Image Region Labelling by Humans and by an Artificial Neural Network

Perception ◽  
1996 ◽  
Vol 25 (1_suppl) ◽  
pp. 27-27
Author(s):  
A Clark ◽  
T Troscianko ◽  
N Campbell ◽  
B Thomas
Keyword(s):  
Neural Network ◽  
Learning Curve ◽  
Human Subjects ◽  
Outer Boundary ◽  
Natural Scenes ◽  
Neural Net ◽  
Image Region ◽  
Average Luminance ◽  
Partial Representation ◽  
Posteriori Analysis

We reported (Troscianko et al, 1995 Perception24 Supplement, 18) that a neural network has been developed which is capable of labelling objects in natural scenes by first segmenting a scene, then obtaining a description of each segment in terms of a set of features. A neural net is then trained to label the segments on the basis of the feature set. The question we are now addressing is: how important is each of these features to overall performance, both in human and machine vision? We carried out an experiment in which human subjects were trained in the same labelling task as the neural net. Individual segments of scenes (sometimes corresponding to a whole object, eg a car, and sometimes an incomplete region, eg part of the sky) were presented on a screen, and the subject asked to label the scene as one of eleven possible types of object (sky, vegetation, vehicle …). Feedback was given and the learning curve monitored. When the learning curve was flat, each subject's performance was investigated with both intact and degraded stimuli. The degradation consisted of partial representation of the information, such as presenting just the outer boundary, or the average colour, or the average luminance, or randomising the size, position, and texture of the segment. The results suggest that this degradation produces significant changes in performance (F9,7=4.4, p=0.0005). A posteriori analysis indicates that certain attributes (particularly texture, boundary-only, colour-averaging) are particularly influential in mediating performance. A similar set of results was obtained by training the network on similarly degraded data. The results imply: (1) that a neural net can provide a useful model of human object labelling processes, and (2) that certain features are more important than others in mediating such performance.

Using artificial neural network condensation to facilitate adaption of machine learning in medical settings by reducing computational burden (Preprint)

2020 ◽  
Author(s):  
Dianbo Liu
Keyword(s):  
Neural Network ◽  
Machine Learning ◽  
Third World ◽  
Mortality Prediction ◽  
Neural Net ◽  
Medical Settings ◽  
Hidden Layer ◽  
Applications Of Machine Learning ◽  
Computational Resources ◽  
Developed Nations

BACKGROUND Applications of machine learning (ML) on health care can have a great impact on people’s lives. At the same time, medical data is usually big, requiring a significant amount of computational resources. Although it might not be a problem for wide-adoption of ML tools in developed nations, availability of computational resource can very well be limited in third-world nations and on mobile devices. This can prevent many people from benefiting of the advancement in ML applications for healthcare. OBJECTIVE In this paper we explored three methods to increase computational efficiency of either recurrent neural net-work(RNN) or feedforward (deep) neural network (DNN) while not compromising its accuracy. We used in-patient mortality prediction as our case analysis upon intensive care dataset. METHODS We reduced the size of RNN and DNN by applying pruning of “unused” neurons. Additionally, we modified the RNN structure by adding a hidden-layer to the RNN cell but reduce the total number of recurrent layers to accomplish a reduction of total parameters in the network. Finally, we implemented quantization on DNN—forcing the weights to be 8-bits instead of 32-bits. RESULTS We found that all methods increased implementation efficiency–including training speed, memory size and inference speed–without reducing the accuracy of mortality prediction. CONCLUSIONS This improvements allow the implementation of sophisticated NN algorithms on devices with lower computational resources.

scholarly journals Efficient Embedded Decoding of Neural Network Language Models in a Machine Translation System

2018 ◽  
Vol 28 (09) ◽  
pp. 1850007
Author(s):  
Francisco Zamora-Martinez ◽  
Maria Jose Castro-Bleda
Keyword(s):  
Neural Network ◽  
Machine Translation ◽  
Language Processing ◽  
Traditional Approach ◽  
Computational Cost ◽  
Integrated Approach ◽  
Language Models ◽  
Translation System ◽  
Neural Net ◽  
Network Language

Neural Network Language Models (NNLMs) are a successful approach to Natural Language Processing tasks, such as Machine Translation. We introduce in this work a Statistical Machine Translation (SMT) system which fully integrates NNLMs in the decoding stage, breaking the traditional approach based on [Formula: see text]-best list rescoring. The neural net models (both language models (LMs) and translation models) are fully coupled in the decoding stage, allowing to more strongly influence the translation quality. Computational issues were solved by using a novel idea based on memorization and smoothing of the softmax constants to avoid their computation, which introduces a trade-off between LM quality and computational cost. These ideas were studied in a machine translation task with different combinations of neural networks used both as translation models and as target LMs, comparing phrase-based and [Formula: see text]-gram-based systems, showing that the integrated approach seems more promising for [Formula: see text]-gram-based systems, even with nonfull-quality NNLMs.

