Section: New Results

Extracting and Representing Information

Text Mining in the Clinical Domain

Participants : Clément Dalloux, Vincent Claveau.

Clinical records cannot be shared, which is a real hurdle to develop and compare information extraction techniques. In the framework of the BigClin Project we have developed annotated corpora, that share the same linguistic properties than records, but can be freely distributed for research purposes. Several corpora and several types of annotation were proposed for French, Portuguese and English. They are made freely available for research purposes and are described in [27], [25]. These corpora will foster reproducible research on clinical text mining.

Thanks to these datasets, we have organized the DeFT text-mining competition in 2019. Several NLP techniques and tools have been developed within the project in order to identify relevant medical or linguistic information [30], [26]. They are all chiefly based on machine learning approaches, and for most of them, more specifically, on deep learning. For instance, we have developed a new Part-of-Speech tagger and lemmatizer for French, especially suited to handle medical texts; it is freely available as a web-service at  https://allgo.inria.fr . The identification of negation and uncertainty is important to precisely understand the clinical texts. Thus, we have continued our work on neural techniques to find the negation/uncertainty cues and their scope (part of sentence concerned by the negation or uncertainty). It achieves state-of-the-art results on English, and is pioneer work for French and Portuguese for which it sets a new standard [4], [21]; it is available at  https://allgo.inria.fr . Other achievements in text-mining include: numerical value extraction (finding concepts that are measured, such as lab results, numerical expressions, their units) in French, English and Portuguese, the identification of gender, age, outcome and admission reasons in French clinical texts, ...

Embedding in hyperbolic spaces

Participants : François Torregrossa, Vincent Claveau, Guillaume Gravier.

During this year, we have studied non-Euclidean spaces into which one can embed data (for instance, words). We have developed the HierarX tool which projects multiple datasources into hyperbolic manifolds: Lorentz or Poincaré. From similarities between word pairs or continuous word representations in high dimensional spaces, HierarX is able to embed knowledge in hyperbolic geometries with small dimensionality. Those shape information into continuous hierarchies. The source code is available on the Inria's GitLab.

Aggregation and embedding for group membership verification

Participants : Marzieh Gheisari Khorasgani, Teddy Furon, Laurent Amsaleg.

This paper proposes a group membership verification protocol preventing the curious but honest server from reconstructing the enrolled signatures and inferring the identity of querying clients [24]. The protocol quantizes the signatures into discrete embeddings, making reconstruction difficult. It also aggregates multiple embeddings into representative values, impeding identification. Theoretical and experimental results show the trade-off between the security and error rates.

Group Membership Verification with Privacy: Sparse or Dense?

Participants : Marzieh Gheisari Khorasgani, Teddy Furon, Laurent Amsaleg.

Group membership verification checks if a biometric trait corresponds to one member of a group without revealing the identity of that member. Recent contributions provide privacy for group membership protocols through the joint use of two mechanisms: quantizing templates into discrete embeddings, and aggregating several templates into one group representation. However, this scheme has one drawback: the data structure representing the group has a limited size and cannot recognize noisy query when many templates are aggregated. Moreover, the sparsity of the embeddings seemingly plays a crucial role on the performance verification. This contribution proposes a mathematical model for group membership verification allowing to reveal the impact of sparsity on both security, compactness, and verification performances [23]. This models bridges the gap towards a Bloom filter robust to noisy queries. It shows that a dense solution is more competitive unless the queries are almost noiseless.

Privacy Preserving Group Membership Verification and Identification

Participants : Marzieh Gheisari Khorasgani, Teddy Furon, Laurent Amsaleg.

