Neural Dataset Generality - Supplementary
نویسندگان
چکیده
Figure 1. Samples of some of the datasets that we used in this analysis. From top to bottom: MNIST [1], MNIST-rotated [2], MNIST-random-background [2], MNIST-rotated-background [2], Google street view house numbers [3], Char 74k English [4], Char 74k Kannada [4]. Last two rows, first five from left are CIFAR 10 and the rest are Caltech101 [5, 6]. The bottom row is the colonoscopy dataset. We used one standard network architecture for all character datasets and experiments, one for Cifar 10 vs. Caltech 101 and another standard for Caltech 101 vs. Colonoscopy. The network architectures, learning rates and other details are provided in the sections below. The experiments were conducted on a Macbook Pro Laptop using an Nvidia GT 750M GPU, for character datasets and on an Nvidia Tesla K40 GPU for the others, with cuDNN v3 and Nvidia CUDA v7. Table 1 shows the train-test-validation splits and the batch sizes used in stochastic gradient descent of all the datasets used. No preprocessing were done on the images themselves except for cropping, resizing, normalizing. The images were all normalized to lie in [0, 1]. The character recognition datasets were all of a constant 28X28 grayscale, the Caltech 101 vs. Cifar 10 experiments were performed ar 32X32, RGB and the Caltech 101 vs. Colonsoscopy were at 128X128, RGB.
منابع مشابه
Supplementary Material for: Road Detection using Convolutional Neural Networks
This dataset contains 154 images in an urban environment originally obtained from the KITTI dataset (see [1]). The images show well a demarcated (white lines) two lane highway road. The detection algorithm/method is requried to only consider the lane the recording platform was driving on (i.e the right lane). Apart from this other challenges include, shadows, variations in lane-markings and pre...
متن کاملA multi-step approach to time series analysis and gene expression clustering
MOTIVATION The huge growth in gene expression data calls for the implementation of automatic tools for data processing and interpretation. RESULTS We present a new and comprehensive machine learning data mining framework consisting in a non-linear PCA neural network for feature extraction, and probabilistic principal surfaces combined with an agglomerative approach based on Negentropy aimed a...
متن کاملUsing Neural Network to Control STATCOM for ImprovingTransient Stability
FACTS technology has considerable applications in power systems, such as; improving the steady stateperformance, damping the power system oscillations, controlling the power flow, and etc. STATCOM is oneof the most important FACTS devices used in the parallel compensation, enhancing transient stability andetc. Since three phase fault is widespread in power systems, in this paper STATCOM is used...
متن کاملA New Model for Classifying DNA Code Inspired by Neural Networks and FSA
This paper introduces a new model of classifiers CL(V,E, `, r) designed for classifying DNA sequences and combining the flexibility of neural networks and the generality of finite state automata. Our careful and thorough verification demonstrates that the classifiers CL(V,E, `, r) are general enough and will be capable of solving all classification tasks for any given DNA dataset. We develop a ...
متن کامل