Search results for: neural net works

Authors: Ning Xue, Srinivas Merugu, Ignacio Delgado Martinez, Tao Sun, John Tsang, Shih-Cheng Yen

Abstract:

We have developed a polyimide based neural interface electrode to record nerve signals from the sciatic nerve of a rat. The neural interface electrode has a split-ring shape, with four protruding gold electrodes for recording, and two reference gold electrodes around the split-ring. The split-ring electrode can be opened up to encircle the sciatic nerve. The four electrodes can be bent to sit on top of the nerve and hold the device in position, while the split-ring frame remains flat. In comparison, while traditional cuff electrodes can only fit certain sizes of the nerve, the developed device can fit a variety of rat sciatic nerve dimensions from 0.6 mm to 1.0 mm, and adapt to the chronic changes in the nerve as the electrode tips are bendable. The electrochemical impedance spectroscopy measurement was conducted. The gold electrode impedance is on the order of 10 kΩ, showing excellent charge injection capacity to record neural signals.

Keywords: impedance, neural interface, split-ring electrode, neural signal recording

Procedia PDF Downloads 357

Authors: Tianle Yang

Abstract:

As a common mental disorder, depression is a leading cause of various diseases worldwide. Early detection and treatment of depression can dramatically promote remission and prevent relapse. However, conventional ways of depression diagnosis require considerable human effort and cause economic burden, while still being prone to misdiagnosis. On the other hand, recent studies report that physical characteristics are major contributors to the diagnosis of depression, which inspires us to mine the internal relationship by neural networks instead of relying on clinical experiences. In this paper, neural networks are constructed to predict depression from physical characteristics. Two initialization methods are examined - Xaiver and Kaiming initialization. Experimental results show that a 3-layers neural network with Kaiming initialization achieves 83% accuracy.

Keywords: depression, neural network, Xavier initialization, Kaiming initialization

Procedia PDF Downloads 109

Authors: Milagros Jaén-Vargas, Javier García Martínez, Karla Miriam Reyes Leiva, María Fernanda Trujillo-Guerrero, Francisco Fernandes, Sérgio Barroso Gonçalves, Miguel Tavares Silva, Daniel Simões Lopes, José Javier Serrano Olmedo

Abstract:

Human Activity Recognition (HAR) is having a growing impact in creating new applications and is responsible for emerging new technologies. Also, the use of wearable sensors is an important key to exploring the human body's behavior when performing activities. Hence, the use of these dispositive is less invasive and the person is more comfortable. In this study, a database that includes three activities is used. The activities were acquired from inertial measurement unit sensors (IMU) and motion capture systems (MOCAP). The main objective is differentiating the performance from four Deep Learning (DL) models: Deep Neural Network (DNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and hybrid model Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM), when considering acceleration, velocity and position and evaluate if integrating the IMU acceleration to obtain velocity and position represent an increment in performance when it works as input to the DL models. Moreover, compared with the same type of data provided by the MOCAP system. Despite the acceleration data is cleaned when integrating, results show a minimal increase in accuracy for the integrated signals.

Keywords: HAR, IMU, MOCAP, acceleration, velocity, position, feature maps

Procedia PDF Downloads 80

Authors: Yahia. Kourd, N. Guersi D. Lefebvre

Abstract:

In this paper we propose to study the faults diagnosis in squirrel-cage induction motor using MLP neural networks. We use neural healthy and faulty models of the behavior in order to detect and isolate some faults in machine. In the first part of this work, we have created a neural model for the healthy state using Matlab and a motor located in LGEB by acquirins data inputs and outputs of this engine. Then we detected the faults in the machine by residual generation. These residuals are not sufficient to isolate the existing faults. For this reason, we proposed additive neural networks to represent the faulty behaviors. From the analysis of these residuals and the choice of a threshold we propose a method capable of performing the detection and diagnosis of some faults in asynchronous machines with squirrel cage rotor.

Keywords: faults diagnosis, neural networks, multi-models, squirrel-cage induction motor

Procedia PDF Downloads 614

Authors: Ahmad Alwosheel, Ahmed Alqaraawi

Abstract:

This work proposes a new speaker segmentation approach for two speakers. It is an online approach that does not require a prior information about speaker models. It has two phases, a conventional approach such as unsupervised BIC-based is utilized in the first phase to detect speaker changes and train a Neural Network, while in the second phase, the output trained parameters from the Neural Network are used to predict next incoming audio stream. Using this approach, a comparable accuracy to similar BIC-based approaches is achieved with a significant improvement in terms of computation time.