When convoking privacy, group membership verification checks if a biometric trait corresponds to one member of a group without revealing the identity of that member. Similarly, group membership identification states which group the individual belongs to, without knowing his/her identity. A recent contribution provides privacy and security for group membership protocols through the joint use of two mechanisms: quantizing biometric templates into discrete embeddings, and aggregating several templates into one group representation. This paper significantly improves that contribution because it jointly learns how to embed and aggregate instead of imposing fixed and hard coded rules [10]. This is demonstrated by exposing the mathematical underpinnings of the learning stage before showing the improvements through an extensive series of experiments targeting face recognition. Overall, experiments show that learning yields an excellent trade-off between security/privacy and the verification/identification performances.

Intrinsic Dimensionality Estimation within Tight Localities

Participants : Laurent Amsaleg, Oussama Chelly [Microsoft Germany] , Michael Houle [National Institute of Informatics, Japan] , Ken-Ichi Kawarabayashi [National Institute of Informatics, Japan] , Miloš Radovanović [Univ. Novi Sad, Serbia] , Weeris Treeratanajaru [Chulalongkorn University, Thailand] .

Accurate estimation of Intrinsic Dimensionality (ID) is of crucial importance in many data mining and machine learning tasks, including dimensionality reduction, outlier detection, similarity search and subspace clustering. However, since their convergence generally requires sample sizes (that is, neighborhood sizes) on the order of hundreds of points, existing ID estimation methods may have only limited usefulness for applications in which the data consists of many natural groups of small size. In this paper, we propose a local ID estimation strategy stable even for ‘tight’ localities consisting of as few as 20 sample points [31]. The estimator applies MLE techniques over all available pairwise distances among the members of the sample, based on a recent extreme-value-theoretic model of intrinsic dimensionality, the Local Intrinsic Dimension (LID). Our experimental results show that our proposed estimation technique can achieve notably smaller variance, while maintaining comparable levels of bias, at much smaller sample sizes than state-of-the-art estimators.

Selective Biogeography-Based Optimizer Considering Resource Allocation for Large-Scale Global Optimization

Participants : Meiji Cui [Tongji University, China] , Li Li [Tongji University, China] , Miaojing Shi.

Biogeography-based optimization (BBO), a recent proposed meta-heuristic algorithm, has been successfully applied to many optimization problems due to its simplicity and efficiency. However, BBO is sensitive to the curse of dimensionality; its performance degrades rapidly as the dimensionality of the search space increases. In [3], a selective migration operator is proposed to scale up the performance of BBO and we name it selective BBO (SBBO). The differential migration operator is selected heuristically to explore the global area as far as possible whist the normal distributed migration operator is chosen to exploit the local area. By the means of heuristic selection, an appropriate migration operator can be used to search the global optimum efficiently. Moreover, the strategy of cooperative co-evolution (CC) is adopted to solve large-scale global optimization problems (LSOPs). To deal with subgroup imbalance contribution to the whole solution in the context of CC, a more efficient computing resource allocation is proposed. Extensive experiments are conducted on the CEC 2010 benchmark suite for large-scale global optimization, and the results show the effectiveness and efficiency of SBBO compared with BBO variants and other representative algorithms for LSOPs. Also, the results confirm that the proposed computing resource allocation is vital to the large-scale optimization within the limited computation budget.

Friend recommendation for cross marketing in online brand community based on intelligent attention allocation link prediction algorithm

Participants : Shugang Li [Shanghai University, China] , Xuewei Song [Shanghai University, China] , Hanyu Lu [Shanghai University, China] , Linyi Zeng [Shanghai University, Industrial and Commercial Bank of China, China] , Miaojing Shi, Fang Liu [Shanghai University, China] .

