In NLP, text classification is one of the primary problems we try to solve and its uses in language analyses are indisputable. The lack of labeled training data made it harder to do these tasks in low resource languages like Amharic. The task of collecting, labeling, annotating, and making valuable this kind of data will encourage junior researchers, schools, and machine learning practitioners to implement existing classification models in their language. In this short paper, we aim to introduce the Amharic text classification dataset that consists of more than 50k news articles that were categorized into 6 classes. This dataset is made available with easy baseline performances to encourage studies and better performance experiments.
Fuzzy Cognitive Maps (FCMs) are considered a soft computing technique combining elements of fuzzy logic and recurrent neural networks. They found multiple application in such domains as modeling of system behavior, prediction of time series, decision making and process control. Less attention, however, has been turned towards using them in pattern classification. In this work we propose an FCM based classifier with a fully connected map structure. In contrast to methods that expect reaching a steady system state during reasoning, we chose to execute a few FCM iterations (steps) before collecting output labels. Weights were learned with a gradient algorithm and logloss or cross-entropy were used as the cost function. Our primary goal was to verify, whether such design would result in a descent general purpose classifier, with performance comparable to off the shelf classical methods. As the preliminary results were promising, we investigated the hypothesis that the performance of $d$-step classifier can be attributed to a fact that in previous $d-1$ steps it transforms the feature space by grouping observations belonging to a given class, so that they became more compact and separable. To verify this hypothesis we calculated three clustering scores for the transformed feature space. We also evaluated performance of pipelines built from FCM-based data transformer followed by a classification algorithm. The standard statistical analyzes confirmed both the performance of FCM based classifier and its capability to improve data. The supporting prototype software was implemented in Python using TensorFlow library.
In this paper, we propose a novel data augmentation strategy named Cut-Thumbnail, that aims to improve the shape bias of the network. We reduce an image to a certain size and replace the random region of the original image with the reduced image. The generated image not only retains most of the original image information but also has global information in the reduced image. We call the reduced image as thumbnail. Furthermore, we find that the idea of thumbnail can be perfectly integrated with Mixed Sample Data Augmentation, so we put one image's thumbnail on another image while the ground truth labels are also mixed, making great achievements on various computer vision tasks. Extensive experiments show that Cut-Thumbnail works better than state-of-the-art augmentation strategies across classification, fine-grained image classification, and object detection. On ImageNet classification, ResNet-50 architecture with our method achieves 79.21\% accuracy, which is more than 2.8\% improvement on the baseline.
Deep object recognition models have been very successful over benchmark datasets such as ImageNet. How accurate and robust are they to distribution shifts arising from natural and synthetic variations in datasets? Prior research on this problem has primarily focused on ImageNet variations (e.g., ImageNetV2, ImageNet-A). To avoid potential inherited biases in these studies, we take a different approach. Specifically, we reanalyze the ObjectNet dataset recently proposed by Barbu et al. containing objects in daily life situations. They showed a dramatic performance drop of the state of the art object recognition models on this dataset. Due to the importance and implications of their results regarding the generalization ability of deep models, we take a second look at their analysis. We find that applying deep models to the isolated objects, rather than the entire scene as is done in the original paper, results in around 20-30% performance improvement. Relative to the numbers reported in Barbu et al., around 10-15% of the performance loss is recovered, without any test time data augmentation. Despite this gain, however, we conclude that deep models still suffer drastically on the ObjectNet dataset. We also investigate the robustness of models against synthetic image perturbations such as geometric transformations (e.g., scale, rotation, translation), natural image distortions (e.g., impulse noise, blur) as well as adversarial attacks (e.g., FGSM and PGD-5). Our results indicate that limiting the object area as much as possible (i.e., from the entire image to the bounding box to the segmentation mask) leads to consistent improvement in accuracy and robustness.