scholarly journals An artificial neural network identifies glyphosate-impacted brackish communities based on 16S rRNA amplicon MiSeq read counts

2019 ◽  
Author(s):  
René Janßen ◽  
Jakob Zabel ◽  
Uwe von Lukas ◽  
Matthias Labrenz
Keyword(s):  
Neural Network ◽  
Neural Networks ◽  
Artificial Neural Network ◽  
Artificial Neural Networks ◽  
Environmental Monitoring ◽  
Microbial Communities ◽  
List Type ◽  
Neural Net ◽  
Next Generation ◽  
Artificial Neural

AbstractArtificial neural networks can be trained on complex data sets to detect, predict, or model specific aspects. Aim of this study was to train an artificial neural network to support environmental monitoring efforts in case of a contamination event by detecting induced changes towards the microbial communities. The neural net was trained on taxonomic cluster count tables obtained via next-generation amplicon sequencing of water column samples originating from a lab microcosm incubation experiment conducted over 140 days to determine the effects of the herbicide glyphosate on succession within brackish-water microbial communities. Glyphosate-treated assemblages were classified correctly; a subsetting approach identified the clusters primarily responsible for this, permitting the reduction of input features. This study demonstrates the potential of artificial neural networks to predict indicator species in cases of glyphosate contamination. The results could empower the development of environmental monitoring strategies with applications limited to neither glyphosate nor amplicon sequence data.Highlight bullet pointsAn artificial neural net was able to identify glyphosate-affected microbial community assemblages based on next generation sequencing dataDecision-relevant taxonomic clusters can be identified by a stochastically subsetting approachJust a fraction of present clusters is needed for classificationFiltering of input data improves classification

scholarly journals Part-of-Speech Classification from Magnetoencephalography Data Using 1-Dimensional Convolutional Neural Network

2020 ◽  
Author(s):  
Alessandro Lopopolo ◽  
Antal van den Bosch
Keyword(s):  
Neural Network ◽  
Convolutional Neural Network ◽  
Mental State ◽  
Human Subjects ◽  
Brain Activity ◽  
Neural Decoding ◽  
Significant Progress ◽  
Part Of Speech ◽  
Syntactic Information ◽  
Speech Classification

Neural decoding of speech and language refers to the extraction of information regarding the stimulus and the mental state of subjects from recordings of their brain activity while performing linguistic tasks. Recent years have seen significant progress in the decoding of speech from cortical activity. This study instead focuses on decoding linguistic information. We present a deep parallel temporal convolutional neural network (1DCNN) trained on part-of-speech (PoS) classification from magnetoencephalography (MEG) data collected during natural language reading. The network is trained on data from 15 human subjects separately, and yields above-chance accuracies on test data for all of them. The level of PoS was targeted because it offers a clean linguistic benchmark level that represents syntactic information and abstracts away from semantic or conceptual representations.

scholarly journals Story Realization: Expanding Plot Events into Sentences

2020 ◽  
Vol 34 (05) ◽  
pp. 7375-7382
Author(s):  
Prithviraj Ammanabrolu ◽  
Ethan Tien ◽  
Wesley Cheung ◽  
Zhaochen Luo ◽  
William Ma ◽  
...  
Keyword(s):  
Neural Network ◽  
Natural Language ◽  
Human Subjects ◽  
Prior Work ◽  
Generation System ◽  
Language Generation ◽  
Story Generation ◽  
Sequence Of Events ◽  
End To End

Neural network based approaches to automated story plot generation attempt to learn how to generate novel plots from a corpus of natural language plot summaries. Prior work has shown that a semantic abstraction of sentences called events improves neural plot generation and and allows one to decompose the problem into: (1) the generation of a sequence of events (event-to-event) and (2) the transformation of these events into natural language sentences (event-to-sentence). However, typical neural language generation approaches to event-to-sentence can ignore the event details and produce grammatically-correct but semantically-unrelated sentences. We present an ensemble-based model that generates natural language guided by events. We provide results—including a human subjects study—for a full end-to-end automated story generation system showing that our method generates more coherent and plausible stories than baseline approaches 1.