Circle structure of online brand communities allows companies to conduct cross-marketing activities by the influence of friends in different circles and build strong and lasting relationships with customers. However, existing works on the friend recommendation in social network do not consider establishing friendships between users in different circles, which has the problems of network sparsity, neither do they study the adaptive generation of appropriate link prediction algorithms for different circle features. In order to fill the gaps in previous works, the intelligent attention allocation link prediction algorithm is proposed to adaptively build attention allocation index (AAI) according to the sparseness of the network and predict the possible friendships between users in different circles. The AAI reflects the amount of attention allocated to the user pair by their common friend in the triadic closure structure, which is decided by the friend count of the common friend. Specifically, for the purpose of overcoming the problem of network sparsity, the AAIs of both the direct common friends and indirect ones are developed. Next, the decision tree (DT) method is constructed to adaptively select the suitable AAIs for the circle structure based on the density of common friends and the dispersion level of common friends’ attention. In addition, for the sake of further improving the accuracy of the selected AAI, its complementary AAIs are identified with support vector machine model according to their similarity in value, direction, and ranking. Finally, the mutually complementary indices are combined into a composite one to comprehensively portray the attention distribution of common friends of users in different circles and predict their possible friendships for cross-marketing activities. Experimental results on Twitter and Google+ show that the model has highly reliable prediction performance [5].

Revisiting the medial axis for planar shape decomposition

Participants : Nikos Papanelopoulos [NTUA, Greece] , Yannis Avrithis, Stefanos Kollias [U. of Lincoln, UK] .

We present a simple computational model for planar shape decomposition that naturally captures most of the rules and salience measures suggested by psychophysical studies, including the minima and short-cut rules, convexity, and symmetry. It is based on a medial axis representation in ways that have not been explored before and sheds more light into the connection between existing rules like minima and convexity. In particular, vertices of the exterior medial axis directly provide the position and extent of negative minima of curvature, while a traversal of the interior medial axis directly provides a small set of candidate endpoints for part-cuts. The final selection follows a prioritized processing of candidate part-cuts according to a local convexity rule that can incorporate arbitrary salience measures. Neither global optimization nor differentiation is involved. We provide qualitative and quantitative evaluation and comparisons on ground-truth data from psycho-physical experiments. With our single computational model, we outperform even an ensemble method on several other competing models [6].

Graph-based Particular Object Discovery

Participants : Oriane Siméoni, Ahmet Iscen [Univ. Prague] , Giorgos Tolias [Univ. Prague] , Yannis Avrithis, Ondra Chum [Univ. Prague] .

Severe background clutter is challenging in many computer vision tasks, including large-scale image retrieval. Global descriptors, that are popular due to their memory and search efficiency, are especially prone to corruption by such a clutter. Eliminating the impact of the clutter on the image descriptor increases the chance of retrieving relevant images and prevents topic drift due to actually retrieving the clutter in the case of query expansion. In this work, we propose a novel salient region detection method. It captures, in an unsupervised manner, patterns that are both discriminative and common in the dataset. Saliency is based on a centrality measure of a nearest neighbor graph constructed from regional CNN representations of dataset images. The proposed method exploits recent CNN architectures trained for object retrieval to construct the image representation from the salient regions. We improve particular object retrieval on challenging datasets containing small objects [7].

Label Propagation for Deep Semi-supervised Learning

Participants : Ahmet Iscen [Univ. Prague] , Giorgos Tolias [Univ. Prague] , Yannis Avrithis, Ondra Chum [Univ. Prague] .

Semi-supervised learning is becoming increasingly important because it can combine data carefully labeled by humans with abundant unlabeled data to train deep neural networks. Classic methods on semi-supervised learning that have focused on transductive learning have not been fully exploited in the inductive framework followed by modern deep learning. The same holds for the manifold assumption—that similar examples should get the same prediction. In this work, we employ a transductive label propagation method that is based on the manifold assumption to make predictions on the entire dataset and use these predictions to generate pseudo-labels for the unlabeled data and train a deep neural network. At the core of the transductive method lies a nearest neighbor graph of the dataset that we create based on the embeddings of the same network. Therefore our learning process iterates between these two steps. We improve performance on several datasets especially in the few labels regime and show that our work is complementary to current state of the art [12], [38].

Dense Classification and Implanting for Few-Shot Learning

Participants : Yann Lifchitz, Yannis Avrithis, Sylvaine Picard [SAFRAN Group] , Andrei Bursuc [Valéo] .