In general, graph representation learning methods assume that the train and test data come from the same distribution. In this work we consider an underexplored area of an otherwise rapidly developing field of graph representation learning: The task of out-of-distribution (OOD) graph classification, where train and test data have different distributions, with test data unavailable during training. Our work shows it is possible to use a causal model to learn approximately invariant representations that better extrapolate between train and test data. Finally, we conclude with synthetic and real-world dataset experiments showcasing the benefits of representations that are invariant to train/test distribution shifts.
Weakly supervised object localization(WSOL) remains an open problem given the deficiency of finding object extent information using a classification network. Although prior works struggled to localize objects through various spatial regularization strategies, we argue that how to extract object structural information from the trained classification network is neglected. In this paper, we propose a two-stage approach, termed structure-preserving activation (SPA), toward fully leveraging the structure information incorporated in convolutional features for WSOL. First, a restricted activation module (RAM) is designed to alleviate the structure-missing issue caused by the classification network on the basis of the observation that the unbounded classification map and global average pooling layer drive the network to focus only on object parts. Second, we designed a post-process approach, termed self-correlation map generating (SCG) module to obtain structure-preserving localization maps on the basis of the activation maps acquired from the first stage. Specifically, we utilize the high-order self-correlation (HSC) to extract the inherent structural information retained in the learned model and then aggregate HSC of multiple points for precise object localization. Extensive experiments on two publicly available benchmarks including CUB-200-2011 and ILSVRC show that the proposed SPA achieves substantial and consistent performance gains compared with baseline approaches.Code and models are available at https://github.com/Panxjia/SPA_CVPR2021
We aim at accelerating super-resolution (SR) networks on large images (2K-8K). The large images are usually decomposed into small sub-images in practical usages. Based on this processing, we found that different image regions have different restoration difficulties and can be processed by networks with different capacities. Intuitively, smooth areas are easier to super-solve than complex textures. To utilize this property, we can adopt appropriate SR networks to process different sub-images after the decomposition. On this basis, we propose a new solution pipeline -- ClassSR that combines classification and SR in a unified framework. In particular, it first uses a Class-Module to classify the sub-images into different classes according to restoration difficulties, then applies an SR-Module to perform SR for different classes. The Class-Module is a conventional classification network, while the SR-Module is a network container that consists of the to-be-accelerated SR network and its simplified versions. We further introduce a new classification method with two losses -- Class-Loss and Average-Loss to produce the classification results. After joint training, a majority of sub-images will pass through smaller networks, thus the computational cost can be significantly reduced. Experiments show that our ClassSR can help most existing methods (e.g., FSRCNN, CARN, SRResNet, RCAN) save up to 50% FLOPs on DIV8K datasets. This general framework can also be applied in other low-level vision tasks.
This article unveils a new relation between the Nishimori temperature parametrizing a distribution P and the Bethe free energy on random Erdos-Renyi graphs with edge weights distributed according to P. Estimating the Nishimori temperature being a task of major importance in Bayesian inference problems, as a practical corollary of this new relation, a numerical method is proposed to accurately estimate the Nishimori temperature from the eigenvalues of the Bethe Hessian matrix of the weighted graph. The algorithm, in turn, is used to propose a new spectral method for node classification in weighted (possibly sparse) graphs. The superiority of the method over competing state-of-the-art approaches is demonstrated both through theoretical arguments and real-world data experiments.
Globally, Skin carcinoma is among the most lethal diseases. Millions of people are diagnosed with this cancer every year. Sill, early detection can decrease the medication cost and mortality rate substantially. The recent improvement in automated cancer classification using deep learning methods has reached a human-level performance requiring a large amount of annotated data assembled in one location, yet, finding such conditions usually is not feasible. Recently, federated learning (FL) has been proposed to train decentralized models in a privacy-preserved fashion depending on labeled data at the client-side, which is usually not available and costly. To address this, we propose \verb!FedPerl!, a semi-supervised federated learning method. Our method is inspired by peer learning from educational psychology and ensemble averaging from committee machines. FedPerl builds communities based on clients' similarities. Then it encourages communities members to learn from each other to generate more accurate pseudo labels for the unlabeled data. We also proposed the peer anonymization (PA) technique to anonymize clients. As a core component of our method, PA is orthogonal to other methods without additional complexity and reduces the communication cost while enhancing performance. Finally, we propose a dynamic peer-learning policy that controls the learning stream to avoid any degradation in the performance, especially for individual clients. Our experimental setup consists of 71,000 skin lesion images collected from 5 publicly available datasets. We test our method in four different scenarios in SSFL. With few annotated data, FedPerl is on par with a state-of-the-art method in skin lesion classification in the standard setup while outperforming SSFLs and the baselines by 1.8% and 15.8%, respectively. Also, it generalizes better to unseen clients while being less sensitive to noisy ones.
