Representation In Layer Research Articles

A semantic guidance-based fusion network for multi-label image classification

Multi-label image classification (MLIC), a fundamental task assigning multiple labels to each image, has been seen notable progress in recent years. Considering simultaneous appearances of objects in the physical world, modeling object correlations is crucial for enhancing classification accuracy. This involves accounting for spatial image feature correlation and label semantic correlation. However, existing methods struggle to establish these correlations due to complex spatial location and label semantic relationships. On the other hand, regarding the fusion of image feature relevance and label semantic relevance, existing methods typically learn a semantic representation in the final CNN layer to combine spatial and label semantic correlations. However, different CNN layers capture features at diverse scales and possess distinct discriminative abilities. To address these issues, in this paper we introduce the Semantic Guidance-Based Fusion Network (SGFN) for MLIC. To model spatial image feature correlation, we leverage the advanced TResNet architecture as the backbone network and employ the Feature Aggregation Module for capturing global spatial correlation. For label semantic correlation, we establish both local and global semantic correlation. We further enrich model features by learning semantic representations across multiple convolutional layers. Our method outperforms current state-of-the-art techniques on PASCAL VOC (2007, 2012) and MS-COCO datasets.

Extreme temperature events monitored by Raman lidar: Consistency and complementarity with spaceborne observations and modelling

AbstractTwo extreme temperature events (ETEs), both triggered by atmospheric blocking situations, were monitored by a pure rotational Raman lidar 20 km south of Paris. The first ETE lasted 6 days in late February 2018 with the onset of a cold wave surging from Siberia, and the second was a short heatwave that occurred at the end of July 2020. After calibration using simultaneous radiosoundings, lidar vertical profiles of temperature are derived with an uncertainty of less than 1 °C within both the planetary boundary layer and the lower/middle free troposphere (~3/6 km above ground level [a.g.l.] for day/night), with a final vertical resolution of 50 m and a temporal resolution of 30 min. Such capabilities fulfil the observational requirements of the World Meteorological Organisation and contribute to fill a gap left behind by current operational instruments, which struggle to follow the diurnal cycle of the planetary boundary layer due to insufficient spatial and temporal resolutions in the low troposphere. These case studies allow to assess how such weather events are represented by numerical weather prediction modelling and whether they can be correctly observed by the Infrared Atmospheric Sounding Interferometer (IASI) spaceborne mission. Our results demonstrate that during these events both ECMWF reanalyses and forecasts are in overall very good accordance with lidar observations (correlation >0.9 in average). Still, they show warm biases (~1°C–2°C) compared with the lidar and improvable temperature field representation in the boundary layer. The radiances measured by IASI, assimilated in the operational model, do not capture temperature profiles adequately below ~2–3 km a.g.l., mainly due to the altitude location of its weighting functions. The bias computed against the lidar reaches 2°C during the heatwave and 4°C during the cold wave case in the first 2 km of the atmosphere, beyond expected errors found in the literature.

MLIM: A CTR prediction model describing evolution law of user interest

With the advent of the digital economy era, business systems such as web advertising and recommendation system have put forward the demand for predicting the click through rate (CTR) of items. However, the current CTR prediction research is not enough to mine user behavior, resulting in the lack of accuracy of user interest representation. In this paper, we propose a CTR prediction model, called MLIM, which can deep mine the evolution law of user interest. Specifically, we first use BiGRU to obtain the low-level user interest representation in the interest extraction layer, and then continue to use attention mechanism, BiGRU and sliding time window multi-components collaborative modeling in the interest evolution layer to obtain multi-level user interest representation with richer information, which can improve the accuracy of CTR prediction to a certain extent. Comprehensive experiments on two real datasets show that the proposed model achieves better performance than the mainstream baselines integrating user behavior analysis.

