Genetic CNN

Abstract

The deep Convolutional Neural Network (CNN) is the state-of-the-art solution for large-scale visual recognition. Following basic principles such as increasing the depth and constructing highway connections, researchers have manually designed a lot of fixed network structures and verified their effectiveness.In this paper, we discuss the possibility of learning deep network structures automatically. Note that the number of possible network structures increases exponentially with the number of layers in the network, which inspires us to adopt the genetic algorithm to efficiently traverse this large search space. We first propose an encoding method to represent each network structure in a fixed-length binary string, and initialize the genetic algorithm by generating a set of randomized individuals. In each generation, we define standard genetic operations, e.g., selection, mutation and crossover, to eliminate weak individuals and then generate more competitive ones. The competitiveness of each individual is defined as its recognition accuracy, which is obtained via training the network from scratch and evaluating it on a validation set. We run the genetic process on two small datasets, i.e., MNIST and CIFAR10, demonstrating its ability to evolve and find high-quality structures which are little studied before. These structures are also transferrable to the large-scale ILSVRC2012 dataset.

Acquisition design for direct reflectivity and velocity estimation from blended and irregularly sampled data

“…In this respect, a typical approach is to explore and design efficient network architectures using GAs (Ahmadizar et al . ; Xie and Yuille ). The other way around, Kramer () outlined ways to support and accelerate evolutionary algorithms with the learning process.…”

Enhancing Evolutionary Couplings with Deep Convolutional Neural Networks

“…The generality of the chosen search space has a major influence on the quality of results that are even possible. Most existing evolutionary NAS approaches [16], [21], [22], [32] search only one aspect of the architecture space-e.g., the connections and/or hyper-parameters. In contrast, NSGANet searches over both operations and connections-the search space is thus more comprehensive, including most of the previous successful architectures designed both by human experts and algorithmically.…”

“…An EA, in general, is an iterative process in which individuals in a population are made gradually better by applying variations to selected individuals and/or recombining parts of multiple individuals. Despite the ease of extending them to handle multiple objectives, most existing EA-based NAS methods [16]- [21] are still single-criteria driven. Even under the explosive growth of general interest in NAS, EA-based NAS approaches have not been well perceived outside the EA community, primarily due to the following two reasons: (i) existing EAbased methods [18], [19] that produce competitive results are extremely computationally inefficient (e.g., one run of [19] takes 7 days on 450 GPUs); or (ii) results from existing EAbased methods [16], [20]- [22] that use limited search budgets are far from state-of-the-art performance and only demonstrated on small-scale datasets.…”

Detection of Optical Characters Using Text Structure Modeling

“…Each of the textual content form trouble detectors is designed to extract the excellent skills of remarkable forms of Chinese language language person shape additives and S. M. Lucas, et al [8], explains to benefit a smooth image of the nation of work of analyzing text in scenes furthermore of the consequences using favored techniques of vicinity precision/recall into account in thoughts and edit distance given with the aid of a. Shahab, F. Shafait, and a. Dengel [9]. Lingxi Xie, Alan Yuille [11], MDD are done with the beneficial useful resource of comparing the diagnosed transcriptions with the canonical ones. Regardless of the reality that this technique performs reason-ably nicely, it's the following problems: (1) studying the error patterns of the target phrases to generate the ERNs stays a hard work.…”

“…The important thing disadvantage of the lower back-propagation again is that it does now not guarantee that the neural community will converge to a worldwide minimal. several strategies are furnished inside the literature that may be used over the as soon as returned-propagation to in show the convergence of the neural network, like gradient-based studying getting to know and splendid reading techniques, utilizing distortions to the pattern from running out set, determine even as the learning price must without a doubt be modified [11] [16]. The text categorization hassle is probably defied as following: several documents and classes are made and we should classify those files to a number of the ones education in step with their contents.…”