The training of deep learning models generally requires a large amount of annotated data for effective convergence and generalisation. However, obtaining high-quality annotations is a laboursome and expensive process due to the need of expert radiologists for the labelling task. The study of semi-supervised learning in medical image analysis is then of crucial importance given that it is much less expensive to obtain unlabelled images than to acquire images labelled by expert radiologists. Essentially, semi-supervised methods leverage large sets of unlabelled data to enable better training convergence and generalisation than using only the small set of labelled images. In this paper, we propose Self-supervised Mean Teacher for Semi-supervised (S$^2$MTS$^2$) learning that combines self-supervised mean-teacher pre-training with semi-supervised fine-tuning. The main innovation of S$^2$MTS$^2$ is the self-supervised mean-teacher pre-training based on the joint contrastive learning, which uses an infinite number of pairs of positive query and key features to improve the mean-teacher representation. The model is then fine-tuned using the exponential moving average teacher framework trained with semi-supervised learning. We validate S$^2$MTS$^2$ on the multi-label classification problems from Chest X-ray14 and CheXpert, and the multi-class classification from ISIC2018, where we show that it outperforms the previous SOTA semi-supervised learning methods by a large margin.
Hierarchical text classification consists in classifying text documents into a hierarchy of classes and sub-classes. Although artificial neural networks have proved useful to perform this task, unfortunately they can leak training data information to adversaries due to training data memorization. Using differential privacy during model training can mitigate leakage attacks against trained models, enabling the models to be shared safely at the cost of reduced model accuracy. This work investigates the privacy-utility trade-off in hierarchical text classification with differential privacy guarantees, and identifies neural network architectures that offer superior trade-offs. To this end, we use a white-box membership inference attack to empirically assess the information leakage of three widely used neural network architectures. We show that large differential privacy parameters already suffice to completely mitigate membership inference attacks, thus resulting only in a moderate decrease in model utility. More specifically, for large datasets with long texts we observed Transformer-based models to achieve an overall favorable privacy-utility trade-off, while for smaller datasets with shorter texts convolutional neural networks are preferable.
This paper discusses the problem of weakly supervised classification, in which instances are given weak labels that are produced by some label-corruption process. The goal is to derive conditions under which loss functions for weak-label learning are proper and lower-bounded -- two essential requirements for the losses used in class-probability estimation. To this end, we derive a representation theorem for proper losses in supervised learning, which dualizes the Savage representation. We use this theorem to characterize proper weak-label losses and find a condition for them to be lower-bounded. From these theoretical findings, we derive a novel regularization scheme called generalized logit squeezing, which makes any proper weak-label loss bounded from below, without losing properness. Furthermore, we experimentally demonstrate the effectiveness of our proposed approach, as compared to improper or unbounded losses. The results highlight the importance of properness and lower-boundedness.
Learning feature representation from discriminative local regions plays a key role in fine-grained visual classification. Employing attention mechanisms to extract part features has become a trend. However, there are two major limitations in these methods: First, they often focus on the most salient part while neglecting other inconspicuous but distinguishable parts. Second, they treat different part features in isolation while neglecting their relationships. To handle these limitations, we propose to locate multiple different distinguishable parts and explore their relationships in an explicit way. In this pursuit, we introduce two lightweight modules that can be easily plugged into existing convolutional neural networks. On one hand, we introduce a feature boosting and suppression module that boosts the most salient part of feature maps to obtain a part-specific representation and suppresses it to force the following network to mine other potential parts. On the other hand, we introduce a feature diversification module that learns semantically complementary information from the correlated part-specific representations. Our method does not need bounding boxes/part annotations and can be trained end-to-end. Extensive experimental results show that our method achieves state-of-the-art performances on several benchmark fine-grained datasets. Source code is available at https://github.com/chaomaer/FBSD.