Deep Matrix Factorization for Trust-Aware Recommendation in Social Networks

Recent years have witnessed remarkable information overload in online social networks, and social network based approaches for recommender systems have been widely studied. The trust information in social networks among users is an important factor for improving recommendation performance. Many successful recommendation tasks are treated as the matrix factorization problems. However, the prediction performance of matrix factorization based methods largely depends on the matrixes initialization of users and items. To address this challenge, we develop a novel trust-aware approach based on deep learning to alleviate the initialization dependence. First, we propose two deep matrix factorization (DMF) techniques, i.e., linear DMF and non-linear DMF to extract features from the user-item rating matrix for improving the initialization accuracy. The trust relationship is integrated into the DMF model according to the preference similarity and the derivations of users on items. Second, we exploit deep marginalized Denoising Autoencoder (Deep-MDAE) to extract the latent representation in the hidden layer from the trust relationship matrix to approximate the user factor matrix factorized from the user-item rating matrix. The community regularization is integrated in the joint optimization function to take neighbours’ effects into consideration. The results of DMF are applied to initialize the updating variables of Deep-MDAE in order to further improve the recommendation performance. Finally, we validate that the proposed approach outperforms state-of-the-art baselines for recommendation, especially for the cold-start users.

Effect of Top-Down Connections in Hierarchical Sparse Coding

Hierarchical sparse coding (HSC) is a powerful model to efficiently represent multidimensional, structured data such as images. The simplest solution to solve this computationally hard problem is to decompose it into independent layer-wise subproblems. However, neuroscientific evidence would suggest interconnecting these subproblems as in predictive coding (PC) theory, which adds top-down connections between consecutive layers. In this study, we introduce a new model, 2-layer sparse predictive coding (2L-SPC), to assess the impact of this interlayer feedback connection. In particular, the 2L-SPC is compared with a hierarchical Lasso (Hi-La) network made out of a sequence of independent Lasso layers. The 2L-SPC and a 2-layer Hi-La networks are trained on four different databases and with different sparsity parameters on each layer. First, we show that the overall prediction error generated by 2L-SPC is lower thanks to the feedback mechanism as it transfers prediction error between layers. Second, we demonstrate that the inference stage of the 2L-SPC is faster to converge and generates a refined representation in the second layer compared to the Hi-La model. Third, we show that the 2L-SPC top-down connection accelerates the learning process of the HSC problem. Finally, the analysis of the emerging dictionaries shows that the 2L-SPC features are more generic and present a larger spatial extension.

Tree recruitment in a drought- and herbivory-stressed oak-beech forest: Implications for future species coexistence

Mixed oak-beech forests are particularly valuable from an ecological and socioeconomic perspective but are increasingly exposed to anthropogenic impacts. We wonder whether anthropogenic change drivers, such as increased drought and ungulate populations affect optimal regeneration niches, species coexistence, and mixed-species forest dynamics. We compared understorey composition, structure and distribution with the overstorey counterpart in an oak-beech forest located at its southernmost European distribution (Central Spain) with increasing drought and ungulate herbivory for the last decades.Results showed that Ilex aquifolium and Prunus avium had greater representation in the recruitment layer as compared to the understorey layer whereas, for the dominant Quercus spp. and Fagus sylvatica, there was a tendency suggesting an expansion of Fagus at the expense of Quercus. Density of tree recruits showed strong differences across microsites. The greatest density was found under tree and under shrub cover whereas the lowest was found under woody debris. By species, Quercus regenerated better in open and shrub microsites whereas Fagus and Ilex regeneration was higher under tree cover. Large patches of shrub cover favoured Prunus avium recruitment, particularly against ungulate herbivory as it was the most heavily browsed species, followed by Quercus, Ilex and Fagus in a descending order. Herbivory occurrence and intensity on tree recruits strongly varied with the microsite, with greater herbivory levels in open microsites. Woody debris on the ground only protected tree recruits from browsing when they were located under tree cover.Overall, fenced areas with no ungulates showed greater tree species coexistence (28% increase in species richness) and recruitment abundance (81% increase) after 7 years of exclusion. We also found a large increase in tree recruit growth and height:diameter ratio (h:d) for all studied species when fenced. Fagus and Prunus avium showed the greatest growth and h:d increase, much greater than those of Quercus and Ilex. The h:d was identified as a reliable indicator of ungulate herbivory and reference values (under null levels of ungulate herbivory) were provided for each species.We conclude that there is an imperative need for adaptive management strategies that address the joint effects of climate change (e.g. increased drought) and ungulate herbivory in mixed oak-beech forests. The provision of different microsites, including patches of shrub cover in open microsites and woody debris under dense tree cover, will facilitate the availability of suitable regeneration niches for most species and will therefore promote future species coexistence.