Few-shot learning for deep neural networks is a highly challenging and key problem in many computer vision tasks. In this context, we are targeting knowledge transfer from a set with abundant data to other sets with few available examples. We propose in [14], [40] two simple and effective solutions: (i) dense classification over feature maps, which for the first time studies local activations in the domain of few- shot learning, and (ii) implanting, that is, attaching new neurons to a previously trained network to learn new, task-specific features. Implanting enables training of multiple layers in the few-shot regime, departing from most related methods derived from metric learning that train only the final layer. Both contributions show consistent gains when used individually or jointly and we report state of the art performance on few-shot classification on miniImageNet.

Point in, Box out: Beyond Counting Persons in Crowds

Participants : Yuting Liu [Sichuan University, China] , Miaojing Shi, Qijun Zhao [Sichuan University, China] , Xiaofang Wang [RAINBOW Team, IRISA] .

Modern crowd counting methods usually employ deep neural networks (DNN) to estimate crowd counts via density regression. Despite their significant improvements, the regression-based methods are incapable of providing the detection of individuals in crowds. The detection-based methods, on the other hand, have not been largely explored in recent trends of crowd counting due to the needs for ex- pensive bounding box annotations. In this work, we instead propose a new deep detection network with only point supervision required [15]. It can simultaneously detect the size and location of human heads and count them in crowds. We first mine useful person size information from point-level annotations and initialize the pseudo ground truth bounding boxes. An online updating scheme is introduced to refine the pseudo ground truth during training; while a locally- constrained regression loss is designed to provide additional constraints on the size of the predicted boxes in a local neighborhood. In the end, we propose a curriculum learning strategy to train the network from images of relatively accurate and easy pseudo ground truth first. Extensive experiments are conducted in both detection and counting tasks on several standard benchmarks, e.g. ShanghaiTech, UCF CC 50, WiderFace, and TRANCOS datasets, and the results show the superiority of our method over the state-of-the-art.

Revisiting Perspective Information for Efficient Crowd Counting

Participants : Miaojing Shi, Zhaohui Yang [Peking University, China] , Chao Xu [Peking University, China] , Qijun Chen [Tongji University, China] .

Crowd counting is the task of estimating people numbers in crowd images. Modern crowd counting methods employ deep neural networks to estimate crowd counts via crowd density regressions. A major challenge of this task lies in the perspective distortion, which results in drastic person scale change in an image. Density regression on the small person area is in general very hard. In this work, we propose a perspective-aware convolutional neural network (PACNN) for efficient crowd counting, which integrates the perspective information into density regression to provide additional knowledge of the person scale change in an image [18]. Ground truth perspective maps are firstly generated for training; PACNN is then specifically designed to predict multi-scale perspective maps, and encode them as perspective-aware weighting layers in the network to adaptively combine the outputs of multi-scale density maps. The weights are learned at every pixel of the maps such that the final density combination is robust to the perspective distortion. We conduct extensive experiments on the ShanghaiTech, WorldExpo’10, UCF CC 50, and UCSD datasets, and demonstrate the effectiveness and efficiency of PACNN over the state-of-the-art.

Local Features and Visual Words Emerge in Activations

Participants : Oriane Siméoni, Yannis Avrithis, Ondra Chum [Univ. Prague] .

We propose a novel method of deep spatial matching (DSM) for image retrieval [19], [41]. Initial ranking is based on image descriptors extracted from convolutional neural network activations by global pooling, as in recent state-of- the-art work. However, the same sparse 3D activation tensor is also approximated by a collection of local features. These local features are then robustly matched to approximate the optimal alignment of the tensors. This hap- pens without any network modification, additional layers or training. No local feature detection happens on the original image. No local feature descriptors and no visual vocabulary are needed throughout the whole process. We experimentally show that the proposed method achieves the state-of-the-art performance on standard benchmarks across different network architectures and different global pooling methods. The highest gain in performance is achieved when diffusion on the nearest-neighbor graph of global descriptors is initiated from spatially verified images.

