Evolving Deep Neural Networks with Explanations for Image Classification

Abstract

<p><b>Image classification problems often face the issues of high dimensionality and large variance within the same class. Deep convolutional neural networks are designed to solve the problem by extracting features using convolutional operations. Researchers have developed complex deep convolutional neural networks to achieve the outstanding performance that outperforms humans. However, the complexity of deep convolutional neural networks brings two side effects. First, the more complex the network architecture is, the harder it is to design. Second, it deteriorates the black-box nature of deep convolutional neural networks, which is harder to explain. To tackle the above two issues, neural architecture search and explainable deep learning have emerged as two promising research areas for automatically designing deep convolutional neural networks and providing explanations of the predictions made by deep learning models, respectively. Evolutionary computation based neural architecture search has been employed to automatically design deep convolutional neural networks that outperform those manually designed, but the computational cost is too high. Surrogate-assisted and transfer learning based methods can be utilised to reduce the computational cost. Furthermore, a branch in explainable deep learning called local approximation does not need machine learning expertise to understand the explanation. The target of local approximation is to find interpretable features. Evolutionary computation has successfully applied to many search problems, but it has never been used in finding local approximation. </b></p> <p>The overall goal of this thesis is to improve the efficiency of evolutionary neural architecture search in image classification tasks and provide explanations in the inference phase. This can mitigate the major issues -- the difficulty of designing deep convolutional neural networks and the lack of explainability, that significantly affected deep learning being widely used in real-world applications. </p> <p>This thesis proposes a surrogate-assisted particle swarm optimisation (an evolutionary computation algorithm) method to efficiently evolve deep convolutional neural networks. A surrogate model is proposed to predict a better solution from a pair of solutions, and a method for sampling a subset of the dataset as a surrogate dataset is proposed to reduce the computational cost to less than 3 GPU-days, but with very competitive classification accuracies across several benchmark datasets. </p> <p>Next, this thesis proposes to use smaller datasets in multiple source domains to evolve deep convolutional neural networks, and then transfer the learned models to the target domain. This has achieved the goal of accelerating the evolutionary neural architecture search process and improving the generalisation performance, which is supported by the experiment results. </p> <p>Furthermore, this thesis proposes a genetic algorithm (an evolutionary computation algorithm) based method to evolve local explanations to explain the predictions of deep convolutional neural networks. By combining the flexible encoding and the proposed fitness evaluation, the proposed method can efficiently evolve meaning interpretable features in the local explanation. It produces competitive explanations, but 10 times faster than Local Interpretable Model-agnostic Explanations (LIME) -- the state-of-the-art method. </p> <p>Lastly, an evolutionary multi-objective approach is explored to evolve local explanations to reduce the human effort of examining the interpretable features in the local explanations. It adds the second objective of minimising the number of interpretable features, which reduces the human effort in checking them. Additionally, it proposes a method to select only two non-dominated solutions from the pareto front to save the labour cost for end users.</p>

Stochastic three-term conjugate gradient method with variance technique for non-convex learning

“…Among them, Yuan [26] obtained the global convergence of PRP through a modified wolf line search. Recently, the adaptive conjugate gradient method proposed by Wang [28] has a better performance in training neural networks for image processing. CGVR is a hybrid of FR and PRP methods on the basis of SVRG [17], and SCGA is a similar work on SAGA [18].…”