Keywords: artificial neural network, diarization, speaker indexing, speaker segmentation

Procedia PDF Downloads 481

Authors: Ali Isapour, Ramin Nateghi

Abstract:

— Markov parameters are unique parameters of the system and remain unchanged under similarity transformations. Markov parameters from a power series that is convergent only if the system matrix’s eigenvalues are inside the unity circle. Therefore, Markov parameters of a stable discrete-time system are convergent. In this study, we aim to realize the system based on Markov parameters by using Artificial Neural Networks (ANN), and this end, we use Flexible Neural Networks. Realization means determining the elements of matrices A, B, C, and D.

Keywords: Markov parameters, realization, activation function, flexible neural network

Procedia PDF Downloads 174

Authors: Aisulu Tileukulova, Zhexebay Dauren

Abstract:

Radar plays an important role because it is widely used in civil and military fields. Target detection is one of the most important radar applications. The accuracy of detecting inconspicuous aerial objects in radar facilities is lower against the background of noise. Convolutional neural networks can be used to improve the recognition of this type of aerial object. The purpose of this work is to develop an algorithm for recognizing aerial objects using convolutional neural networks, as well as training a neural network. In this paper, the structure of a convolutional neural network (CNN) consists of different types of layers: 8 convolutional layers and 3 layers of a fully connected perceptron. ReLU is used as an activation function in convolutional layers, while the last layer uses softmax. It is necessary to form a data set for training a neural network in order to detect a target. We built a Confusion Matrix of the CNN model to measure the effectiveness of our model. The results showed that the accuracy when testing the model was 95.7%. Classification of echo signals using CNN shows high accuracy and significantly speeds up the process of predicting the target.

Keywords: radar, neural network, convolutional neural network, echo signals

Procedia PDF Downloads 327

Authors: Siddhant Rao

Abstract:

Studies estimate that there will be 266,120 new cases of invasive breast cancer and 40,920 breast cancer induced deaths in the year of 2018 alone. Despite the pervasiveness of this affliction, the current process to obtain an accurate breast cancer prognosis is tedious and time consuming. It usually requires a trained pathologist to manually examine histopathological images and identify the features that characterize various cancer severity levels. We propose MITOS-RCNN: a region based convolutional neural network (RCNN) geared for small object detection to accurately grade one of the three factors that characterize tumor belligerence described by the Nottingham Grading System: mitotic count. Other computational approaches to mitotic figure counting and detection do not demonstrate ample recall or precision to be clinically viable. Our models outperformed all previous participants in the ICPR 2012 challenge, the AMIDA 2013 challenge and the MITOS-ATYPIA-14 challenge along with recently published works. Our model achieved an F- measure score of 0.955, a 6.11% improvement in accuracy from the most accurate of the previously proposed models.

Keywords: breast cancer, mitotic count, machine learning, convolutional neural networks

Procedia PDF Downloads 202

Authors: Suk-Hyun Yu, Heeyong Kwon

Abstract:

Recently PM-10 has become a social and global issue. It is one of major air pollutants which affect human health. Therefore, it needs to be forecasted rapidly and precisely. However, PM-10 comes from various emission sources, and its level of concentration is largely dependent on meteorological and geographical factors of local and global region, so the forecasting of PM-10 concentration is very difficult. Neural network model can be used in the case. But, there are few cases of high concentration PM-10. It makes the learning of the neural network model difficult. In this paper, we suggest a simple input balancing method when the data distribution is uneven. It is based on the probability of appearance of the data. Experimental results show that the input balancing makes the neural networks’ learning easy and improves the forecasting rates.

