A formal selection and pruning algorithm for feedforward artificial neural network optimization

Learning beyond sensations: How dreams organize neuronal representations

Semantic representations in higher sensory cortices form the basis for robust, yet flexible behavior. These representations are acquired over the course of development in an unsupervised fashion and continuously maintained over an organism's lifespan. Predictive processing theories propose that these representations emerge from predicting or reconstructing sensory inputs. However, brains are known to generate virtual experiences, such as during imagination and dreaming, that go beyond previously experienced inputs. Here, we suggest that virtual experiences may be just as relevant as actual sensory inputs in shaping cortical representations. In particular, we discuss two complementary learning principles that organize representations through the generation of virtual experiences. First, "adversarial dreaming" proposes that creative dreams support a cortical implementation of adversarial learning in which feedback and feedforward pathways engage in a productive game of trying to fool each other. Second, "contrastive dreaming" proposes that the invariance of neuronal representations to irrelevant factors of variation is acquired by trying to map similar virtual experiences together via a contrastive learning process. These principles are compatible with known cortical structure and dynamics and the phenomenology of sleep thus providing promising directions to explain cortical learning beyond the classical predictive processing paradigm.

Efficient Harmonic Neural Networks With Compound Discrete Cosine Transform Filters and Shared Reconstruction Filters

The harmonic neural network (HNN) learns a combination of discrete cosine transform (DCT) filters to obtain an integrated feature from all spectra in the frequency domain. HNN, however, faces two challenges in learning and inference processes. First, the spectrum feature learned by HNN is insufficient and limited because the number of DCT filters is much smaller than that of feature maps. In addition, the number of parameters and the computation costs of HNN are significantly high because the intermediate spectrum layers are expanded multiple times. These two challenges will severely harm the performance and efficiency of HNN. To solve these problems, we first propose the compound DCT (C-DCT) filters integrating the nearest DCT filters to retrieve rich spectrum features to improve the performance. To significantly reduce the model size and computation complexity for improving the efficiency, the shared reconstruction filter is then proposed to share and dynamically drop the meta-filters in every frequency branch. Integrating the C-DCT filters with the shared reconstruction filters, the efficient harmonic network (EH-Net) is introduced. Extensive experiments on different datasets demonstrate that the proposed EH-Nets can effectively reduce the model size and computation complexity while maintaining the model performance. The code has been released at https://github.com/zhangle408/EH-Nets.

SRGC-Nets: Sparse Repeated Group Convolutional Neural Networks

Group convolution is widely used in many mobile networks to remove the filter's redundancy from the channel extent. In order to further reduce the redundancy of group convolution, this article proposes a novel repeated group convolutional (RGC) kernel, which has M primary groups, and each primary group includes N tiny groups. In every primary group, the same convolutional kernel is repeated in all the tiny groups. The RGC filter is the first kernel to remove the redundancy from group extent. Based on RGC, a sparse RGC (SRGC) kernel is also introduced in this article, and its corresponding network is called SRGC neural networks (SRGC-Net). The SRGC kernel is the summation of RGC kernel and pointwise group convolutional (PGC) kernel. The number of PGC's groups is M . Accordingly, in each primary group, besides the center locations in all channels, the values of parameters located in other N-1 tiny groups are all zero. Therefore, SRGC can significantly reduce the parameters. Moreover, it can also effectively retrieve spatial and channel-difference features by utilizing RGC and PGC to preserve the richness of produced features. Comparative experiments were performed on the benchmark classification data sets. Compared with the traditional popular networks, SRGC-Nets can perform better with timely reducing the model size and computational complexity. Furthermore, it can also achieve better performances than other latest state-of-the-art mobile networks on most of the databases and effectively decrease the test and training runtime.

Continuously Constructive Deep Neural Networks

Traditionally, deep learning algorithms update the network weights, whereas the network architecture is chosen manually using a process of trial and error. In this paper, we propose two novel approaches that automatically update the network structure while also learning its weights. The novelty of our approach lies in our parameterization, where the depth, or additional complexity, is encapsulated continuously in the parameter space through control parameters that add additional complexity. We propose two methods. In tunnel networks, this selection is done at the level of a hidden unit, and in budding perceptrons, this is done at the level of a network layer; updating this control parameter introduces either another hidden unit or layer. We show the effectiveness of our methods on the synthetic two-spiral data and on three real data sets of MNIST, MIRFLICKR, and CIFAR, where we see that our proposed methods, with the same set of hyperparameters, can correctly adjust the network complexity to the task complexity.

Pruning algorithms of neural networks — a comparative study

The neural network with optimal architecture speeds up the learning process and generalizes the problem well for further knowledge extraction. As a result researchers have developed various techniques for pruning the neural networks. This paper provides a survey of existing pruning techniques that optimize the architecture of neural networks and discusses their advantages and limitations. Also the paper evaluates the effectiveness of various pruning techniques by comparing the performance of some traditional and recent pruning algorithms based on sensitivity analysis, mutual information and significance on four real datasets namely Iris, Wisconsin breast cancer, Hepatitis Domain and Pima Indian Diabetes.

Waking and dreaming consciousness: neurobiological and functional considerations

This paper presents a theoretical review of rapid eye movement sleep with a special focus on pontine-geniculate-occipital waves and what they might tell us about the functional anatomy of sleep and consciousness. In particular, we review established ideas about the nature and purpose of sleep in terms of protoconsciousness and free energy minimization. By combining these theoretical perspectives, we discover answers to some fundamental questions about sleep: for example, why is homeothermy suspended during sleep? Why is sleep necessary? Why are we not surprised by our dreams? What is the role of synaptic regression in sleep? The imperatives for sleep that emerge also allow us to speculate about the functional role of PGO waves and make some empirical predictions that can, in principle, be tested using recent advances in the modeling of electrophysiological data.

AN INCREMENTAL FRAMEWORK BASED ON CROSS-VALIDATION FOR ESTIMATING THE ARCHITECTURE OF A MULTILAYER PERCEPTRON

We define the problem of optimizing the architecture of a multilayer perceptron (MLP) as a state space search and propose the MOST (Multiple Operators using Statistical Tests) framework that incrementally modifies the structure and checks for improvement using cross-validation. We consider five variants that implement forward/backward search, using single/multiple operators and searching depth-first/breadth-first. On 44 classification and 30 regression datasets, we exhaustively search for the optimal and evaluate the goodness based on: (1) Order, the accuracy with respect to the optimal and (2) Rank, the computational complexity. We check for the effect of two resampling methods (5 × 2, ten-fold cv), four statistical tests (5 × 2 cv t, ten-fold cv t, Wilcoxon, sign) and two corrections for multiple comparisons (Bonferroni, Holm). We also compare with Dynamic Node Creation (DNC) and Cascade Correlation (CC). Our results show that: (1) On most datasets, networks with few hidden units are optimal, (2) forward searching finds simpler architectures, (3) variants using single node additions (deletions) generally stop early and get stuck in simple (complex) networks, (4) choosing the best of multiple operators finds networks closer to the optimal, (5) MOST variants generally find simpler networks having lower or comparable error rates than DNC and CC.

Feature selection using a piecewise linear network

We present an efficient feature selection algorithm for the general regression problem, which utilizes a piecewise linear orthonormal least squares (OLS) procedure. The algorithm 1) determines an appropriate piecewise linear network (PLN) model for the given data set, 2) applies the OLS procedure to the PLN model, and 3) searches for useful feature subsets using a floating search algorithm. The floating search prevents the "nesting effect." The proposed algorithm is computationally very efficient because only one data pass is required. Several examples are given to demonstrate the effectiveness of the proposed algorithm.