Combining convolutional side-outputs for road image segmentation

Participants : Raquel Almeida, Simon Malinowski, Ewa Kijak, Silvio Guimaraes [PUC Minas] .

Image segmentation consists in creating partitions within an image into meaningful areas and objects. It can be used in scene understanding and recognition, in fields like biology, medicine, robotics, satellite imaging, amongst others. In this work [17], we take advantage of the learned model in a deep architecture, by extracting side-outputs at different layers of the network for the task of image segmentation. We study the impact of the amount of side-outputs and evaluate strategies to combine them. A post-processing filtering based on mathematical morphology idempotent functions is also used in order to remove some undesirable noises. Experiments were performed on the publicly available KITTI Road Dataset for image segmentation. Our comparison shows that the use of multiples side outputs can increase the overall performance of the network, making it easier to train and more stable when compared with a single output in the end of the network. Also, for a small number of training epochs (500), we achieved a competitive performance when compared to the best algorithm in KITTI Evaluation Server.

BRIEF-based mid-level representations for time series classification

Participants : Raquel Almeida, Simon Malinowski, Silvio Guimaraes [PUC Minas] .

Time series classification has been widely explored over the last years. Amongst the best approaches for that task, many are based on the Bag-of-Words framework, in which time series are transformed into a histogram of word occurrences. These words represent quantized features that are extracted beforehand. In this work [20], we aim to evaluate the use of accurate mid-level representation called BossaNova in order to enhance the Bag-of-Words representation and to propose a new binary time series descriptor, called BRIEF-based descriptor. More precisely, this kind of representation enables to reduce the loss induced by feature quantization. Experiments show that this representation in conjunction to BRIEF-based descriptor is statistically equivalent to traditional Bag-of-Words, in terms time series classification accuracy, being about 4 times faster. Furthermore, it is very competitive when compared to the state-of-the-art.

Toward a Framework for Seasonal Time Series Forecasting Using Clustering

Participants : Simon Malinowski, Thomas Guyet [LACODAM Team] , Colin Leverger [LACODAM Team] , Alexandre Termier [LACODAM Team] .

Seasonal behaviours are widely encountered in various applications. For instance, requests on web servers are highly influenced by our daily activities. Seasonal forecasting consists in forecasting the whole next season for a given seasonal time series. It may help a service provider to provision correctly the potentially required resources, avoiding critical situations of over- or under provision. In this article, we propose a generic framework to make seasonal time series forecasting. The framework combines machine learning techniques (1) to identify the typical seasons and (2) to forecast the likelihood of having a season type in one season ahead. We study in[13] this framework by comparing the mean squared errors of forecasts for various settings and various datasets. The best setting is then compared to state-of-the-art time series forecasting methods. We show that it is competitive with them.

Smooth Adversarial Examples

Participants : Hanwei Zhang, Yannis Avrithis, Teddy Furon, Laurent Amsaleg.

This paper investigates the visual quality of the adversarial examples. Recent papers propose to smooth the perturbations to get rid of high frequency artefacts. In this work, smoothing has a different meaning as it perceptually shapes the perturbation according to the visual content of the image to be attacked [44]. The perturbation becomes locally smooth on the flat areas of the input image, but it may be noisy on its textured areas and sharp across its edges. This operation relies on Laplacian smoothing, well-known in graph signal processing, which we integrate in the attack pipeline. We benchmark several attacks with and without smoothing under a white-box scenario and evaluate their transferability. Despite the additional constraint of smoothness, our attack has the same probability of success at lower distortion.

Walking on the Edge: Fast, Low-Distortion Adversarial Examples

Participants : Hanwei Zhang, Yannis Avrithis, Teddy Furon, Laurent Amsaleg.

