Learning Optimized Structure of Neural Networks by Hidden Node Pruning With L1 Regularization

Structured Pruning for Deep Convolutional Neural Networks: A Survey

The remarkable performance of deep Convolutional neural networks (CNNs) is generally attributed to their deeper and wider architectures, which can come with significant computational costs. Pruning neural networks has thus gained interest since it effectively lowers storage and computational costs. In contrast to weight pruning, which results in unstructured models, structured pruning provides the benefit of realistic acceleration by producing models that are friendly to hardware implementation. The special requirements of structured pruning have led to the discovery of numerous new challenges and the development of innovative solutions. This article surveys the recent progress towards structured pruning of deep CNNs. We summarize and compare the state-of-the-art structured pruning techniques with respect to filter ranking methods, regularization methods, dynamic execution, neural architecture search, the lottery ticket hypothesis, and the applications of pruning. While discussing structured pruning algorithms, we briefly introduce the unstructured pruning counterpart to emphasize their differences. Furthermore, we provide insights into potential research opportunities in the field of structured pruning. A curated list of neural network pruning papers can be found at: https://github.com/he-y/Awesome-Pruning. A dedicated website offering a more interactive comparison of structured pruning methods can be found at: https://huggingface.co/spaces/he-yang/Structured-Pruning-Survey.

Double Sparse Deep Reinforcement Learning via Multilayer Sparse Coding and Nonconvex Regularized Pruning

Deep reinforcement learning (DRL), which highly depends on the data representation, has shown its potential in many practical decision-making problems. However, the process of acquiring representations in DRL is easily affected by interference from models, and moreover leaves unnecessary parameters, leading to control performance reduction. In this article, we propose a double sparse DRL via multilayer sparse coding and nonconvex regularized pruning. To alleviate interference in DRL, we propose a multilayer sparse-coding-structural network to obtain deep sparse representation for control in reinforcement learning. Furthermore, we employ a nonconvex log regularizer to promote strong sparsity, efficiently removing the unnecessary weights with a regularizer-based pruning scheme. Hence, a double sparse DRL algorithm is developed, which can not only learn deep sparse representation to reduce the interference but also remove redundant weights while keeping the robust performance. The experimental results in five benchmark environments of the deep q network (DQN) architecture demonstrate that the proposed method with deep sparse representations from the multilayer sparse-coding structure can outperform existing sparse-coding-based DRL in control, for example, completing Mountain Car with 140.81 steps, achieving near 10% reward increase from the single-layer sparse-coding DRL algorithm, and obtaining 286.08 scores in Catcher, which are over two times the rewards of the other algorithms. Moreover, the proposed algorithm can reduce over 80% parameters while keeping performance improvements from deep sparse representations.

A fractional gradient descent algorithm robust to the initial weights of multilayer perceptron

For multilayer perceptron (MLP), the initial weights will significantly influence its performance. Based on the enhanced fractional derivative extend from convex optimization, this paper proposes a fractional gradient descent (RFGD) algorithm robust to the initial weights of MLP. We analyze the effectiveness of the RFGD algorithm. The convergence of the RFGD algorithm is also analyzed. The computational complexity of the RFGD algorithm is generally larger than that of the gradient descent (GD) algorithm but smaller than that of the Adam, Padam, AdaBelief, and AdaDiff algorithms. Numerical experiments show that the RFGD algorithm has strong robustness to the order of fractional calculus which is the only added parameter compared to the GD algorithm. More importantly, compared to the GD, Adam, Padam, AdaBelief, and AdaDiff algorithms, the experimental results show that the RFGD algorithm has the best robust performance for the initial weights of MLP. Meanwhile, the correctness of the theoretical analysis is verified.

Weighted Gate Layer Autoencoders

A single dataset could hide a significant number of relationships among its feature set. Learning these relationships simultaneously avoids the time complexity associated with running the learning algorithm for every possible relationship, and affords the learner with an ability to recover missing data and substitute erroneous ones by using available data. In our previous research, we introduced the gate-layer autoencoders (GLAEs), which offer an architecture that enables a single model to approximate multiple relationships simultaneously. GLAE controls what an autoencoder learns in a time series by switching on and off certain input gates, thus, allowing and disallowing the data to flow through the network to increase network's robustness. However, GLAE is limited to binary gates. In this article, we generalize the architecture to weighted gate layer autoencoders (WGLAE) through the addition of a weight layer to update the error according to which variables are more critical and to encourage the network to learn these variables. This new weight layer can also be used as an output gate and uses additional control parameters to afford the network with abilities to represent different models that can learn through gating the inputs. We compare the architecture against similar architectures in the literature and demonstrate that the proposed architecture produces more robust autoencoders with the ability to reconstruct both incomplete synthetic and real data with high accuracy.

An Advanced Pruning Method in the Architecture of Extreme Learning Machines Using L1-Regularization and Bootstrapping

Extreme learning machines (ELMs) are efficient for classification, regression, and time series prediction, as well as being a clear solution to backpropagation structures to determine values in intermediate layers of the learning model. One of the problems that an ELM may face is due to a large number of neurons in the hidden layer, making the expert model a specific data set. With a large number of neurons in the hidden layer, overfitting is more likely and thus unnecessary information can deterioriate the performance of the neural network. To solve this problem, a pruning method is proposed, called Pruning ELM Using Bootstrapped Lasso BR-ELM, which is based on regularization and resampling techniques, to select the most representative neurons for the model response. This method is based on an ensembled variant of Lasso (achieved through bootstrap replications) and aims to shrink the output weight parameters of the neurons to 0 as many and as much as possible. According to a subset of candidate regressors having significant coefficient values (greater than 0), it is possible to select the best neurons in the hidden layer of the ELM. Finally, pattern classification tests and benchmark regression tests of complex real-world problems are performed by comparing the proposed approach to other pruning models for ELMs. It can be seen that statistically BR-ELM can outperform several related state-of-the-art methods in terms of classification accuracies and model errors (while performing equally to Pruning-ELM P-ELM), and this with a significantly reduced number of finally selected neurons.