SmaQ: Smart Quantization for DNN Training by Exploiting Value Clustering

Some Thoughts About Data Type, Distribution, and Statistical Significance

Neural network pruning has shown promising performance in reducing computational complexity and facilitate the deployment of deep neural networks on resource-limited edge devices. Most existing pruning methods focus on the indicators of the filter's weight, gradient, or feature map and regard the weak or similar filters as network redundancy. In contrast, the representation of discriminative power is also a fundamental attribute that analog neural networks to have extraordinary performance in various tasks. However, such representation is neglected in existing works. Alternatively, we propose a novel filter pruning strategy via class-wise discriminative power (CDP). Unlike the previous methods, CDP treats the filters that always yield large or small activation values as redundant and reserves the filters that show different magnitudes in activations as they yield high discriminative power. We further propose to obtain such discriminative power by employing the widely-used Term Frequency-Inverse Document Frequency (TF-IDF) on feature representations across classes. Specifically, the output of a filter is considered as a word, and the whole feature map is considered as a document. Then, TF-IDF is used to generate the relevant score between words and all documents. If a filter has low TF-IDF scores is less discriminate and can be pruned. Thus, the filters with high TF-IDF scores are reserved. To our best knowledge, this is the first work that prunes neural networks through class-wise discriminative power and measures such power by introducing TF-IDF in feature representation among different classes. Without any iterative process, CDP achieves better compression trade-offs comparing to the state-of-the-art compression algorithms. For instance, in VGG-16, we achieve a 68.05%-FLOPs reduction, with a 94.86% Top-1 accuracy on CIFAR-10. Specifically, we compress a 90.12%-FLOPs reduction VGG-16, even retains 93.30% Top-1 accuracy on CIFAR-10. The code is available at https://github.com/Tianshuo-Xu/CDP-Towards-Optimal-Filter-Pruning-via-Class-wise-Discriminative-Power.git

We propose an end-to-end approach for modeling polyphonic music with a novel graphical representation, based on music theory, in a deep neural network. Despite the success of deep learning in various applications, it remains a challenge to incorporate existing domain knowledge in a network without affecting its training routines. In this paper we present a novel approach for predictive music modeling and music generation that incorporates domain knowledge in its representation. In this work, music is transformed into a 2D representation, inspired by tonnetz from music theory, which graphically encodes musical relationships between pitches. This representation is incorporated in a deep network structure consisting of multilayered convolutional neural networks (CNN, for learning an efficient abstract encoding of the representation) and recurrent neural networks with long short-term memory cells (LSTM, for capturing temporal dependencies in music sequences). We empirically evaluate the nature and the effectiveness of the network by using a dataset of classical music from various composers. We investigate the effect of parameters including the number of convolution feature maps, pooling strategies, and three configurations of the network: LSTM without CNN, LSTM with CNN (pre-trained vs. not pre-trained). Visualizations of the feature maps and filters in the CNN are explored, and a comparison is made between the proposed tonnetz-inspired representation and pianoroll, a commonly used representation of music in computational systems. Experimental results show that the tonnetz representation produces musical sequences that are more tonally stable and contain more repeated patterns than sequences generated by pianoroll-based models, a finding that is directly useful for tackling current challenges in music and AI such as smart music generation.

Deep learning is finding its way into high energy physics by replacing traditional Monte Carlo simulations. However, deep learning still requires an excessive amount of computational resources. A promising approach to make deep learning more efficient is to quantize the parameters of the neural networks to reduced precision. Reduced precision computing is extensively used in modern deep learning and results to lower execution inference time, smaller memory footprint and less memory bandwidth. In this paper we analyse the effects of low precision inference on a complex deep generative adversarial network model. The use case which we are addressing is calorimeter detector simulations of subatomic particle interactions in accelerator based high energy physics. We employ the novel Intel low precision optimization tool (iLoT) for quantization and compare the results to the quantized model from TensorFlow Lite. In the performance benchmark we gain a speed-up of 1.73x on Intel hardware for the quantized iLoT model compared to the initial, not quantized, model. With different physics-inspired self-developed metrics, we validate that the quantized iLoT model shows a lower loss of physical accuracy in comparison to the TensorFlow Lite model.

Recent advances in deep learning allow on-demand reduction of model complexity, without a need for re-training, thus enabling a dynamic trade-off between the inference accuracy and the energy savings. Approximate mobile computing, on the other hand, adapts the computation approximation level as the context of usage, and consequently the computation needs or result accuracy needs, vary. In this work, we propose a synergy between the two directions and develop a context-aware method for dynamically adjusting the width of an on-device neural network based on the input and context-dependent classification confidence. We implement our method on a human activity recognition neural network and through measurements on a real-world embedded device demonstrate that such a network would save up to 37.8% energy and induce only 1% loss of accuracy, if used for continuous activity monitoring in the field of elderly care.