Multi-Element Hierarchical Attention Capsule Network for Stock Prediction

Stock prediction is a challenging task concerned by researchers due to its considerable returns. It is difficult because of the high randomness in the stock market. Stock price movement is mainly related to the capital situation and hot events. In recent years, researchers improved prediction accuracy with news and social media. However, the existing methods do not take into account the different influences of events. To solve this problem, we propose a multi-element hierarchical attention capsule network, which consists of two components. The former component, multi-element hierarchical attention, quantifies the importance of valuable information contained in multiple news and social media through its weights assignment process. And the latter component, capsule network, learns more context information from the events through its vector representation in the hidden layer. Moreover, we construct a combined data set to maintain the complementarity between social media and news. Finally, we achieve better results than baselines, and experiments show that our model improves prediction accuracy by quantifying the different influences of events.

Evaluation of Convolution Operation Based on the Interpretation of Deep Learning on 3-D Point Cloud

The interpretation of deep learning network is an important part in understanding the convolutional neural networks (CNNs). As an exploratory research, this article explored the interpretation method in 3-D point cloud deep learning networks, for the purpose of evaluating the performance of convolution functions in 3-D point cloud CNNs. Specifically, a 3-D point cloud classification network with two branches is used as the interpretation network in two aspects; 1) information entropy is introduced to diagnose the internal representation in the middle layer of CNN; and 2) the external consistency of convolution function is measured by per-point classification accuracy with class activation mapping technique. Four typical convolution functions are tested by the interpretation network on ModelNet40 dataset and the experimental results demonstrate that the proposed evaluation method is reliable. Feature transformation ability and feature recognition ability of convolution functions are extracted by visualization evaluation and proposed measurable metrics evaluation.

Layer-specific representation of long-lasting sustained activity in the rat auditory cortex

In the auditory system, distinct and reproducible transient activities responding to the onset of sound have long been the focus when characterizing the auditory cortex, i.e., tonotopic maps, subregions, and layer-specific representation. There is limited information on sustained activities because the rapid adaptation impairs reproducibility and the signal-to-noise ratio. We recently overcame this problem by focusing on neural synchrony and machine learning demonstrated that band-specific power and the phase locking value (PLV) represent sound information in a tonotopic and region-specific manner. Here, we attempted to reveal the layer-specific representation of sustained activities. A microelectrode array recorded sustained activities from layers 2/3, 4, and 5/6 of the rat auditory cortex. We characterized band-specific power and PLV patterns and applied sparse logistic regression (SLR) to discriminate (1) between the sound-induced and spontaneous activities and (2) five test frequencies from the sound-induced activities in each layer. SLR achieved the highest discrimination performance in high-gamma activities in layers 4 and 5/6, higher than in layer 2/3, indicating poor sound representation in layer 2/3. Moreover, the recording sites that contributed to the discrimination in layers 4 and 5/6 had a characteristic frequency similar to the test frequency and were often located in the belt area, indicating tonotopic and region-specific representation. These results indicate that information processing of sustained activities may depend on high-gamma oscillators, i.e., cortical inhibitory interneurons, and reflects layer-specific thalamocortical and corticocortical neural circuits in the auditory system, which may contribute to associative information processing beyond sound frequency in auditory perception.

Distillation of the clinical algorithm improves prognosis by multi-task deep learning in high-risk Neuroblastoma.