Keywords: artificial intelligence, air quality prediction, neural networks, pattern recognition, PM-10

Procedia PDF Downloads 212

Authors: Po-Fang Hsu, Chiching Wei

Abstract:

In this paper, we present a novel neural graph matching approach applied to document comparison. Document comparison is a common task in the legal and financial industries. In some cases, the most important differences may be the addition or omission of words, sentences, clauses, or paragraphs. However, it is a challenging task without recording or tracing the whole edited process. Under many temporal uncertainties, we explore the potentiality of our approach to proximate the accurate comparison to make sure which element blocks have a relation of edition with others. In the beginning, we apply a document layout analysis that combines traditional and modern technics to segment layouts in blocks of various types appropriately. Then we transform this issue into a problem of layout graph matching with textual awareness. Regarding graph matching, it is a long-studied problem with a broad range of applications. However, different from previous works focusing on visual images or structural layout, we also bring textual features into our model for adapting this domain. Specifically, based on the electronic document, we introduce an encoder to deal with the visual presentation decoding from PDF. Additionally, because the modifications can cause the inconsistency of document layout analysis between modified documents and the blocks can be merged and split, Sinkhorn divergence is adopted in our neural graph approach, which tries to overcome both these issues with many-to-many block matching. We demonstrate this on two categories of layouts, as follows., legal agreement and scientific articles, collected from our real-case datasets.

Keywords: document comparison, graph matching, graph neural network, modification similarity, multi-modal

Procedia PDF Downloads 159

Authors: Lamyaa Gamal El-Deen Taha, Ashraf Sharawi

Abstract:

China launched satellite GF-2 in 2014. This study deals with comparing nearest neighbor object-based classification and neural network classification methods for classification of the fused GF-2 image. Firstly, rectification of GF-2 image was performed. Secondly, a comparison between nearest neighbor object-based classification and neural network classification for classification of fused GF-2 was performed. Thirdly, the overall accuracy of classification and kappa index were calculated. Results indicate that nearest neighbor object-based classification is better than neural network classification for urban mapping.

Keywords: GF-2 images, feature extraction-rectification, nearest neighbour object based classification, segmentation algorithms, neural network classification, multilayer perceptron

Procedia PDF Downloads 372

Authors: Salah Ghamari

Abstract:

In this study, the ability of self organizing map artificial (SOM) neural networks in clustering the wheat seeds varieties according to morphological properties of them was considered. The SOM is one type of unsupervised competitive learning. Experimentally, five morphological features of 300 seeds (including three varieties: gaskozhen, Md and sardari) were obtained using image processing technique. The results show that the artificial neural network has a good performance (90.33% accuracy) in classification of the wheat varieties despite of high similarity in them. The highest classification accuracy (100%) was achieved for sardari.

Keywords: artificial neural networks, clustering, self organizing map, wheat variety

Procedia PDF Downloads 626

Authors: Javier Fernandez De Canete, Alvaro Fernandez-Quintero

Abstract:

Nowadays, adaptive control schemes are being used when model based control schemes are applied in presence of uncertainty and model mismatches. Artificial neural networks have been employed both in modelling and control of non-linear dynamic systems with unknown dynamics. In fact, these are powerful tools to solve this control problem when only input-output operational data are available. A neural network controller under SIMULINK together with the ARDUINO hardware platform has been used to perform real-time speed control of a computer case fan. Comparison of performance with a PID controller has also been presented in order to show the efficacy of neural control under different command signals tracking and also when disturbance signals are present in the speed control loops.

Keywords: neural networks, ARDUINO platform, SIMULINK, adaptive speed control

Procedia PDF Downloads 338

Authors: Danladi Ali

Abstract:

In this paper, we measured GSM signal strength in the Dnepropetrovsk city in order to predict path loss in study area using nonlinear autoregressive neural network prediction and we also, used neural network clustering to determine average GSM signal strength receive at the study area. The nonlinear auto-regressive neural network predicted that the GSM signal is attenuated with the mean square error (MSE) of 2.6748dB, this attenuation value is used to modify the COST 231 Hata and the Okumura-Hata models. The neural network clustering revealed that -75dB to -95dB is received more frequently. This means that the signal strength received at the study is mostly weak signal

Keywords: one-dimensional multilevel wavelets, path loss, GSM signal strength, propagation, urban environment and model

Procedia PDF Downloads 368

Authors: Ilker Ali Ozkan

Abstract:

In this study, an artificial neural network model has been developed to estimate chronic kidney failure which is a common disease. The patients’ age, their blood and biochemical values, and 24 input data which consists of various chronic diseases are used for the estimation process. The input data have been subjected to preprocessing because they contain both missing values and nominal values. 147 patient data which was obtained from the preprocessing have been divided into as 70% training and 30% testing data. As a result of the study, artificial neural network model with 25 neurons in the hidden layer has been found as the model with the lowest error value. Chronic kidney failure disease has been able to be estimated accurately at the rate of 99.3% using this artificial neural network model. The developed artificial neural network has been found successful for the estimation of chronic kidney failure disease using clinical data.