Self-supervised learning (SSL) is rapidly closing the gap with supervised methods on large computer vision benchmarks. A successful approach to SSL is to learn embeddings which are invariant to distortions of the input sample. However, a recurring issue with this approach is the existence of trivial constant solutions. Most current methods avoid such solutions by careful implementation details. We propose an objective function that naturally avoids collapse by measuring the cross-correlation matrix between the outputs of two identical networks fed with distorted versions of a sample, and making it as close to the identity matrix as possible. This causes the embedding vectors of distorted versions of a sample to be similar, while minimizing the redundancy between the components of these vectors. The method is called Barlow Twins, owing to neuroscientist H. Barlow's redundancy-reduction principle applied to a pair of identical networks. Barlow Twins does not require large batches nor asymmetry between the network twins such as a predictor network, gradient stopping, or a moving average on the weight updates. Intriguingly it benefits from very high-dimensional output vectors. Barlow Twins outperforms previous methods on ImageNet for semi-supervised classification in the low-data regime, and is on par with current state of the art for ImageNet classification with a linear classifier head, and for transfer tasks of classification and object detection.
We propose a novel methodology for general multi-class classification in arbitrary feature spaces, which results in a potentially well-calibrated classifier. Calibrated classifiers are important in many applications because, in addition to the prediction of mere class labels, they also yield a confidence level for each of their predictions. In essence, the training of our classifier proceeds in two steps. In a first step, the training data is represented in a latent space whose geometry is induced by a regular $(n-1)$-dimensional simplex, $n$ being the number of classes. We design this representation in such a way that it well reflects the feature space distances of the datapoints to their own- and foreign-class neighbors. In a second step, the latent space representation of the training data is extended to the whole feature space by fitting a regression model to the transformed data. With this latent-space representation, our calibrated classifier is readily defined. We rigorously establish its core theoretical properties and benchmark its prediction and calibration properties by means of various synthetic and real-world data sets from different application domains.
The convolutional neural network (CNN) is vulnerable to degraded images with even very small variations (e.g. corrupted and adversarial samples). One of the possible reasons is that CNN pays more attention to the most discriminative regions, but ignores the auxiliary features when learning, leading to the lack of feature diversity for final judgment. In our method, we propose to dynamically suppress significant activation values of CNN by group-wise inhibition, but not fixedly or randomly handle them when training. The feature maps with different activation distribution are then processed separately to take the feature independence into account. CNN is finally guided to learn richer discriminative features hierarchically for robust classification according to the proposed regularization. Our method is comprehensively evaluated under multiple settings, including classification against corruptions, adversarial attacks and low data regime. Extensive experimental results show that the proposed method can achieve significant improvements in terms of both robustness and generalization performances, when compared with the state-of-the-art methods. Code is available at https://github.com/LinusWu/TENET_Training.
In this paper, we propose a progressive margin loss (PML) approach for unconstrained facial age classification. Conventional methods make strong assumption on that each class owns adequate instances to outline its data distribution, likely leading to bias prediction where the training samples are sparse across age classes. Instead, our PML aims to adaptively refine the age label pattern by enforcing a couple of margins, which fully takes in the in-between discrepancy of the intra-class variance, inter-class variance and class center. Our PML typically incorporates with the ordinal margin and the variational margin, simultaneously plugging in the globally-tuned deep neural network paradigm. More specifically, the ordinal margin learns to exploit the correlated relationship of the real-world age labels. Accordingly, the variational margin is leveraged to minimize the influence of head classes that misleads the prediction of tailed samples. Moreover, our optimization carefully seeks a series of indicator curricula to achieve robust and efficient model training. Extensive experimental results on three face aging datasets demonstrate that our PML achieves compelling performance compared to state of the arts. Code will be made publicly.