We introduce the CDRP (Concatenated Diagnostic-Relapse Prognostic) architecture for multi-task deep learning that incorporates a clinical algorithm, e.g., a risk stratification schema to improve prognostic profiling. We present the first application to survival prediction in High-Risk (HR) Neuroblastoma from transcriptomics data, a task that studies from the MAQC consortium have shown to remain the hardest among multiple diagnostic and prognostic endpoints predictable from the same dataset. To obtain a more accurate risk stratification needed for appropriate treatment strategies, CDRP combines a first component (CDRP-A) synthesizing a diagnostic task and a second component (CDRP-N) dedicated to one or more prognostic tasks. The approach leverages the advent of semi-supervised deep learning structures that can flexibly integrate multimodal data or internally create multiple processing paths. CDRP-A is an autoencoder trained on gene expression on the HR/non-HR risk stratification by the Children’s Oncology Group, obtaining a 64-node representation in the bottleneck layer. CDRP-N is a multi-task classifier for two prognostic endpoints, i.e., Event-Free Survival (EFS) and Overall Survival (OS). CDRP-A provides the HR embedding input to the CDRP-N shared layer, from which two branches depart to model EFS and OS, respectively. To control for selection bias, CDRP is trained and evaluated using a Data Analysis Protocol (DAP) developed within the MAQC initiative. CDRP was applied on Illumina RNA-Seq of 498 Neuroblastoma patients (HR: 176) from the SEQC study (12,464 Entrez genes) and on Affymetrix Human Exon Array expression profiles (17,450 genes) of 247 primary diagnostic Neuroblastoma of the TARGET NBL cohort. On the SEQC HR patients, CDRP achieves Matthews Correlation Coefficient (MCC) 0.38 for EFS and MCC = 0.19 for OS in external validation, improving over published SEQC models. We show that a CDRP-N embedding is indeed parametrically associated to increasing severity and the embedding can be used to better stratify patients’ survival.

The English of Bosnia and Herzegovina

AbstractWith the number of world Englishes steadily increasing, the impact of ‘core’ English varieties is weakening. However, the influence of two varieties, namely standardised American English and standardised British English, seems to be dispersed across all areas, in particular across peripheral varieties. Due to the difference in their demographic weight and the institutional support they receive, these two varieties do not secure a balanced representation in the peripheral layer. The current study investigates whether 132 university‐level Bosnian learners of English report using particular items of American or British English at the levels of pronunciation, spelling, grammar and vocabulary.

Serotonin regulates dynamics of cerebellar granule cell activity by modulating tonic inhibition.

Understanding how afferent information is integrated by cortical structures requires identifying the factors shaping excitation and inhibition within their input layers. The input layer of the cerebellar cortex integrates diverse sensorimotor information to enable learned associations that refine the dynamics of movement. Specifically, mossy fiber afferents relay sensorimotor input into the cerebellum to excite granule cells, whose activity is regulated by inhibitory Golgi cells. To test how this integration can be modulated, we have used an acute brain slice preparation from young adult rats and found that encoding of mossy fiber input in the cerebellar granule cell layer can be regulated by serotonin (5-hydroxytryptamine, 5-HT) via a specific action on Golgi cells. We find that 5-HT depolarizes Golgi cells, likely by activating 5-HT2A receptors, but does not directly act on either granule cells or mossy fibers. As a result of Golgi cell depolarization, 5-HT significantly increases tonic inhibition onto both granule cells and Golgi cells. 5-HT-mediated Golgi cell depolarization is not sufficient, however, to alter the probability or timing of mossy fiber-evoked feed-forward inhibition onto granule cells. Together, increased granule cell tonic inhibition paired with normal feed-forward inhibition acts to reduce granule cell spike probability without altering spike timing. Hence, these data provide a circuit mechanism by which 5-HT can reduce granule cell activity without altering temporal representations of mossy fiber input. Such changes in network integration could enable flexible, state-specific suppression of cerebellar sensorimotor input that should not be learned or enable reversal learning for unwanted associations. NEW & NOTEWORTHY Serotonin (5-hydroxytryptamine, 5-HT) regulates synaptic integration at the input stage of cerebellar processing by increasing tonic inhibition of granule cells. This circuit mechanism reduces the probability of granule cell spiking without altering spike timing, thus suppressing cerebellar input without altering its temporal representation in the granule cell layer.