Adversarial examples of deep neural networks are receiving ever increasing attention because they help in understanding and reducing the sensitivity to their input. This is natural given the increasing applications of deep neural networks in our everyday lives. When white-box attacks are almost always successful, it is typically only the distortion of the perturbations that matters in their evaluation. In this work [45], we argue that speed is important as well, especially when considering that fast attacks are required by adversarial training. Given more time, iterative methods can always find better solutions. We investigate this speed-distortion trade-off in some depth and introduce a new attack called boundary projection (BP) that improves upon existing methods by a large margin. Our key idea is that the classification boundary is a manifold in the image space: we therefore quickly reach the boundary and then optimize distortion on this manifold.

Accessing watermarking information: Error exponents in the noisy case

Participant : Teddy Furon.

The study of the error exponents of zero-bit watermarking is addressed in the article by Comesana, Merhav, and Barni, under the assumption that the detector relies solely on second order joint empirical statistics of the received signal and the watermark. This restriction leads to the well-known dual hypercone detector, whose score function is the absolute value of the normalized correlation. They derive the false negative error exponent and the optimum embedding rule. However, they only focus on high SNR regime, i.e. the noiseless scenario. This work extends this theoretical study to the noisy scenario. It introduces a new definition of watermarking robustness based on the false negative error exponent, derives this quantity for the dual hypercone detector, and shows that its performances is almost equal to Costa's lower bound [22].

Detecting fake news and image forgeries

Participants : Cédric Maigrot, Vincent Claveau, Ewa Kijak.

Social networks make it possible to share information rapidly and massively. Yet, one of their major drawback comes from the absence of verification of the piece of information, especially with viral messages. Based on the work already presented in the previous years, C. Maigrot defended his thesis on the detection of image forgeries, classification of reinformation websites, and on the late fusion of models based on the text, image and source analysis [1]. This work was also given a large visibility thanks to numerous interviews in Press and TV (see the dedicated section about popularization).

Learning Interpretable Shapelets for Time Series Classification through Adversarial Regularization

Times series classification can be successfully tackled by jointly learning a shapelet-based representation of the series in the dataset and classifying the series according to this representation. However, although the learned shapelets are discriminative, they are not always similar to pieces of a real series in the dataset. This makes it difficult to interpret the decision, i.e. difficult to analyze if there are particular behaviors in a series that triggered the decision. In this work [29], we make use of a simple convolutional network to tackle the time series classification task and we introduce an adversarial regularization to constrain the model to learn more interpretable shapelets. Our classification results on all the usual time series benchmarks are comparable with the results obtained by similar state-of-the-art algorithms but our adversarially regularized method learns shapelets that are, by design, interpretable.

Using Knowledge Base Semantics in Context-Aware Entity Linking

Participants : Cheikh Brahim El Vaigh, Guillaume Gravier, Pascale Sébillot.

Done as part of the IPL iCODA, in collaboration with CEDAR Inria team.

Entity linking is a core task in textual document processing, which consists in identifying the entities of a knowledge base (KB) that are mentioned in a text. Approaches in the literature consider either independent linking of individual mentions or collective linking of all mentions. Regardless of this distinction, most approaches rely on the Wikipedia encyclopedic KB in order to improve the linking quality, by exploiting its entity descriptions (web pages) or its entity interconnections (hyperlink graph of web pages). We devised a novel collective linking technique which departs from most approaches in the literature by relying on a structured RDF KB [9]. This allows exploiting the semantics of the interrelationships that candidate entities may have at disambiguation time rather than relying on raw structural approximation based on Wikipedia’s hyperlink, graph. The few approaches that also use an RDF KB simply rely on the existence of a relation between the candidate entities to which mentions may be linked. Instead, we weight such relations based on the RDF KB structure and propose an efficient decoding strategy for collective linking. Experiments on standard benchmarks show significant improvement over the state of the art.

Neural-based lexico-syntactic relation extraction in news archives

Participants : Guillaume Gravier, Cyrielle Mallart, Pascale Sébillot.

