Feature blindness: A challenge for understanding and modelling visual object recognition

Humans rely heavily on the shape of objects to recognise them. Recently, it has been argued that Convolutional Neural Networks (CNNs) can also show a shape-bias, provided their learning environment contains this bias. This has led to the proposal that CNNs provide good mechanistic models of shape-bias and, more generally, human visual processing. However, it is also possible that humans and CNNs show a shape-bias for very different reasons, namely, shape-bias in humans may be a consequence of architectural and cognitive constraints whereas CNNs show a shape-bias as a consequence of learning the statistics of the environment. We investigated this question by exploring shape-bias in humans and CNNs when they learn in a novel environment. We observed that, in this new environment, humans (i) focused on shape and overlooked many non-shape features, even when non-shape features were more diagnostic, (ii) learned based on only one out of multiple predictive features, and (iii) failed to learn when global features, such as shape, were absent. This behaviour contrasted with the predictions of a statistical inference model with no priors, showing the strong role that shape-bias plays in human feature selection. It also contrasted with CNNs that (i) preferred to categorise objects based on non-shape features, and (ii) increased reliance on these non-shape features as they became more predictive. This was the case even when the CNN was pre-trained to have a shape-bias and the convolutional backbone was frozen. These results suggest that shape-bias has a different source in humans and CNNs: while learning in CNNs is driven by the statistical properties of the environment, humans are highly constrained by their previous biases, which suggests that cognitive constraints play a key role in how humans learn to recognise novel objects.     

Noninvasive cellular internalization of silver molecules by chitosan nanoneedles: a novel nanocarrier

We explore with molecular modeling, dynamics simulations, and a statistical model the ability of chitosan nanoneedles (CNNs) to be internalized into a model lipid bilayer as a function of their length, keeping in view of their applications in the field of biomedicine for advanced targeted drug delivery. In this study, we have computationally modeled and studied the structural geometry and the stability of CNNs formed by 4, 6, and 8 subunits. We reported the molecular surface analysis of the modeled CNNs along with molecular dynamic (MD) simulations studies toward revealing the noninvasive cellular internalization potential of these CNNs and a case study has been carried to study the ability of CNNs to translocate silver nanoparticles across membrane. The present results are strongly in support of further exploration of 8 subunits based CNNs for their application as target drug delivery vehicles. The hydrophilicity of the CNNs has been attributed as one of the key factors responsible for the internalization process. Moreover, our MD simulation studies marched the ability of CNNs to translocate silver nanoparticles through biological membrane in a similar manner that resembles cell-penetrating peptides.     

Distributed sensor networks: a cellular nonlinear network perspective

Large-scale networks of integrated wireless sensors become increasingly tractable. Advances in hardware technology and engineering design have led to dramatic reductions in size, power consumption, and cost for digital circuitry, and wireless communications. Networking, self-organization, and distributed operation are crucial ingredients to harness the sensing, computing, and computational capabilities of the nodes into a complete system. This article shows that those networks can be considered as cellular nonlinear networks (CNNs), and that their analysis and design may greatly benefit from the rich theoretical results available for CNNs.     

Learning the local landscape of protein structures with convolutional neural networks

One fundamental problem of protein biochemistry is to predict protein structure from amino acid sequence. The inverse problem, predicting either entire sequences or individual mutations that are consistent with a given protein structure, has received much less attention even though it has important applications in both protein engineering and evolutionary biology. Here, we ask whether 3D convolutional neural networks (3D CNNs) can learn the local fitness landscape of protein structure to reliably predict either the wild-type amino acid or the consensus in a multiple sequence alignment from the local structural context surrounding site of interest. We find that the network can predict wild type with good accuracy, and that network confidence is a reliable measure of whether a given prediction is likely going to be correct or not. Predictions of consensus are less accurate and are primarily driven by whether or not the consensus matches the wild type. Our work suggests that high-confidence mis-predictions of the wild type may identify sites that are primed for mutation and likely targets for protein engineering.

     

A new method for the re-implementation of threshold logic functions with cellular neural networks

A new strategy is presented for the implementation of threshold logic functions with binary-output Cellular Neural Networks (CNNs). The objective is to optimize the CNNs weights to develop a robust implementation. Hence, the concept of generative set is introduced as a convenient representation of any linearly separable Boolean function. Our analysis of threshold logic functions leads to a complete algorithm that automatically provides an optimized generative set. New weights are deduced and a more robust CNN template assuming the same function can thus be implemented. The strategy is illustrated by a detailed example.     

