform manual staining operations, clustering of nuclei, and blurry and overlapping nuclei boundaries. Existing approaches involve segmenting nuclei by drawing their polygon representations or by measuring the distances between nuclei centroids. In contrast, we leverage the fact that morphological features (appearance, shape, and texture) of nuclei in a tissue vary greatly depending upon the tissue type. We exploit this information by extracting tissue specific (TS) features from raw histopathology images using the proposed tissue specific feature distillation (TSFD) backbone. The bi-directional feature pyramid network (BiFPN) within TSFD-Net generates a robust hierarchical feature pyramid utilizing TS features where the interlinked decoders jointly optimize and fuse these features to generate final predictions. We also propose a novel combinational loss function for joint optimization and faster convergence of our proposed network. Extensive ablation studies are performed to validate the effectiveness of each component of TSFD-Net. The proposed network outperforms state-of-the-art networks such as StarDist, Micro-Net, Mask-RCNN, Hover-Net, and CPP-Net on the PanNuke dataset, which contains 19 different tissue types and 5 clinically important tumor classes, achieving 50.4% and 63.77% mean and binary panoptic quality, respectively. The code is available at: https://github.com/MrTalhaIlyas/TSFD. (c) 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
查看更多>>摘要：Recent theoretical and experimental works have connected Hebbian plasticity with the reinforcement learning (RL) paradigm, producing a class of trial-and-error learning in artificial neural networks known as neo-Hebbian plasticity. Inspired by the role of the neuromodulator dopamine in synaptic modification, neo-Hebbian RL methods extend unsupervised Hebbian learning rules with value-based modulation to selectively reinforce associations. This reinforcement allows for learning exploitative behaviors and produces RL models with strong biological plausibility. The review begins with coverage of fundamental concepts in rate-and spike-coded models. We introduce Hebbian correlation detection as a basis for modification of synaptic weighting and progress to neo-Hebbian RL models guided solely by extrinsic rewards. We then analyze state-of-the-art neo-Hebbian approaches to the exploration- exploitation balance under the RL paradigm, emphasizing works that employ additional mechanics to modulate that dynamic. Our review of neo-Hebbian RL methods in this context indicates substantial potential for novel improvements in exploratory learning, primarily through stronger incorporation of intrinsic motivators. We provide a number of research suggestions for this pursuit by drawing from modern theories and results in neuroscience and psychology. The exploration-exploitation balance is a central issue in RL research, and this review is the first to focus on it under the neo-Hebbian RL framework. (C) 2022 Elsevier Ltd. All rights reserved
查看更多>>摘要：Rapid advancements in deep learning have led to many recent breakthroughs. While deep learning models achieve superior performance, often statistically better than humans, their adoption into safety critical settings, such as healthcare or self-driving cars is hindered by their inability to provide safety guarantees or to expose the inner workings of the model in a human understandable form. We present MoET, a novel model based on Mixture of Experts, consisting of decision tree experts and a generalized linear model gating function. Thanks to such gating function the model is more expressive than the standard decision tree. To support non-differentiable decision trees as experts, we formulate a novel training procedure. In addition, we introduce a hard thresholding version, MoETh, in which predictions are made solely by a single expert chosen via the gating function. Thanks to that property, MoETh allows each prediction to be easily decomposed into a set of logical rules in a form which can be easily verified. While MoET is a general use model, we illustrate its power in the reinforcement learning setting. By training MoET models using an imitation learning procedure on deep RL agents we outperform the previous state-of-the-art technique based on decision trees while preserving the verifiability of the models. Moreover, we show that MoET can also be used in real-world supervised problems on which it outperforms other verifiable machine learning models. (c) 2022 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Recent research has used margin theory to analyze the generalization performance for deep neural networks (DNNs). The existed results are almost based on the spectrally-normalized minimum margin. However, optimizing the minimum margin ignores a mass of information about the entire margin distribution, which is crucial to generalization performance. In this paper, we prove a generalization upper bound dominated by the statistics of the entire margin distribution. Compared with the minimum margin bounds, our bound highlights an important measure for controlling the complexity, which is the ratio of the margin standard deviation to the expected margin. We utilize a convex margin distribution loss function on the deep neural networks to validate our theoretical results by optimizing the margin ratio. Experiments and visualizations confirm the effectiveness of our approach and the correlation between generalization gap and margin ratio. (c) 2022 Elsevier Ltd. All rights reserved.
Yang, YejiangWang, TaoWoolard, Jefferson P.Xiang, Weiming...
9页
查看更多>>摘要：Approximation error is a key measure in the process of model validation and verification for neural networks. In this paper, the problems of guaranteed error estimation of neural networks and applications to assured system modeling and assured neural network compression are addressed. First, a concept called guaranteed error estimation of feedforward neural networks is proposed, which intends to provide the worst-case approximation error of a trained neural network with respect to a compact input set essentially containing an infinite number of values. Given different prior information about the original system, two approaches including Lipschitz constant analysis and set-valued reachability analysis methods are developed to efficiently compute upper-bounds of approximation errors. Based on the guaranteed approximation error estimation framework, an optimization for obtaining parameter values from data set is proposed. A robotic arm and neural network compression examples are presented to illustrate the effectiveness of our approach.