Refinement of learned skilled movement representation in motor cortex deep output layer

The mechanisms underlying the emergence of learned motor skill representation in primary motor cortex (M1) are not well understood. Specifically, how motor representation in the deep output layer 5b (L5b) is shaped by motor learning remains virtually unknown. In rats undergoing motor skill training, we detect a subpopulation of task-recruited L5b neurons that not only become more movement-encoding, but their activities are also more structured and temporally aligned to motor execution with a timescale of refinement in tens-of-milliseconds. Field potentials evoked at L5b in vivo exhibit persistent long-term potentiation (LTP) that parallels motor performance. Intracortical dopamine denervation impairs motor learning, and disrupts the LTP profile as well as the emergent neurodynamical properties of task-recruited L5b neurons. Thus, dopamine-dependent recruitment of L5b neuronal ensembles via synaptic reorganization may allow the motor cortex to generate more temporally structured, movement-encoding output signal from M1 to downstream circuitry that drives increased uniformity and precision of movement during motor learning.

A Biologically Inspired Appearance Model for Robust Visual Tracking.

In this paper, we propose a biologically inspired appearance model for robust visual tracking. Motivated in part by the success of the hierarchical organization of the primary visual cortex (area V1), we establish an architecture consisting of five layers: whitening, rectification, normalization, coding, and pooling. The first three layers stem from the models developed for object recognition. In this paper, our attention focuses on the coding and pooling layers. In particular, we use a discriminative sparse coding method in the coding layer along with spatial pyramid representation in the pooling layer, which makes it easier to distinguish the target to be tracked from its background in the presence of appearance variations. An extensive experimental study shows that the proposed method has higher tracking accuracy than several state-of-the-art trackers.

The role of the different layers of primary visual cortex in working memory

Imaging studies have revealed a neuronal correlate of working memory in primary visual cortex (Harrison & Tong, Nature, 2009). However, it is unknown if working memories influence spiking activity in the primary visual cortex. To address this question, we recorded neuronal activity in the primary visual cortex of monkeys trained to perform attentional and working memory tasks with a probe that records activity in all the cortical layers. We found a consistent working memory trace in the spiking activity in the superficial and deep layers of monkey V1, and only a weak memory representation in input layer 4. This V1 memory trace could be disrupted with a visual mask, but it then quickly recovered. The advantage of the laminar probe is that is also gives insight into the current-source density, which reveals the putative synaptic sources of memory activity. The current-source density measurements revealed a characteristic signature of feedback processing with putative synaptic inputs in the superficial and deep layers for working memory. This signature resembles the signature of selective attention, supporting the view that top-down modulation of activity in primary visual cortex underlies both working memory and attention. Our results provide new insights into the role of early visual cortex in working memory.

Correction: Clancy et al., Structure of a Single Whisker Representation in Layer 2 of Mouse Somatosensory Cortex

Correction: Clancy et al., Structure of a Single Whisker Representation in Layer 2 of Mouse Somatosensory Cortex

Rare diseases in ICD11: making rare diseases visible in health information systems through appropriate coding.

BackgroundBecause of their individual rarity, genetic diseases and other types of rare diseases are under-represented in healthcare coding systems; this contributes to a lack of ascertainment and recognition of their importance for healthcare planning and resource allocation, and prevents clinical research from being performed.MethodsOrphanet was given the task to develop an inventory of rare diseases and a classification system which could serve as a template to update International terminologies. When the World Health Organization (WHO) launched the revision process of the International Classification of Diseases (ICD), a Topic Advisory Group for rare diseases was established, managed by Orphanet and funded by the European Commission.ResultsSo far 5,400 rare diseases listed in the Orphanet database have an endorsed representation in the foundation layer of ICD-11, and are thus provided with a unique identifier in the Beta version of ICD-11, which is 10 times more than in ICD10. A rare disease linearization is also planned. The current beta version is open for public consultation and comments, and to be used for field testing. The adoption by the World Health Assembly is planned for 2017.ConclusionsThe overall revision process was carried out with very limited means considering its scope, ambition and strategic significance, and experienced significant hurdles and setbacks. The lack of funding impacted the level of professionalism that could be attained. The contrast between the initially declared goals and the currently foreseen final product is disappointing. In the context of uncertainty around the outcome of the field testing and the potential willingness of countries to adopt this new version, the European Commission Expert Group on Rare Diseases adopted in November 2014 a recommendation for health care coding systems to consider using ORPHA codes in addition to ICD10 codes for rare diseases having no specific ICD10 codes. The Orphanet terminology, classifications and mappings with other terminologies are freely available at www.orphadata.org.