Characterizing robustness and sensitivity of convolutional neural networks for quantitative analysis of mitochondrial morphology

Background: Quantitative analysis of mitochondrial morphology plays important roles in studies of mitochondrial biology. The analysis depends critically on segmentation of mitochondria, the image analysis process of extracting mitochondrial morphology from images. The main goal of this study is to characterize the performance of convolutional neural networks (CNNs) in segmentation of mitochondria from fluorescence microscopy images. Recently, CNNs have achieved remarkable success in challenging image segmentation tasks in several disciplines. So far, however, our knowledge of their performance in segmenting biological images remains limited. In particular, we know little about their robustness, which defines their capability of segmenting biological images of different conditions, and their sensitivity, which defines their capability of detecting subtle morphological changes of biological objects. Methods: We have developed a method that uses realistic synthetic images of different conditions to characterize the robustness and sensitivity of CNNs in segmentation of mitochondria. Using this method, we compared performance of two widely adopted CNNs: the fully convolutional network (FCN) and the U-Net. We further compared the two networks against the adaptive active-mask (AAM) algorithm, a representative of high-performance conventional segmentation algorithms. Results: The FCN and the U-Net consistently outperformed the AAM in accuracy, robustness, and sensitivity, often by a significant margin. The U-Net provided overall the best performance. Conclusions: Our study demonstrates superior performance of the U-Net and the FCN in segmentation of mitochondria. It also provides quantitative measurements of the robustness and sensitivity of these networks that are essential to their applications in quantitative analysis of mitochondrial morphology.     

Toxic gas dispersion prediction for point source emission using deep learning method

Abstract

Accurate and rapid toxic gas concentration prediction model plays an important role in emergency aid of sudden gas leak. However, it is difficult for existing dispersion model to achieve accuracy and efficiency requirements at the same time. Although some researchers have considered developing new forecasting models with traditional machine learning, such as back propagation (BP) neural network, support vector machine (SVM), the prediction results obtained from such models need to be improved still in terms of accuracy. Then new prediction models based on deep learning are proposed in this paper. Deep learning has obvious advantages over traditional machine learning in prediction and classification. Deep belief networks (DBNs) as well as convolution neural networks (CNNs) are used to build new dispersion models here. Both models are compared with Gaussian plume model, computation fluid dynamics (CFD) model and models based on traditional machine learning in terms of accuracy, prediction time, and computation time. The experimental results turn out that CNNs model performs better considering all evaluation indexes.     

Passive acoustic monitoring of animal populations with transfer learning

Progress in deep learning, more specifically in using convolutional neural networks (CNNs) for the creation of classification models, has been tremendous in recent years. Within bioacoustics research, there has been a large number of recent studies that use CNNs. Designing CNN architectures from scratch is non-trivial and requires knowledge of machine learning. Furthermore, hyper-parameter tuning associated with CNNs is extremely time consuming and requires expensive hardware. In this paper we assess whether it is possible to build good bioacoustic classifiers by adapting and re-using existing CNNs pre-trained on the ImageNet dataset – instead of designing them from scratch, a strategy known as transfer learning that has proved highly successful in other domains. This study is a first attempt to conduct a large-scale investigation on how transfer learning can be used for passive acoustic monitoring (PAM), to simplify the implementation of CNNs and the design decisions when creating them, and to remove time consuming hyper-parameter tuning phases. We compare 12 modern CNN architectures across 4 passive acoustic datasets that target calls of the Hainan gibbon Nomascus hainanus, the critically endangered black-and-white ruffed lemur Varecia variegata, the vulnerable Thyolo alethe Chamaetylas choloensis, and the Pin-tailed whydah Vidua macroura. We focus our work on data scarcity issues by training PAM binary classification models very small datasets, with as few as 25 verified examples. Our findings reveal that transfer learning can result in up to 82% F1 score while keeping CNN implementation details to a minimum, thus rendering this approach accessible, easier to design, and speeding up further vocalisation annotations to create PAM robust models.     