Keywords: estimation, artificial neural network, chronic kidney failure disease, disease diagnosis

Procedia PDF Downloads 429

Authors: Aydin Teymourifar, Gurkan Ozturk

Abstract:

In this paper, a new hybrid Hopfield neural network is proposed for the dynamic, flexible job shop scheduling problem. A new heuristic based and easy to implement energy function is designed for the Hopfield neural network, which penalizes the constraints violation and decreases makespan. Moreover, for enhancing the performance, several heuristics are integrated to it that achieve active, and non-delay schedules also, prevent early convergence of the neural network. The suggested algorithm that is designed as a generalization of the previous studies for the flexible and dynamic scheduling problems can be used for solving real scheduling problems. Comparison of the presented hybrid method results with the previous studies results proves its efficiency.

Keywords: dynamic flexible job shop scheduling, neural network, heuristics, constrained optimization

Procedia PDF Downloads 400

Authors: Lu Zhao, Nadir Farhi, Yeltsin Valero, Zoi Christoforou, Nadia Haddadou

Abstract:

In this paper, a long short-term memory (LSTM) neural network model is proposed to replicate simultaneously car-following and lane-changing behaviors in road networks. By combining two kinds of LSTM layers and three input designs of the neural network, six variants of the LSTM model have been created. These models were trained and tested on the NGSIM 101 dataset, and the results were evaluated in terms of longitudinal speed and lateral position, respectively. Then, we compared the LSTM model with a classical car-following model (the intelligent driving model (IDM)) in the part of speed decision. In addition, the LSTM model is compared with a model using classical neural networks. After the comparison, the LSTM model demonstrates higher accuracy than the physical model IDM in terms of car-following behavior and displays better performance with regard to both car-following and lane-changing behavior compared to the classical neural network model.

Keywords: traffic modeling, neural networks, LSTM, car-following, lane-change

Procedia PDF Downloads 230

Authors: R. B. Mishra

Abstract:

Hidden Markov Model (HMM) has been used in prediction and determination of states that generate different neural activations as well as mental working conditions. This paper addresses two applications of HMM; one to determine the optimal sequence of states for two neural states: Active (AC) and Inactive (IA) for the three emission (observations) which are for No Working (NW), Waiting (WT) and Working (W) conditions of human beings. Another is for the determination of optimal sequence of intentionality i.e. Believe (B), Desire (D), and Intention (I) as the states and three observational sequences: NW, WT and W. The computational results are encouraging and useful.

Keywords: hiden markov model, believe desire intention, neural activation, simulation

Procedia PDF Downloads 361

Authors: B. Afsharipoor, E. Nazemi

Abstract:

Artificial neural networks (ANNs) are powerful tool for prediction which can be trained based on a set of examples and thus, it would be useful for nonlinear image processing. The present paper reviews several paper regarding applications of ANN in image processing to shed the light on advantage and disadvantage of ANNs in this field. Different steps in the image processing chain including pre-processing, enhancement, segmentation, object recognition, image understanding and optimization by using ANN are summarized. Furthermore, results on using multi artificial neural networks are presented.

Keywords: neural networks, image processing, segmentation, object recognition, image understanding, optimization, MANN

Procedia PDF Downloads 375

Authors: Li Kewen, Su Zhaoxin, Wang Xingmou, Zhu Jian Bing

Abstract:

Conventional reservoir prediction methods ar not sufficient to explore the implicit relation between seismic attributes, and thus data utilization is low. In order to improve the predictive classification accuracy of reservoir lithology, this paper proposes a deep learning lithology prediction method based on ResNet (Residual Neural Network) and SENet (Squeeze-and-Excitation Neural Network). The neural network model is built and trained by using seismic attribute data and lithology data of Shengli oilfield, and the nonlinear mapping relationship between seismic attribute and lithology marker is established. The experimental results show that this method can significantly improve the classification effect of reservoir lithology, and the classification accuracy is close to 70%. This study can effectively predict the lithology of undrilled area and provide support for exploration and development.