High-resolution remote sensing images usually contain complex semantic information and confusing targets, so their semantic segmentation is an important and challenging task. To resolve the problem of inadequate utilization of multilayer features by existing methods, a semantic segmentation method for remote sensing images based on convolutional neural network and mask generation is proposed. In this method, the boundary box is used as the initial foreground segmentation profile, and the edge information of the foreground object is obtained by using the multilayer feature of the convolutional neural network. In order to obtain the rough object segmentation mask, the general shape and position of the foreground object are estimated by using the high-level features in the process of layer-by-layer iteration. Then, based on the obtained rough mask, the mask is updated layer by layer using the neural network characteristics to obtain a more accurate mask. In order to solve the difficulty of deep neural network training and the problem of degeneration after convergence, a framework based on residual learning was adopted, which can simplify the training of those very deep networks and improve the accuracy of the network. For comparison with other advanced algorithms, the proposed algorithm was tested on the Potsdam and Vaihingen datasets. Experimental results show that, compared with other algorithms, the algorithm in this article can effectively improve the overall precision of semantic segmentation of high-resolution remote sensing images and shorten the overall training time and segmentation time.

In recent years, deep neural networks (DNNs) have achieved great successes in many areas, such as cognitive computation, pattern recognition, and computer vision. Although many hand-crafted deep networks have been proposed in the literature, designing a well-behaved neural network for a specific application requires high-level expertise yet. Hence, the automatic architecture design of DNNs has become a challenging and important problem. In this paper, we propose a new reinforcement learning method, whose action policy is to select neural blocks and construct deep networks. We define the action search space with three types of neural blocks, i.e., dense block, residual block, and inception-like block. Additionally, we have also designed several variants for the residual and inception-like blocks. The optimal network is automatically learned by a Q-learning agent, which is iteratively trained to generate well-performed deep networks. To evaluate the proposed method, we have conducted experiments on three datasets, MNIST, SVHN, and CIFAR-10, for image classification applications. Compared with existing hand-crafted and auto-generated neural networks, our auto-designed neural network delivers promising results. Moreover, the proposed reinforcement learning algorithm for deep networks design only runs on one GPU, demonstrating much higher efficiency than most of the previous deep network search approaches.

Quantization-aware training for low precision photonic neural networks

The rapid growth of genomic data presents new challenges and opportunities in the field of bioinformatics. Deep learning has emerged as a powerful tool for analyzing large and complex genomic datasets, offering unprecedented capabilities for identifying patterns, predicting gene functions, and understanding genetic variations. This article explores the latest advancements in deep learning techniques applied to genomic data analysis. It discusses key approaches such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and autoencoders, and their applications in various aspects of genomics, including gene expression analysis, variant calling, and drug discovery. The article also highlights the integration of deep learning with other emerging technologies, such as CRISPR, and its potential to revolutionize personalized medicine and genetic research..

Recent advances in deep learning have bolstered our ability to forecast the evolution of dynamical systems, but common neural networks do not adhere to physical laws, critical information that could lead to sounder state predictions. This contribution addresses this concern by proposing a neural network to polynomial (NN-Poly) approximation, a method that furnishes algorithmic guarantees of adhering to physics while retaining state prediction accuracy. To achieve these goals, this article shows how to represent a trained fully connected perceptron, convolution, and recurrent neural networks of various activation functions as Taylor polynomials of arbitrary order. This solution is not only analytic in nature but also least squares optimal. The NN-Poly system identification or state prediction method is evaluated against a single-layer neural network and a polynomial trained on data generated by dynamic systems. Across our test cases, the proposed method maintains minimal root mean-squared state error, requires few parameters to form, and enables model structure for verification and safety. Future work will incorporate safety constraints into state predictions, with this new model structure and test high-dimensional dynamical system data.

Deep learning (DL) has shown explosive growth in its application to bioinformatics and has demonstrated thrillingly promising power to mine the complex relationship hidden in large-scale biological and biomedical data. A number of comprehensive reviews have been published on such applications, ranging from high-level reviews with future perspectives to those mainly serving as tutorials. These reviews have provided an excellent introduction to and guideline for applications of DL in bioinformatics, covering multiple types of machine learning (ML) problems, different DL architectures, and ranges of biological/biomedical problems. However, most of these reviews have focused on previous research, whereas current trends in the principled DL field and perspectives on their future developments and potential new applications to biology and biomedicine are still scarce. We will focus on modern DL, the ongoing trends and future directions of the principled DL field, and postulate new and major applications in bioinformatics.