Complexity and diversity in sparse code priors improve receptive field characterization of Macaque V1 neurons

System identification techniques—projection pursuit regression models (PPRs) and convolutional neural networks (CNNs)—provide state-of-the-art performance in predicting visual cortical neurons’ responses to arbitrary input stimuli. However, the constituent kernels recovered by these methods are often noisy and lack coherent structure, making it difficult to understand the underlying component features of a neuron’s receptive field. In this paper, we show that using a dictionary of diverse kernels with complex shapes learned from natural scenes based on efficient coding theory, as the front-end for PPRs and CNNs can improve their performance in neuronal response prediction as well as algorithmic data efficiency and convergence speed. Extensive experimental results also indicate that these sparse-code kernels provide important information on the component features of a neuron’s receptive field. In addition, we find that models with the complex-shaped sparse code front-end are significantly better than models with a standard orientation-selective Gabor filter front-end for modeling V1 neurons that have been found to exhibit complex pattern selectivity. We show that the relative performance difference due to these two front-ends can be used to produce a sensitive metric for detecting complex selectivity in V1 neurons.     

Nonhanded chirality in octahedral metal complexes

Chiral molecules can either be handed (i.e., "shoes") or nonhanded ("potatoes"). The only chiral ligand partition for tetrahedral metal complexes (or for a tetrahedral carbon atom such as that found in amino acids and other chiral biological molecules) is the fully unsymmetrical degree 6 partition (1(4)), which leads to handed metal complexes of the type MABCD with a lowest-degree chirality polynomial consisting of the product of all six possible linear factors of the type (s(i)-s(j)) where 1 < or = i,j < or = 4. The lowest-degree chiral ligand partitions for octahedral metal complexes are the degree 6 partitions (31(3)) and (2(3)) leading to handed chiral metal complexes of the types fac-MA(3)BCD and cis-MA(2)B(2)C(2). The form of the lowest-degree chirality polynomial for the (31(3)) chiral ligand partition of the octahedron resembles that of the (1(4)) chiral ligand partition of the tetrahedron, likewise with four different ligands. However, the form of the lowest-degree chirality polynomial for the (2(3)) chiral ligand partition of the octahedron corresponds to the square of the chirality polynomial of the (1(3)) chiral ligand partition of the polarized triangle, which likewise has three different ligands. Ligand partitions for octahedral metal complexes such as (2(2)1(2)), (21(4)), and (1(6)), which are less symmetrical than the lowest-degree chiral ligand partitions (31(3)) and (2(3)), lead to chiral octahedral metal complexes which are nonhanded. In such complexes, pairs of enantiomers can be interconverted by simple ligand interchanges without ever going through an achiral intermediate.     

Conditions for the various types of co-operativity allowed by the adair,equation and their relation with the algebraic character of binding polynomials

A convenient way to obtain for any number, n, of sites, the functions of the constants of the Adair equation that decide the type of co-operativity of ligand binding to a non-dissociating protein is given and is illustrated by the examples n = 4 and n = 5. These functions are invariants of the binding polynomial and various of its derivatives.Although there are some simple sufficient conditions (inequalities relating successive Adair constants) for some co-operativity types, the full necessary and sufficient conditions even for uniform positive and negative co-operativity depend on very complicated functions of the constants for n > 4.However there are alternative ways of writing binding polynomials known as canonical forms. Up to at least n = 5, and probably beyond, the conditions that are complicated in terms of Adair constants are very simple in terms of the constants of canonical forms. For instance any fourth-degree polynomial can be written in the form p(x - α)⁴ + q(x - β)⁴ + 6μ (x - α)²(x - β)² although in three different ways. For one of these ways, the sign of μ distinguishes between mixed and uniform co-operativity. For any kind of mixed co-operativity μ > 0, while μ < 0 corresponds to uniform co-operativity. Advantages of the use of canonical forms are briefly commented on.     

Osmometric analysis of quasi-indefinite mixed associations of the first kind

Mixed associations of the type A + B----AB, A + AB----A2B, ..., A + Ai-1 B----AiB, ... are readily analyzed by osmometric methods. The equilibrium molar concentration of A, mA, is obtained very simply from mA = meq-m0B; here meq = c/Meqn is the equilibrium molar concentration of all associating species and m0B denotes the stoichiometric or original molar concentration of B. The quantity mB can then be obtained from methods developed by Steiner. The value of the binding polynomial lambda is given by lambda = m0B/mB; lambda is a function of mA only. In principle, one can evaluate the equilibrium constants (kA,B,etc.) by fitting lambda to the appropriate polynomial in mA of degree n (n = 2, 3, ...). The binding polynomial lambda is analogous to polynomials encountered in the analysis of self-associations. By making some simple assumptions one can develop four analogs of two sequential, equal equilibrium constant (SEK) or two attenuated equilibrium constant (AK) models. With the aid of r (the number average degree of binding), g (the osmotic coefficient), lambda, as well as mA and mB, one can evaluate the equilibrium constant or constants. The methods developed here can be extended to the nonideal case.     