Keywords: convolutional neural network, lithology, prediction of reservoir, seismic attributes

Procedia PDF Downloads 159

Authors: Magdi. M. Nabi, Ding-Li Yu

Abstract:

Control of a semi-batch polymerization reactor using an adaptive radial basis function (RBF) neural network method is investigated in this paper. A neural network inverse model is used to estimate the valve position of the reactor; this method can identify the controlled system with the RBF neural network identifier. The weights of the adaptive PID controller are timely adjusted based on the identification of the plant and self-learning capability of RBFNN. A PID controller is used in the feedback control to regulate the actual temperature by compensating the neural network inverse model output. Simulation results show that the proposed control has strong adaptability, robustness and satisfactory control performance and the nonlinear system is achieved.

Keywords: Chylla-Haase polymerization reactor, RBF neural networks, feed-forward, feedback control

Procedia PDF Downloads 682

Authors: Faezeh Aghazadeh, Mohammad Javad Sharifi

Abstract:

The application of neural networks to model a full-scale industrial acetylene hydrogenation in olefin plant has been studied. The operating variables studied are the, input-temperature of the reactor, output-temperature of the reactor, hydrogen ratio of the reactor, [C₂H₂]input, and [C₂H₆]input. The studied operating variables were used as the input to the constructed neural network to predict the [C₂H₆]output at any time as the output or the target. The constructed neural network was found to be highly precise in predicting the quantity of [C₂H₆]output for the new input data, which are kept unaware of the trained neural network showing its applicability to determine the [C₂H₆]output for any operating conditions. The enhancement of [C₂H₆]output as compared with [C₂H₆]input was a consequence of low selective acetylene hydrogenation to ethylene.

Keywords: acetylene hydrogenation, Pd-Ag/Al₂O₃, artificial neural network, modeling, optimal design

Procedia PDF Downloads 252

Authors: Mohammad Saleh Vahdatpour, Yanqing Zhang

Abstract:

Understanding the neural mechanisms underlying motion detection in the human visual system has long been a fascinating challenge in neuroscience and artificial intelligence. This paper presents a spiking neural network model inspired by the processing of motion information in the primate visual system, particularly focusing on the Middle Temporal (MT) area. In our study, we propose a multi-layer spiking neural network model to perform motion detection tasks, leveraging the idea that synaptic delays in neuronal communication are pivotal in motion perception. Synaptic delay, determined by factors like axon length and myelin insulation, affects the temporal order of input spikes, thereby encoding motion direction and speed. Overall, our spiking neural network model demonstrates the feasibility of capturing motion detection principles observed in the primate visual system. The combination of synaptic delays, learning mechanisms, and shared weights and delays in SMD provides a promising framework for motion perception in artificial systems, with potential applications in computer vision and robotics.

Keywords: neural network, motion detection, signature detection, convolutional neural network

Procedia PDF Downloads 63

Authors: Yoke-Mui Lim, Yahaya Ahmad

Abstract:

Conservation works in Malaysia that is procured by public organisation usually follow the traditional approach where the works are tendered based on Bills of Quantities (BQ). One of the purposes of tendering is to enable the selection of a competent contractor that offers a competitive price. While competency of the contractors are assessed by their technical knowledge, experience and track records, the assessment of pricing will be dependent on the tender amount. However, the issue currently faced by the conservation works sector is the difficulty in assessing the competitiveness and reasonableness of the tender amount due to the high variance between the tenders amount. Thus, this paper discusses the factors that cause difficulty to the tenderers in pricing competitively in a bidding exercise for conservation tenders. Data on tendering is collected from interviews with conservation works contractors to gain in-depth understanding of the barriers faced in pricing tenders of conservation works. Findings from the study lent support to the contention that the variance of tender amount is very high amongst tenderers. The factors identified in the survey are the format of BQ, hidden works, experience and labour and material costs.

Keywords: building conservation, Malaysia, bill of quantities, tender

Procedia PDF Downloads 361

Authors: Si Mon Kueh, Tom J. Kazmierski

Abstract:

There are about 1% of the world population suffering from the hidden disability known as epilepsy and major developing countries are not fully equipped to counter this problem. In order to reduce the inconvenience and danger of epilepsy, different methods have been researched by using a artificial neural network (ANN) classification to distinguish epileptic waveforms from normal brain waveforms. This paper outlines the aim of achieving massive ANN parallelization through a dedicated hardware using bit-serial processing. The design of this bit-serial Neural Processing Element (NPE) is presented which implements the functionality of a complete neuron using variable accuracy. The proposed design has been tested taking into consideration non-idealities of a hardware ANN. The NPE consists of a bit-serial multiplier which uses only 16 logic elements on an Altera Cyclone IV FPGA and a bit-serial ALU as well as a look-up table. Arrays of NPEs can be driven by a single controller which executes the neural processing algorithm. In conclusion, the proposed compact NPE design allows the construction of complex hardware ANNs that can be implemented in a portable equipment that suits the needs of a single epileptic patient in his or her daily activities to predict the occurrences of impending tonic conic seizures.