The goal in label-imbalanced and group-sensitive classification is to optimize relevant metrics such as balanced error and equal opportunity. Classical methods, such as weighted cross-entropy, fail when training deep nets to the terminal phase of training (TPT), that is training beyond zero training error. This observation has motivated recent flurry of activity in developing heuristic alternatives following the intuitive mechanism of promoting larger margin for minorities. In contrast to previous heuristics, we follow a principled analysis explaining how different loss adjustments affect margins. First, we prove that for all linear classifiers trained in TPT, it is necessary to introduce multiplicative, rather than additive, logit adjustments so that the interclass margins change appropriately. To show this, we discover a connection of the multiplicative CE modification to the cost-sensitive support-vector machines. Perhaps counterintuitively, we also find that, at the start of training, the same multiplicative weights can actually harm the minority classes. Thus, while additive adjustments are ineffective in the TPT, we show that they can speed up convergence by countering the initial negative effect of the multiplicative weights. Motivated by these findings, we formulate the vector-scaling (VS) loss, that captures existing techniques as special cases. Moreover, we introduce a natural extension of the VS-loss to group-sensitive classification, thus treating the two common types of imbalances (label/group) in a unifying way. Importantly, our experiments on state-of-the-art datasets are fully consistent with our theoretical insights and confirm the superior performance of our algorithms. Finally, for imbalanced Gaussian-mixtures data, we perform a generalization analysis, revealing tradeoffs between balanced / standard error and equal opportunity.
The rapid advances in Deep Learning (DL) techniques have enabled rapid detection, localisation, and recognition of objects from images or videos. DL techniques are now being used in many applications related to agriculture and farming. Automatic detection and classification of weeds can play an important role in weed management and so contribute to higher yields. Weed detection in crops from imagery is inherently a challenging problem because both weeds and crops have similar colours ('green-on-green'), and their shapes and texture can be very similar at the growth phase. Also, a crop in one setting can be considered a weed in another. In addition to their detection, the recognition of specific weed species is essential so that targeted controlling mechanisms (e.g. appropriate herbicides and correct doses) can be applied. In this paper, we review existing deep learning-based weed detection and classification techniques. We cover the detailed literature on four main procedures, i.e., data acquisition, dataset preparation, DL techniques employed for detection, location and classification of weeds in crops, and evaluation metrics approaches. We found that most studies applied supervised learning techniques, they achieved high classification accuracy by fine-tuning pre-trained models on any plant dataset, and past experiments have already achieved high accuracy when a large amount of labelled data is available.
Environmental Sound Classification (ESC) is a challenging field of research in non-speech audio processing. Most of current research in ESC focuses on designing deep models with special architectures tailored for specific audio datasets, which usually cannot exploit the intrinsic patterns in the data. However recent studies have surprisingly shown that transfer learning from models trained on ImageNet is a very effective technique in ESC. Herein, we propose SoundCLR, a supervised contrastive learning method for effective environment sound classification with state-of-the-art performance, which works by learning representations that disentangle the samples of each class from those of other classes. Our deep network models are trained by combining a contrastive loss that contributes to a better probability output by the classification layer with a cross-entropy loss on the output of the classifier layer to map the samples to their respective 1-hot encoded labels. Due to the comparatively small sizes of the available environmental sound datasets, we propose and exploit a transfer learning and strong data augmentation pipeline and apply the augmentations on both the sound signals and their log-mel spectrograms before inputting them to the model. Our experiments show that our masking based augmentation technique on the log-mel spectrograms can significantly improve the recognition performance. Our extensive benchmark experiments show that our hybrid deep network models trained with combined contrastive and cross-entropy loss achieved the state-of-the-art performance on three benchmark datasets ESC-10, ESC-50, and US8K with validation accuracies of 99.75\%, 93.4\%, and 86.49\% respectively. The ensemble version of our models also outperforms other top ensemble methods. The code is available at https://github.com/alireza-nasiri/SoundCLR.