HOW CHANGES IN THE DATASET AFFECTS THE ACCURACY OF THE BODY CLASSIFIER

Convolutional neural networks have revolutionized computer vision and pattern recognition. They are used to recognize speech, generate various images, process audio signals, process time series, and analyze the meaning of texts. An increasingly complex and deep architecture of neural networks is being developed, along with the undoubted advantages of the common problems of this approach, there are some disadvantages. One of them is the hidden internal principle of the neural network. A properly trained network does not provide researchers with information about identified data dependencies and the structure of the problem. A trained neural network is a set of weight matrices. From this point of view, neural networks are only a tool for solving a specific machine learning problem, but they do not provide experts with analytical information to study the problem. As has long been known, the principle of neural network operation was taken from the principle of neurons in our brain. Inside the brain, we learned to look through ultrasound, PET, MRI and fMRI. And for convolutional neural networks, such indicators as accuracy, precision and heat maps will be used for visualization. The purpose of the work is to find out the effect of a training dataset on the accuracy of a neural network. And how much data will significantly change the stability of the neural network. First of all, they were selected by hyperparameters: learning speed, batch size and number of eras, as well as image size. Then a series of trainings were carried out using the original data, and only then were the images that had been pre-processed added. As the results of the work showed, the dataset, which contains about 15% of the pre-processed data, has a positive effect on the accuracy of the model. When using more data, there was no significant increase in accuracy.

1. Tensor processing unit. Available at: https://cloud.google.com/blog/products/gcp/cloud-tpu-machine-learning-accelerators-now-available-in-beta

2. Christoph Molnar. Interpretable Machine Learning : A Guide for Making Black Box Models Explainable, 2018.

3. ImageNet. Available at: http://www.image-net.org/

4. Labeled Faces in the Wild. Availabe at: http://vis-www.cs.umass.edu/lfw/

5. LabelImg. Available at: https://github.com/tzutalin/labelImg

6. Augmentation sequence. Available at: https://imgaug.readthedocs.io/en/latest/source/examples_basics.html

7. VGGFace2. Available at: http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/

8. Alex Krizhevsky. ImageNet Classification with Deep Convolutional Neural Networks

9. Aharon Azulay and Yair Weiss. Why do deep convolutional networks generalize so poorly to small image transformations?, 2018

10. Dan Hendrycks, Thomas G. Dietterich. Benchmarking Neural Network Robustness to Common Corruptions and Surface Variations, 2019.