Image Classification Using Generative Neuro Evolution for Deep Learning

Verbancsics, Phillip; Harguess, Josh

doi:10.1109/wacv.2015.71

Cited by 47 publications

(25 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Evolutionary Algorithms (EA) are advantageous for architecture search, as any function (not necessarily differentiable) can be optimized using these methods. HyperNEAT was the first EA successfully applied [234] to deep learning, used for training weights and DNN architecture at the same time; and CoDeepNEAT [169] defines a variant of the NEAT algorithm to optimize hyperparameters and architecture, using the self-similarity feature of DNNs by optimizing "blueprints" that are composed of modules. Genetic CNNs [242] uses Genetic Algorithms (GAs) by encoding the DNN connections as binary genes (as required in GAs, shown in Fig.…”

Section: 52mentioning

confidence: 99%

Demystifying Parallel and Distributed Deep Learning

2019

View full text Add to dashboard Cite

Deep Neural Networks (DNNs) are becoming an important tool in modern computing applications. Accelerating their training is a major challenge and techniques range from distributed algorithms to low-level circuit design. In this survey, we describe the problem from a theoretical perspective, followed by approaches for its parallelization. We present trends in DNN architectures and the resulting implications on parallelization strategies. We then review and model the different types of concurrency in DNNs: from the single operator, through parallelism in network inference and training, to distributed deep learning. We discuss asynchronous stochastic optimization, distributed system architectures, communication schemes, and neural architecture search. Based on those approaches, we extrapolate potential directions for parallelism in deep learning.

show abstract

Section: 52mentioning

confidence: 99%

Demystifying Parallel and Distributed Deep Learning

2019

View full text Add to dashboard Cite

show abstract

“…Weight inheritance was found to improve the MSE in an image reconstruction experiment. Verbancsics and Harguess [18] test HyperNEAT [15] as a way to train CNNs for image classification. However, results were mediocre and could be substantially improved using backpropagation.…”

Section: Related Workmentioning

confidence: 99%

Lamarckian Evolution of Convolutional Neural Networks

Prellberg

Krämer

2018

Parallel Problem Solving From Nature – PPSN XV

View full text Add to dashboard Cite

Convolutional neural networks belong to the most successul image classifiers, but the adaptation of their network architecture to a particular problem is computationally expensive. We show that an evolutionary algorithm saves training time during the network architecture optimization, if learned network weights are inherited over generations by Lamarckian evolution. Experiments on typical image datasets show similar or significantly better test accuracies and improved convergence speeds compared to two different baselines without weight inheritance. On CIFAR-10 and CIFAR-100 a 75 % improvement in data efficiency is observed.

show abstract

“…Their approach is based on long offline learning sessions with subsequent testing procedures. In [18], neuro-evolution for deep learning is investigated, showing good results in training a feature extractor for use by other ML approaches. The authors of [19] apply neuro-evolution to general Atari game playing in the ALE.…”

Section: Related Workmentioning

confidence: 99%

Applying and Augmenting Deep Reinforcement Learning in Serious Games through Interaction

Dobrovsky

Borghoff

Hofmann

2017

Period. Polytech. Elec. Eng. Comp. Sci.

View full text Add to dashboard Cite

IntroductionThe term serious game (SG) origins from [1] and refers to games not exclusively developed for mere entertainment, but primarily for creating educational value. SGs are counted among the current e-learning trends and are gaining more and more acceptance and influence [2]. SG development slightly differs from that of entertainment games: SGs are often individual products for a restricted target audience, industrial branch or company. This results in high expenditure and deficient reusability, although the market shows growing interest in cost-efficient and customized applications [2]. One of the most relevant, but also most time-consuming tasks is the creation of reasonable, human-like non-player character (NPC) behaviour [3][4][5]. Thus, simplifying authoring and adaptation of AI in SGs seems desirable and profitable.Machine learning, especially reinforcement learning (RL), is occasionally used for automated NPC behaviour generation. Nevertheless, several issues arise when applying RL to create believable and diverse behaviour in complex scenarios. We indicate a way to overcome these issues by combining RL with human guidance, including effective collaboration between a learning system and human experts.In this paper, we give a short introduction to some challenges of behaviour generation in serious games and to the background of RL, deep reinforcement learning (DRL) and interactive reinforcement learning (iRL). We show related approaches and depict current issues of applying DRL methods to SGs. Furthermore, we introduce SanTrain as a SG providing challenging scenarios for NPC behaviour generation. Finally, we show how our approach of interactive deep reinforcement learning (iDRL), integrated into a flexible framework, could enhance AI development in SGs and exemplarily indicate valuable application opportunities in SanTrain.

show abstract

Image Classification Using Generative Neuro Evolution for Deep Learning

Cited by 47 publications

References 21 publications

Demystifying Parallel and Distributed Deep Learning

Demystifying Parallel and Distributed Deep Learning

Lamarckian Evolution of Convolutional Neural Networks

Applying and Augmenting Deep Reinforcement Learning in Serious Games through Interaction

Contact Info

Product

Resources

About