A Hybrid Neural Network Approach for the Development of Friction Component Dynamic Model

2004 ◽  
Vol 126 (1) ◽  
pp. 144-153 ◽  
Author(s):  
M. Cao ◽  
K. W. Wang ◽  
Y. Fujii ◽  
W. E. Tobler
Keyword(s):  
Neural Network ◽  
First Principles ◽  
Physical Parameters ◽  
Neural Net ◽  
Asperity Contact ◽  
Neural Network Approach ◽  
Hybrid Neural Network ◽  
Friction Component ◽  
Improved Performance ◽  
Viscous Torque

In this research, a new hybrid neural network is developed to model engagement behaviors of automotive transmission wet friction component. Utilizing known first principles on the physics of engagement, special modules are created to estimate viscous torque and asperity contact torque as preprocessors to a two-layer neural network. Inside these modules, all the physical parameters are represented by neurons with various activation functions derived from first principles. These new features contribute to the improved performance and trainability over a conventional two-layer network model. Both the hybrid and conventional neural net models are trained and tested with experimental data collected from an SAE#2 test stand. The results show that the performance of the hybrid model is much superior to that of the conventional model. It successfully captures detailed characteristics of the friction component engagement torque as a function of time over a wide operating range.

A Recurrent Neural Network Approach for Automated Neural Tracing in Multispectral 3D Images

2018 ◽  
Author(s):  
Yan Yan ◽  
Douglas H. Roossien ◽  
Benjamin V. Sadis ◽  
Jason J. Corso ◽  
Dawen Cai
Keyword(s):  
Neural Network ◽  
Deep Learning ◽  
Neuronal Cell ◽  
Ground Truth ◽  
Neuronal Morphology ◽  
Spectral Information ◽  
Neural Net ◽  
Learning Approaches ◽  
3D Images ◽  
Neural Tracing

AbstractNeuronal morphology reconstruction in fluorescence microscopy 3D images is essential for analyzing neuronal cell type and connectivity. Manual tracing of neurons in these images is time consuming and subjective. Automated tracing is highly desired yet is one of the foremost challenges in computational neuroscience. The multispectral labeling technique, Brainbow utilizes high dimensional spectral information to distinguish intermingled neuronal processes. It is particular interesting to develop new algorithms to include the spectral information into the tracing process. Recently, deep learning approaches achieved state-of-the-art in different computer vision and medical imaging applications. To benefit from the power of deep learning, in this paper, we propose an automated neural tracing approach in multispectral 3D Brainbow images based on recurrent neural net-work. We first adopt VBM4D approach to denoise multispectral 3D images. Then we generate cubes as training samples along the ground truth, manually traced paths. These cubes are the input to the recur-rent neural network. The proposed approach is simple and effective. The approach can be implemented with the deep learning toolbox ‘Keras’ in 100 lines. Finally, to evaluate our approach, we computed the average and standard deviation of DIADEM metric from the ground truth results to our tracing results, and from our tracing results to the ground truth results. Extensive experimental results on the collected dataset demonstrate that the proposed approach performs well in Brainbow labeled mouse brain images.

Continuous Learning Framework for Freeway Incident Detection

1998 ◽  
Vol 1644 (1) ◽  
pp. 124-131 ◽  
Author(s):  
Srinivas Peeta ◽  
Debjit Das
Keyword(s):  
Neural Network ◽  
Back Propagation ◽  
Superior Performance ◽  
Incident Detection ◽  
Neural Net ◽  
Continuous Learning ◽  
Back Propagation Algorithm ◽  
Learning Capability ◽  
Detection Algorithms ◽  
On Line

Existing freeway incident detection algorithms predominantly require extensive off-line training and calibration precluding transferability to new sites. Also, they are insensitive to demand and supply changes in the current site without recalibration. We propose two neural network-based approaches that incorporate an on-line learning capability, thereby ensuring transferability, and adaptability to changes at the current site. The least-squares technique and the error back propagation algorithm are used to develop on-line neural network-trained versions of the popular California algorithm and the more recent McMaster algorithm. Simulated data from the integrated traffic simulation model is used to analyze performance of the neural network-based versions of the California and McMaster algorithms over a broad spectrum of operational scenarios. The results illustrate the superior performance of the neural net implementations in terms of detection rate, false alarm rate, and time to detection. Of implications to current practice, they suggest that just introducing a continuous learning capability to commonly used detection algorithms in practice such as the California algorithm enhances their performance with time in service, allows transferability, and ensures adaptability to changes at the current site. An added advantage of this strategy is that existing traffic measures used (such as volume, occupancy, and so forth.) in those algorithms are sufficient, circumventing the need for new traffic measures, new threshold parameters, and variables that require subjective decisions.

Neural network studies 2. Use of a neural net to estimate oxidation energies for substituted dihydropyridines and related heterocycles

Tetrahedron ◽  
1992 ◽  
Vol 48 (17) ◽  
pp. 3463-3472 ◽  
Author(s):  
Marcus E. Brewster ◽  
Ming-Ju Huang ◽  
Alan Harget ◽  
Nicholas Bodor
Keyword(s):  
Neural Network ◽  
Neural Net ◽  
Network Studies
Sign in / Sign up

Export Citation Format

Share Document