Fusing multi-season UAS images with convolutional neural networks to map tree species in Amazonian forests

Remote sensing images obtained by unoccupied aircraft systems (UAS) across different seasons enabled capturing of species-specific phenological patterns of tropical trees. The application of UAS multi-season images to classify tropical tree species is still poorly understood. In this study, we used RGB images from different seasons obtained by a low-cost UAS and convolutional neural networks (CNNs) to map tree species in an Amazonian forest. Individual tree crowns (ITC) were outlined in the UAS images and identified to the species level using forest inventory data. The CNN model was trained with images obtained in February, May, August, and November. The classification accuracy in the rainy season (November and February) was higher than in the dry season (May and August). Fusing images from multiple seasons improved the average accuracy of tree species classification by up to 21.1 percentage points, reaching 90.5%. The CNN model can learn species-specific phenological characteristics that impact the classification accuracy, such as leaf fall in the dry season, which highlights its potential to discriminate species in various conditions. We produced high-quality individual tree crown maps of the species using a post-processing procedure. The combination of multi-season UAS images and CNNs has the potential to map tree species in the Amazon, providing valuable insights for forest management and conservation initiatives.     

Classifying T cell activity in autofluorescence intensity images with convolutional neural networks

The importance of T cells in immunotherapy has motivated developing technologies to improve therapeutic efficacy. One objective is assessing antigen‐induced T cell activation because only functionally active T cells are capable of killing the desired targets. Autofluorescence imaging can distinguish T cell activity states in a non‐destructive manner by detecting endogenous changes in metabolic co‐enzymes such as NAD(P)H. However, recognizing robust activity patterns is computationally challenging in the absence of exogenous labels. We demonstrate machine learning methods that can accurately classify T cell activity across human donors from NAD(P)H intensity images. Using 8260 cropped single‐cell images from six donors, we evaluate classifiers ranging from traditional models that use previously‐extracted image features to convolutional neural networks (CNNs) pre‐trained on general non‐biological images. Adapting pre‐trained CNNs for the T cell activity classification task provides substantially better performance than traditional models or a simple CNN trained with the autofluorescence images alone. Visualizing the images with dimension reduction provides intuition into why the CNNs achieve higher accuracy than other approaches. Our image processing and classifier training software is available at https://github.com/gitter‐lab/t‐cell‐classification .     

Convolutional neural networks improve species distribution modelling by capturing the spatial structure of the environment

Convolutional Neural Networks (CNNs) are statistical models suited for learning complex visual patterns. In the context of Species Distribution Models (SDM) and in line with predictions of landscape ecology and island biogeography, CNN could grasp how local landscape structure affects prediction of species occurrence in SDMs. The prediction can thus reflect the signatures of entangled ecological processes. Although previous machine-learning based SDMs can learn complex influences of environmental predictors, they cannot acknowledge the influence of environmental structure in local landscapes (hence denoted “punctual models”). In this study, we applied CNNs to a large dataset of plant occurrences in France (GBIF), on a large taxonomical scale, to predict ranked relative probability of species (by joint learning) to any geographical position. We examined the way local environmental landscapes improve prediction by performing alternative CNN models deprived of information on landscape heterogeneity and structure (“ablation experiments”). We found that the landscape structure around location crucially contributed to improve predictive performance of CNN-SDMs. CNN models can classify the predicted distributions of many species, as other joint modelling approaches, but they further prove efficient in identifying the influence of local environmental landscapes. CNN can then represent signatures of spatially structured environmental drivers. The prediction gain is noticeable for rare species, which open promising perspectives for biodiversity monitoring and conservation strategies. Therefore, the approach is of both theoretical and practical interest. We discuss the way to test hypotheses on the patterns learnt by CNN, which should be essential for further interpretation of the ecological processes at play.     

Growth curve models of repeated binary response

Experimental designs that include repeated measures of binary response variables over time and under different conditions are common in biology. In such settings, it is often desirable to characterize the response pattern over time. When response variables are continuous, this characterization can be made in terms of a growth model such as the Potthoff-Roy growth curve model. We illustrate how a similar growth curve modeling strategy can be implemented using weighted least squares (WLS) methods for binary response data. The growth models are constructed in terms of polynomial functions across marginal response. However, when growth models are fit to repeated binary response, the nonsignificant higher-order polynomial functions are dropped from the model, rather than used as covariates. Dropping the nonsignificant polynomials from the model will reduce the number of response functions, and help avoid small-sample problems that can occur when the number of correlated response functions is large and sample sizes are small. The reduced set of response functions are then modeled using WLS methods. We illustrate such models with an example of binary fly oviposition response (accept or reject) exhibited by two populations of flies at four ages to two types of fruit.     