Spiking neural networks (SNNs) have attracted many researchers’ interests due to its biological plausibility and event-driven characteristic. In particular, recently, many studies on high-performance SNNs comparable to the conventional analog-valued neural networks (ANNs) have been reported by converting weights trained from ANNs into SNNs. However, unlike ANNs, SNNs have an inherent latency that is required to reach the best performance because of differences in operations of neuron. In SNNs, not only spatial integration but also temporal integration exists, and the information is encoded by spike trains rather than values in ANNs. Therefore, it takes time to achieve a steady-state of the performance in SNNs. The latency is worse in deep networks and required to be reduced for the practical applications. In this work, we propose a pre-charged membrane potential (PCMP) for the latency reduction in SNN. A variety of neural network applications (e.g., classification, autoencoder using MNIST and CIFAR-10 datasets) are trained and converted to SNNs to demonstrate the effect of the proposed approach. The latency of SNNs is successfully reduced without accuracy loss. In addition, we propose a delayed evaluation method (DE), by which the errors during the initial transient are discarded. The error spikes occurring in the initial transient is removed by DE, resulting in the further latency reduction. DE can be used in combination with PCMP for further latency reduction. Finally, we also show the advantages of the proposed methods in improving the number of spikes required to reach a steady-state of the performance in SNNs for energy-efficient computing.

Neural network pruning, an important method to reduce the computational complexity of deep models, can be well applied to devices with limited resources. However, most current methods focus on some kind of information about the filter itself to prune the network, rarely exploring the relationship between the feature maps and the filters. In this paper, two novel pruning methods are proposed. First, a new pruning method is proposed, which reflects the importance of filters by exploring the information in the feature maps. Based on the premise that the more information there is, more important the feature map is, the information entropy of feature maps is used to measure information, which is used to evaluate the importance of each filter in the current layer. Further, normalization is used to realize cross layer comparison. As a result, based on the method mentioned above, the network structure is efficiently pruned while its performance is well reserved. Second, we proposed a parallel pruning method using the combination of our pruning method above and slimming pruning method which has better results in terms of computational cost. Our methods perform better in terms of accuracy, parameters, and FLOPs compared to most advanced methods. On ImageNet, it is achieved 72.02% top1 accuracy for ResNet50 with merely 11.41M parameters and 1.12B FLOPs.For DenseNet40, it is obtained 94.04% accuracy with only 0.38M parameters and 110.72M FLOPs on CIFAR10, and our parallel pruning method makes the parameters and FLOPs are just 0.37M and 100.12M, respectively, with little loss of accuracy.

Recent advances in deep learning have had a methodological and practical impact on brain-computer interface (BCI) research. Among the various deep network architectures, convolutional neural networks (CNNs) have been well suited for spatio-spectral-temporal electroencephalogram (EEG) signal representation learning. Most of the existing CNN-based methods described in the literature extract features at a sequential level of abstraction with repetitive nonlinear operations and involve densely connected layers for classification. However, studies in neurophysiology have revealed that EEG signals carry information in different ranges of frequency components. To better reflect these multi-frequency properties in EEGs, we propose a novel deep multi-scale neural network that discovers feature representations in multiple frequency/time ranges and extracts relationships among electrodes, i.e., spatial representations, for subject intention/condition identification. Furthermore, by completely representing EEG signals with spatio-spectral-temporal information, the proposed method can be utilized for diverse paradigms in both active and passive BCIs, contrary to existing methods that are primarily focused on single-paradigm BCIs. To demonstrate the validity of our proposed method, we conducted experiments on various paradigms of active/passive BCI datasets. Our experimental results demonstrated that the proposed method achieved performance improvements when judged against comparable state-of-the-art methods. Additionally, we analyzed the proposed method using different techniques, such as PSD curves and relevance score inspection to validate the multi-scale EEG signal information capturing ability, activation pattern maps for investigating the learned spatial filters, and t-SNE plotting for visualizing represented features. Finally, we also demonstrated our method's application to real-world problems. Based on our experimental results and analyses, we believe that the proposed multi-scale neural network can be useful for various BCI paradigms, as a starting model or as a backbone network in any new BCI experiments.

Filter pruning with uniqueness mechanism in the frequency domain for efficient neural networks