Structure of a single whisker representation in layer 2 of mouse somatosensory cortex.

Layer (L)2 is a major output of primary sensory cortex that exhibits very sparse spiking, but the structure of sensory representation in L2 is not well understood. We combined two-photon calcium imaging with deflection of many whiskers to map whisker receptive fields, characterize sparse coding, and quantitatively define the point representation in L2 of mouse somatosensory cortex. Neurons within a column-sized imaging field showed surprisingly heterogeneous, salt-and-pepper tuning to many different whiskers. Single whisker deflection elicited low-probability spikes in highly distributed, shifting neural ensembles spanning multiple cortical columns. Whisker-evoked response probability correlated strongly with spontaneous firing rate, but weakly with tuning properties, indicating a spectrum of inherent responsiveness across pyramidal cells. L2 neurons projecting to motor and secondary somatosensory cortex differed in whisker tuning and responsiveness, and carried different amounts of information about columnar whisker deflection. From these data, we derive a quantitative, fine-scale picture of the distributed point representation in L2.

A feedforward inhibitory circuit mediates lateral refinement of sensory representation in upper layer 2/3 of mouse primary auditory cortex.

Sensory information undergoes ordered and coordinated processing across cortical layers. Whereas cortical layer (L) 4 faithfully acquires thalamic information, the superficial layers appear well staged for more refined processing of L4-relayed signals to generate corticocortical outputs. However, the specific role of superficial layer processing and how it is specified by local synaptic circuits remains not well understood. Here, in the mouse primary auditory cortex, we showed that upper L2/3 circuits play a crucial role in refining functional selectivity of excitatory neurons by sharpening auditory tonal receptive fields and enhancing contrast of frequency representation. This refinement is mediated by synaptic inhibition being more broadly recruited than excitation, with the inhibition predominantly originating from interneurons in the same cortical layer. By comparing the onsets of synaptic inputs as well as of spiking responses of different types of neuron, we found that the broadly tuned, fast responding inhibition observed in excitatory cells can be primarily attributed to feedforward inhibition originating from parvalbumin (PV)-positive neurons, whereas somatostatin (SOM)-positive interneurons respond much later compared with the onset of inhibitory inputs to excitatory neurons. We propose that the feedforward circuit-mediated inhibition from PV neurons, which has an analogous function to lateral inhibition, enables upper L2/3 excitatory neurons to rapidly refine auditory representation.

Training-dependent associative learning induced neocortical structural plasticity: a trace eyeblink conditioning analysis.

Studies utilizing general learning and memory tasks have suggested the importance of neocortical structural plasticity for memory consolidation. However, these learning tasks typically result in learning of multiple different tasks over several days of training, making it difficult to determine the synaptic time course mediating each learning event. The current study used trace-eyeblink conditioning to determine the time course for neocortical spine modification during learning. With eyeblink conditioning, subjects are presented with a neutral, conditioned stimulus (CS) paired with a salient, unconditioned stimulus (US) to elicit an unconditioned response (UR). With multiple CS-US pairings, subjects learn to associate the CS with the US and exhibit a conditioned response (CR) when presented with the CS. Trace conditioning is when there is a stimulus free interval between the CS and the US. Utilizing trace-eyeblink conditioning with whisker stimulation as the CS (whisker-trace-eyeblink: WTEB), previous findings have shown that primary somatosensory (barrel) cortex is required for both acquisition and retention of the trace-association. Additionally, prior findings demonstrated that WTEB acquisition results in an expansion of the cytochrome oxidase whisker representation and synaptic modification in layer IV of barrel cortex. To further explore these findings and determine the time course for neocortical learning-induced spine modification, the present study utilized WTEB conditioning to examine Golgi-Cox stained neurons in layer IV of barrel cortex. Findings from this study demonstrated a training-dependent spine proliferation in layer IV of barrel cortex during trace associative learning. Furthermore, findings from this study showing that filopodia-like spines exhibited a similar pattern to the overall spine density further suggests that reorganization of synaptic contacts set the foundation for learning-induced neocortical modifications through the different neocortical layers.