Double evaluations of chemicals using a cocktail of fused recombinant receptors

Some chemicals have multipotential as endocrine-disrupting chemicals (EDCs). For example, some chemicals act as both estrogens and antiandrogens. Numbers of such chemicals should be evaluated from many aspects; however, labor and expenses are generally limited. We have developed two expression systems for the wild type of human estrogen receptor alpha and the wild type of human androgen receptor fused with a maltose binding protein. They are soluble and have binding activities. They showed dose-responses to natural hormones and well-known potential EDCs. After we established each assay condition for a competitive binding assay using each receptor, we found that two assay systems can be carried out simultaneously under limited and harmonized conditions. Under harmonized conditions using a cocktail of two types of receptors, we could estimate natural hormones and potential EDCs. Interference between two assay systems was not observed under these conditions. We believe that some competitive binding assays can be carried out using a cocktail of receptors at the same time if interference among different assay systems can be avoided by choosing ideal conditions.     

Chemostat-cultivated Escherichia coli at high dilution rate: multiple steady states and drift

The representation of metabolic network reaction kinetics in a scaled, polynomial form can allow for the prediction of multiple steady states. The polynomial formalism is used to study chemostat-cultured Escherichia coli which has been observed to exhibit two multiple steady states under ammonium ion-limited growth conditions: a high cell density-low ammonium ion concentration steady state and a low cell density-high ammonium ion concentration steady state. Additionally, the low-cell-density steady state has been observed to drift to the high-cell-density steady state. Inspection of the steady-state rate expressions for the ammonium ion transport/assimilation network (in polynomial form) suggests that at low ammonium ion concentrations, two steady states are possible. One corresponds to heavy use of the glutamine synthetase-glutamate synthase (GLNS-GS) branch and the second to heavy use of the glutamate dehydrogenase (GDH) branch. Realization of the predicted intracellular steady states is also found to be dependent on the parameters of the transport process. Moreover, the two steady states differ in where their energy intensity lies. To explain the drift, GLNS, which is inducible under low ammonium ion concentrations, is suggested to be a "memory element." A chemostat-based model is developed to illustrate that perturbations in dilution rate can lead to drift between the two steady states provided that the disturbance in dilution rate is sufficiently large and/or long in duration.     

A support vector machine approach to the identification of phosphorylation sites

We describe a bioinformatics tool that can be used to predict the position of phosphorylation sites in proteins based only on sequence information. The method uses the support vector machine (SVM) statistical learning theory. The statistical models for phosphorylation by various types of kinases are built using a dataset of short (9-amino acid long) sequence fragments. The sequence segments are dissected around post-translationally modified sites of proteins that are on the current release of the Swiss-Prot database, and that were experimentally confirmed to be phosphorylated by any kinase. We represent them as vectors in a multidimensional abstract space of short sequence fragments. The prediction method is as follows. First, a given query protein sequence is dissected into overlapping short segments. All the fragments are then projected into the multidimensional space of sequence fragments via a collection of different representations. Those points are classified with pre-built statistical models (the SVM method with linear, polynomial and radial kernel functions) either as phosphorylated or inactive ones. The resulting list of plausible sites for phosphorylation by various types of kinases in the query protein is returned to the user. The efficiency of the method for each type of phosphorylation is estimated using leave-one-out tests and presented here. The sensitivities of the models can reach over 70%, depending on the type of kinase. The additional information from profile representations of short sequence fragments helps in gaining a higher degree of accuracy in some phosphorylation types. The further development of an automatic phosphorylation site annotation predictor based on our algorithm should yield a significant improvement when using statistical algorithms in order to quantify the results.     

Realization of couplings in a polynomial type mixed-mode cellular neural network

In this paper realization of couplings between cells in a polynomial type mixed-mode cellular neural network (CNN) is analyzed. The choice of the multiplier is discussed and two multiplier types are analyzed. Also, two circuits for generating the second and third order polynomial terms of cell output are described. The accuracy of the multipliers and polynomial circuits at the presence of device mismatch is analyzed.