查看更多>>摘要：Graph classification aims to predict the property of the whole graph, which has attracted growing attention in the graph learning community. This problem has been extensively studied in the literature of both graph convolutional networks and graph kernels. Graph convolutional networks can learn effective node representations via message passing to mine graph topology in an implicit way, whereas graph kernels can explicitly utilize graph structural knowledge for classification. Due to the scarcity of labeled data in real-world applications, semi-supervised algorithms are anticipated for this problem. In this paper, we propose Graph Harmonic Neural Network (GHNN) which combines the advantages of both worlds to sufficiently leverage the unlabeled data, and thus overcomes label scarcity in semi-supervised scenarios. Specifically, our GHNN consists of a graph convolutional network (GCN) module and a graph kernel network (GKN) module that explore graph topology information from complementary perspectives. To fully leverage the unlabeled data, we develop a novel harmonic contrastive loss and a harmonic consistency loss to harmonize the training of two modules by giving priority to high-quality unlabeled data, thereby reconciling prediction consistency between both of them. In this manner, the two modules mutually enhance each other to sufficiently explore the graph topology of both labeled and unlabeled data. Extensive experiments on a variety of benchmarks demonstrate the effectiveness of our approach over competitive baselines. (c) 2022 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Complex-valued neural networks have attracted increasing attention in recent years, while it remains open on the advantages of complex-valued neural networks in comparison with real-valued networks. This work takes one step on this direction by introducing the complex-reaction network with fully-connected feed-forward architecture. We prove the universal approximation property for complex-reaction networks, and show that a class of radial functions can be approximated by a complex-reaction network using the polynomial number of parameters, whereas real-valued networks need at least exponential parameters to reach the same approximation level. For empirical risk minimization, we study the landscape and convergence of complex gradient descents. Our theoretical result shows that the critical point set of complex-reaction networks is a proper subset of that of real-valued networks, which may show some insights on finding the optimal solutions more easily for complex-reaction networks. (C)& nbsp;2022 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Performing pattern analysis over the symmetric positive definite (SPD) manifold requires specific mathematical computations, characterizing the non-Euclidian property of the involved data points and learning tasks, such as the image set classification problem. Accompanied with the advanced neural networking techniques, several architectures for processing the SPD matrices have recently been studied to obtain fine-grained structured representations. However, existing approaches are challenged by the diversely changing appearance of the data points, begging the question of how to learn invariant representations for improved performance with supportive theories. Therefore, this paper designs two Riemannian operation modules for SPD manifold neural network. Specifically, a Riemannian batch regularization (RBR) layer is firstly proposed for the purpose of training a discriminative manifold-to-manifold transforming network with a novelly-designed metric learning regularization term. The second module realizes the Riemannian pooling operation with geometric computations on the Riemannian manifolds, notably the Riemannian barycenter, metric learning, and Riemannian optimization. Extensive experiments on five benchmarking datasets show the efficacy of the proposed approach.(C)& nbsp; 2022 Published by Elsevier Ltd.
查看更多>>摘要：Electroencephalographic measurement of cortical activity subserving motor behavior varies among different individuals, restricting the potential of brain computer interfaces (BCIs) based on motor imagery (MI). How to deal with this variability and thereby improve the accuracy of BCI classification remains a key issue. This paper proposes a deep learning-based approach to transfer the data distribution from BCI-friendly - "golden subjects"to the data from more typical BCI-illiterate users. In this work, we use the perceptual loss to align the dimensionality-reduced BCI-illiterate data with the data of golden subjects in low dimensions, by which a subject transfer neural network (STNN) is proposed. The network consists of two parts: 1) a generator, which generates the transferred BCIilliterate features, and 2) a CNN classifier, which is used for the classification of the transferred features, thus outperforming traditional classification methods both in terms of accuracy and robustness. Electroencephalography (EEG) signals from 25 healthy subjects performing MI of the right hand and foot were classified with an average accuracy of 88.2% +/- 5.1%. The proposed model was further validated on the BCI Competition IV dataset 2b, and was demonstrated to be robust to inter-subject variations. The advantages of STNN allow it to bridge the gap between the golden subjects and the BCI-illiterate ones, paving the way to real-time BCI applications. (c) 2022 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Despite considerable progress in the field of automatic multi-target tracking, several problems such as data association remained challenging. On the other hand, cognitive studies have reported that humans can robustly track several objects simultaneously. Such circumstances happen regularly in daily life, and humans have evolved to handle the associated problems. Accordingly, using braininspired processing principles may contribute to significantly increase the performance of automatic systems able to follow the trajectories of multiple objects. In this paper, we propose a multipleobject tracking algorithm based on dynamic neural field theory which has been proven to provide neuro-plausible processing mechanisms for cognitive functions of the brain. We define several input neural fields responsible for representing previous location and orientation information as well as instantaneous linear and angular speed of the objects in successive video frames. Image processing techniques are applied to extract the critical object features including target location and orientation. Two prediction fields anticipate the objects' locations and orientations in the upcoming frame after receiving excitatory and inhibitory inputs from the input fields in a feed-forward architecture. This information is used in the data association and labeling process. We tested the proposed algorithm on a zebrafish larvae segmentation and tracking dataset and an ant-tracking dataset containing non-rigid objects with spiky movements and frequently occurring occlusions. The results showed a significant improvement in tracking metrics compared to state-of-the-art algorithms.(c) 2022 Elsevier Ltd. All rights reserved.
NSTL
Elsevier