Keywords: Artificial Neural Networks (ANN), bit-serial neural processor, FPGA, Neural Processing Element (NPE)

Procedia PDF Downloads 300

Authors: J. K. Alhassan, C. S. Actsu

Abstract:

This study is on the prediction of feature performance of undergraduate students with Artificial Neural Networks (ANN). With the growing decline in the quality academic performance of undergraduate students, it has become essential to predict the students’ feature academic performance early in their courses of first and second years and to take the necessary precautions using such prediction-based information. The feed forward multilayer neural network model was used to train and develop a network and the test carried out with some of the input variables. A result of 80% accuracy was obtained from the test which was carried out, with an average error of 0.009781.

Keywords: academic performance, artificial neural network, prediction, students

Procedia PDF Downloads 441

Authors: B. S.Akhmetov, S. T. Akhmetova, D. N. Nadeyev, V. Yu. Yegorov, V. V. Smogoonov

Abstract:

Estimated probabilities of errors of the first and second kind for nonideal biometrics-neural transducers 256 outputs, the construction of nomograms based error probability of 'own' and 'alien' from the mathematical expectation and standard deviation of the normalized measures Hamming.

Keywords: modeling, errors, probability, biometrics, neural network, authentication

Procedia PDF Downloads 472

Authors: Liu Zhiyuan, Sun Zongdi

Abstract:

In this paper, the BP neural network model is established to predict the carbon trading price and carbon trading volume in Shanghai City. First of all, we find the data of carbon trading price and carbon trading volume in Shanghai City from September 30, 2015 to December 23, 2016. The carbon trading price and trading volume data were processed to get the average value of each 5, 10, 20, 30, and 60 carbon trading price and trading volume. Then, these data are used as input of BP neural network model. Finally, after the training of BP neural network, the prediction values of Shanghai carbon trading price and trading volume are obtained, and the model is tested.

Keywords: Carbon trading price, carbon trading volume, BP neural network model, Shanghai City

Procedia PDF Downloads 333

Authors: Baeza S. Roberto

Abstract:

The knitted fabric suffers a deformation in its dimensions due to stretching and tension factors, transverse and longitudinal respectively, during the process in rectilinear knitting machines so it performs a dry relaxation shrinkage procedure and thermal action of prefixed to obtain stable conditions in the knitting. This paper presents a dry relaxation shrinkage prediction of Bordeaux fiber using a feed forward neural network and linear regression models. Six operational alternatives of shrinkage were predicted. A comparison of the results was performed finding neural network models with higher levels of explanation of the variability and prediction. The presence of different reposes are included. The models were obtained through a neural toolbox of Matlab and Minitab software with real data in a knitting company of Southern Guanajuato. The results allow predicting dry relaxation shrinkage of each alternative operation.

Keywords: neural network, dry relaxation, knitting, linear regression

Procedia PDF Downloads 562

Authors: Alexander Goncharenko, Andrey Denisov, Sergey Alyamkin, Evgeny Terentev

Abstract:

The neural network quantization is highly desired procedure to perform before running neural networks on mobile devices. Quantization without fine-tuning leads to accuracy drop of the model, whereas commonly used training with quantization is done on the full set of the labeled data and therefore is both time- and resource-consuming. Real life applications require simplification and acceleration of quantization procedure that will maintain accuracy of full-precision neural network, especially for modern mobile neural network architectures like Mobilenet-v1, MobileNet-v2 and MNAS. Here we present a method to significantly optimize training with quantization procedure by introducing the trained scale factors for discretization thresholds that are separate for each filter. Using the proposed technique, we quantize the modern mobile architectures of neural networks with the set of train data of only ∼ 10% of the total ImageNet 2012 sample. Such reduction of train dataset size and small number of trainable parameters allow to fine-tune the network for several hours while maintaining the high accuracy of quantized model (accuracy drop was less than 0.5%). Ready-for-use models and code are available in the GitHub repository.

Keywords: distillation, machine learning, neural networks, quantization

Procedia PDF Downloads 300