-
Optimizing the Hyper-parameters of Multi-layer Perceptron with Greedy Search
Issue:
Volume 4, Issue 4, December 2021
Pages:
90-96
Received:
13 September 2021
Accepted:
4 October 2021
Published:
15 October 2021
Abstract: The core of deep learning network is hyper-parameters which are updated through learning process with samples. Whenever a sample is fed into deep learning network, parameters change according to gradient value. At this point, the number of samples and the amount of learning are crucial, which are batch size and learning rate. To find the optimal batch size and learning rate, lots of trial is inevitable so it takes so much time and effort. Therefore, there have been lots of papers to enhance the efficiency of its optimization process by automatically tuning the single parameter. However, global optimization can’t be guaranteed by simply combining separately optimized parameters. This paper propose brand new effective method for hyperparameter optimization in which greedy search is adopted to find the optimal batch size and learning rate. In experiment with Fashion MNIST and Kuzushiji MNIST dataset, the proposed algorithm shows the similar performance as compared to complete search, which means the proposed algorithm can be a potential alternative to complete search.
Abstract: The core of deep learning network is hyper-parameters which are updated through learning process with samples. Whenever a sample is fed into deep learning network, parameters change according to gradient value. At this point, the number of samples and the amount of learning are crucial, which are batch size and learning rate. To find the optimal ba...
Show More
-
Scaling up an Unsupervised Image-to-Image Translation Framework from Basic to Complex Scenes
Daniel Filipe Nunes Silva,
Samuel Chassot,
Luca Barras,
Deblina Bhattacharjee,
Sabine Süsstrunk
Issue:
Volume 4, Issue 4, December 2021
Pages:
97-105
Received:
27 July 2021
Accepted:
31 August 2021
Published:
29 October 2021
Abstract: Unsupervised image-to-image translation methods have received a lot of attention in the last few years. Multiple techniques emerged to tackle the initial challenge from different perspectives. Some focus on learning as much as possible from the target-style using several images of that style for each translation while others make use of object detection in order to produce more realistic results on content-rich scenes. In this paper, we explore multiple frameworks that rely on different paradigms and assess how one of these that has initially been developed for single object translation performs on more diverse and content-rich images. Our work is based on an already existing framework. We explore its versatility by training it with a more diverse dataset than the one it was designed and tuned for. This helps understanding how such methods behave beyond their original application. We explore how to make the most out of the datasets despite our computational power limitations. We present a way to extend a dataset by passing it through an object detector. The latter provides us with new and diverse dataset classes. Moreover, we propose a way to adapt the framework in order to leverage the power of object detection by integrating it in the architecture as one can see in other methods.
Abstract: Unsupervised image-to-image translation methods have received a lot of attention in the last few years. Multiple techniques emerged to tackle the initial challenge from different perspectives. Some focus on learning as much as possible from the target-style using several images of that style for each translation while others make use of object dete...
Show More
-
Research and Application of Rectified-NAdam Optimization Algorithm in Data Classification
Issue:
Volume 4, Issue 4, December 2021
Pages:
106-110
Received:
27 September 2021
Accepted:
25 October 2021
Published:
5 November 2021
Abstract: Data classification exists in various practical applications, such as the classification of words in natural language processing, classification of meteorological conditions, classification of environmental pollution degree, and so on. Artificial neural network is a basic method of data classification. A reasonable optimization algorithm will get better results for a loss function in the neural network. The research and improvement of these optimization algorithms has been a focus in this field. Because of the various optimizers developing in building the neural networks, an improved NAdam Algorithm (RNAdam) is proposed in this paper, on the basis of discussing and comparing several Algorithms with Adam Algorithm. This algorithm not only combines the advantages of RAdam algorithm, but also keeps the convergence of NAdam algorithm. A classification experiment is carried out on the data set composed of 300 sample points generated by the Make moon function. The experimental results show that the RNAdam algorithm is better than SGDM, Adam and Nadam algorithm in terms of the loss and accuracy between the output and the actual results, when the data are classified by the three-layer neural network. Therefore, the classification effect will be improved when this algorithm is applied to neural network for various practical data classification problems.
Abstract: Data classification exists in various practical applications, such as the classification of words in natural language processing, classification of meteorological conditions, classification of environmental pollution degree, and so on. Artificial neural network is a basic method of data classification. A reasonable optimization algorithm will get b...
Show More
-
Pedestrian Tracking Algorithm Combining Contextual Information and Attention Mechanism
Shunliang Xiao,
Zanxia Qiang,
Weiguang Liu,
Xianfu Bao
Issue:
Volume 4, Issue 4, December 2021
Pages:
111-118
Received:
8 October 2021
Accepted:
25 October 2021
Published:
5 November 2021
Abstract: In the real scene, because pedestrians are occluded or the size of pedestrians is small, the convolutional neural network cannot fully extract their features, resulting in poor detection results. In two adjacent frames, the same pedestrian is prone to errors when doing data association, which makes the pedestrian tracking effect unsatisfactory. In order to solve this problem, the pedestrian tracking algorithm based on Anchor-free idea is improved. A fusion context information module is proposed to enhance the model's feature extraction ability for different receptive fields, and improve the model's detection and tracking performance when the pedestrian size is small. In addition, in order to let the model learn to pay attention to the effective information of the feature layer. A coordinated attention mechanism is introduced to guide the model to learn the weights of different channels and different regions of the feature layer, and to improve the tracking performance of the model when pedestrians are occluded. In the experiment, the tracking performance of the model was verified on the MOT16 dataset. Experimental results show that compared with other main popular person tracking algorithms, the improved algorithm has higher tracking accuracy and lower pedestrian ID switching times. Its tracking accuracy is 70.74.
Abstract: In the real scene, because pedestrians are occluded or the size of pedestrians is small, the convolutional neural network cannot fully extract their features, resulting in poor detection results. In two adjacent frames, the same pedestrian is prone to errors when doing data association, which makes the pedestrian tracking effect unsatisfactory. In ...
Show More
-
Applications Based on a Novel Sudoku Solver Algorithm and Grid Based Models
Abhishake Kundu,
Anand Sunder
Issue:
Volume 4, Issue 4, December 2021
Pages:
119-128
Received:
18 August 2021
Accepted:
24 November 2021
Published:
2 December 2021
Abstract: Numerous algorithms for solving sudoku puzzles have been explored, most of which use a backtracking approach. Thus computational efficiency of such algorithms can sometimes yield poor results. We propose a probabilistic solver algorithm which, iteratively fills the sudoku grid and solves the same. In this approach we make use of a dynamic random number set, we identify unassigned sudoku grids for a given puzzle where only one possible value can be filled in and iteratively identify and assign cells with least number of possible values. We not only elaborate on our solver algorithm logic, but also explore application areas based on algorithm devised, after reviewing relevant similar approaches illustrated in the referenced articles. We believe by extension of this algorithm, many combinatorial problems in the field of material characterization, cryptography, cybersecurity can be solved and advanced. We also envision that with application of neural networks, Machine Learning techniques the algorithm will take a very adaptive and robust form, useful for solving complex problems in accurate estimation of missing data, discrete event analysis and prediction. Uniqueness is the ability to use high probability for faster computation and low execution time. With cyberattacks of varied vectors and types, its important to devise a mechanism to create a deliberate mismatch every time a possible attack is detected.
Abstract: Numerous algorithms for solving sudoku puzzles have been explored, most of which use a backtracking approach. Thus computational efficiency of such algorithms can sometimes yield poor results. We propose a probabilistic solver algorithm which, iteratively fills the sudoku grid and solves the same. In this approach we make use of a dynamic random nu...
Show More