Done as part of the IPL iCODA, in collaboration with Ouest France

Relation extraction is the task of finding and classifying the relationship between two entities in a text. We pursued work on the detection of relations between entities, seen as a binary classification problem. In the context of large-scale news archives, we argue that detection is paramount before even considering classification, where most approaches consider the two tasks jointly with a null garbage class. This does hardly allow for the detection of relations for unseen categories, which are all considered as garbage. We designed a bi-LSTM sequence neural model acting on features extracted from the surface realization, the part-of-speech tags and the dependency parse tree and compared with a state-of-the-art relation detection LSTM-based approach. Experimental evaluations rely on a dataset derived from 200k Wikipedia articles in French containing 4M linked mentions of entities: 330k pairs of entities co-occur in the same sentence, of which 1 % are actual relations according to Wikidata. Results show the benefit of our binary detection approach over previous methods and over joint detection and classification.

Graph Convolutional Networks for Learning with Few Clean and Many Noisy Labels

Participants : Ahmet Iscen [Google Research] , Giorgos Tolias [Univ. Prague] , Yannis Avrithis, Ondra Chum [Univ. Prague] , Cordelia Schmid [Google Research] .

In this work we consider the problem of learning a classifier from noisy labels when a few clean labeled examples are given [39]. The structure of clean and noisy data is modeled by a graph per class and Graph Convolutional Networks (GCN) are used to predict class relevance of noisy examples. For each class, the GCN is treated as a binary classifier learning to discriminate clean from noisy examples using a weighted binary cross-entropy loss function, and then the GCN-inferred "clean" probability is exploited as a relevance measure. Each noisy example is weighted by its relevance when learning a classifier for the end task. We evaluate our method on an extended version of a few-shot learning problem, where the few clean examples of novel classes are supplemented with additional noisy data. Experimental results show that our GCN-based cleaning process significantly improves the classification accuracy over not cleaning the noisy data and standard few-shot classification where only few clean examples are used. The proposed GCN-based method outperforms the transductive approach (Douze et al., 2018) that is using the same additional data without labels.

Rethinking deep active learning: Using unlabeled data at model training

Participants : Oriane Siméoni, Mateusz Budnik, Yannis Avrithis, Guillaume Gravier.

Active learning typically focuses on training a model on few labeled examples alone, while unlabeled ones are only used for acquisition. In this work we depart from this setting by using both labeled and unlabeled data during model training across active learning cycles [42]. We do so by using unsupervised feature learning at the beginning of the active learning pipeline and semi-supervised learning at every active learning cycle, on all available data. The former has not been investigated before in active learning, while the study of latter in the context of deep learning is scarce and recent findings are not conclusive with respect to its benefit. Our idea is orthogonal to acquisition strategies by using more data, much like ensemble methods use more models. By systematically evaluating on a number of popular acquisition strategies and datasets, we find that the use of unlabeled data during model training brings a spectacular accuracy improvement in image classification, compared to the differences between acquisition strategies. We thus explore smaller label budgets, even one label per class.

Training Object Detectors from Few Weakly-Labeled and Many Unlabeled Images

Participants : Zhaohui Yang [Peking University] , Miaojing Shi, Yannis Avrithis, Chao Xu [Peking University] , Vittorio Ferrari [Google Research] .

Weakly-supervised object detection attempts to limit the amount of supervision by dispensing the need for bounding boxes, but still assumes image-level labels on the entire training set are available. In this work, we study the problem of training an object detector from one or few clean images with image-level labels and a larger set of completely unlabeled images [43]. This is an extreme case of semi-supervised learning where the labeled data are not enough to bootstrap the learning of a classifier or detector. Our solution is to use a standard weakly-supervised pipeline to train a student model from image-level pseudo-labels generated on the unlabeled set by a teacher model, bootstrapped by region-level similarities to clean labeled images. By using the recent pipeline of PCL and more unlabeled images, we achieve performance competitive or superior to many state of the art weakly-supervised detection solutions.