ICPR2020 Paper Browser

Paper download is intended for registered attendees only, and is subjected to the IEEE Copyright Policy. Any other use is strongly forbidden.

Exploiting Non-Linear Redundancy for Neural Model Compression

Muhammad Ahmed Shah, Raphael Olivier, Bhiksha Raj

Auto-TLDR; Compressing Deep Neural Networks with Linear Dependency

Abstract Slides Poster

Deploying deep learning models with millions, even billions, of parameters is challenging given real world memory, power and compute constraints. In an effort to make these models more practical, in this paper, we propose a novel model compression approach that exploits linear dependence between the activations in a layer to eliminate entire structural units (neurons/convolutional filters). Our approach also adjusts the weights of the layer in a manner that is provably lossless while training if the removed neuron was perfectly predictable. We combine this approach with an annealing algorithm that may be applied during training, or even on a trained model, and demonstrate, using popular datasets, that our technique can reduce the parameters of VGG and AlexNet by more than 97\% on \cifar, 85\% on \caltech, and 19\% on ImageNet at less than 2\% loss in accuracy. Furthermore, we provide theoretical results showing that in overparametrized, locally linear (ReLU) neural networks where redundant features exist, and with correct hyperparameter selection, our method is indeed able to capture and suppress those dependencies.

Similar papers

HFP: Hardware-Aware Filter Pruning for Deep Convolutional Neural Networks Acceleration

Fang Yu, Chuanqi Han, Pengcheng Wang, Ruoran Huang, Xi Huang, Li Cui

Auto-TLDR; Hardware-Aware Filter Pruning for Convolutional Neural Networks

Abstract Slides Poster Similar

Convolutional Neural Networks (CNNs) are powerful but computationally demanding and memory intensive, thus impeding their practical applications on resource-constrained hardware. Filter pruning is an efficient approach for deep CNN compression and acceleration, which aims to eliminate some filters with tolerable performance degradation. In the literature, the majority of approaches prune networks by defining the redundant filters or training the networks with a sparsity prior loss function. These approaches mainly use FLOPs as their speed metric. However, the inference latency of pruned networks cannot be directly controlled on the hardware platform, which is an important dimension of practicality. To address this issue, we propose a novel Hardware-aware Filter Pruning method (HFP) which can produce pruned networks that satisfy the actual latency budget on the hardwares of interest. In addition, we propose an iterative pruning framework called Opti-Cut to decrease the accuracy degradation of pruning process and accelerate the pruning procedure whilst meeting the hardware budget. More specifically, HFP first builds up a lookup table for fast estimating the latency of target network about filter configuration layer by layer. Then, HFP leverages information gain (IG) to globally evaluate the filters' contribution to network output distribution. HFP utilizes the Opti-Cut framework to globally prune filters with the minimum IG one by one until the latency budget is satisfied. We verify the effectiveness of the proposed method on CIFAR-10 and ImageNet. Compared with the state-of-the-art pruning methods, HFP demonstrates superior performances on VGGNet, ResNet and MobileNet V1/V2.

Neuron-Based Network Pruning Based on Majority Voting

Ali Alqahtani, Xianghua Xie, Ehab Essa, Mark W. Jones

Auto-TLDR; Large-Scale Neural Network Pruning using Majority Voting

Abstract Slides Poster Similar

The achievement of neural networks in a variety of applications is accompanied by a dramatic increase in computational costs and memory requirements. In this paper, we propose an efficient method to simultaneously identify the critical neurons and prune the model during training without involving any pre-training or fine-tuning procedures. Unlike existing methods, which accomplish this task in a greedy fashion, we propose a majority voting technique to compare the activation values among neurons and assign a voting score to quantitatively evaluate their importance.This mechanism helps to effectively reduce model complexity by eliminating the less influential neurons and aims to determine a subset of the whole model that can represent the reference model with much fewer parameters within the training process. Experimental results show that majority voting efficiently compresses the network with no drop in model accuracy, pruning more than 79\% of the original model parameters on CIFAR10 and more than 91\% of the original parameters on MNIST. Moreover, we show that with our proposed method, sparse models can be further pruned into even smaller models by removing more than 60\% of the parameters, whilst preserving the reference model accuracy.

On the Information of Feature Maps and Pruning of Deep Neural Networks

Mohammadreza Soltani, Suya Wu, Jie Ding, Robert Ravier, Vahid Tarokh

Auto-TLDR; Compressing Deep Neural Models Using Mutual Information

Abstract Slides Poster Similar

A technique for compressing deep neural models achieving competitive performance to state-of-the-art methods is proposed. The approach utilizes the mutual information between the feature maps and the output of the model in order to prune the redundant layers of the network. Extensive numerical experiments on both CIFAR-10, CIFAR-100, and Tiny ImageNet data sets demonstrate that the proposed method can be effective in compressing deep models, both in terms of the numbers of parameters and operations. For instance, by applying the proposed approach to DenseNet model with 0.77 million parameters and 293 million operations for classification of CIFAR-10 data set, a reduction of 62.66% and 41.00% in the number of parameters and the number of operations are respectively achieved, while increasing the test error only by less than 1%.

Learning Sparse Deep Neural Networks Using Efficient Structured Projections on Convex Constraints for Green AI

Michel Barlaud, Frederic Guyard

Auto-TLDR; Constrained Deep Neural Network with Constrained Splitting Projection

Abstract Slides Poster Similar

In recent years, deep neural networks (DNN) have been applied to different domains and achieved dramatic performance improvements over state-of-the-art classical methods. These performances of DNNs were however often obtained with networks containing millions of parameters and which training required heavy computational power. In order to cope with this computational issue a huge literature deals with proximal regularization methods which are time consuming.\\ In this paper, we propose instead a constrained approach. We provide the general framework for our new splitting projection gradient method. Our splitting algorithm iterates a gradient step and a projection on convex sets. We study algorithms for different constraints: the classical $\ell_1$ unstructured constraint and structured constraints such as the nuclear norm, the $\ell_{2,1} $ constraint (Group LASSO). We propose a new $\ell_{1,1} $ structured constraint for which we provide a new projection algorithm We demonstrate the effectiveness of our method on three popular datasets (MNIST, Fashion MNIST and CIFAR). Experiments on these datasets show that our splitting projection method with our new $\ell_{1,1} $ structured constraint provides the best reduction of memory and computational power. Experiments show that fully connected linear DNN are more efficient for green AI.

Activation Density Driven Efficient Pruning in Training

Timothy Foldy-Porto, Yeshwanth Venkatesha, Priyadarshini Panda

Auto-TLDR; Real-Time Neural Network Pruning with Compressed Networks

Abstract Slides Poster Similar

Neural network pruning with suitable retraining can yield networks with considerably fewer parameters than the original with comparable degrees of accuracy. Typical pruning methods require large, fully trained networks as a starting point from which they perform a time-intensive iterative pruning and retraining procedure to regain the original accuracy. We propose a novel pruning method that prunes a network real-time during training, reducing the overall training time to achieve an efficient compressed network. We introduce an activation density based analysis to identify the optimal relative sizing or compression for each layer of the network. Our method is architecture agnostic, allowing it to be employed on a wide variety of systems. For VGG-19 and ResNet18 on CIFAR-10, CIFAR-100, and TinyImageNet, we obtain exceedingly sparse networks (up to $200 \times$ reduction in parameters and over $60 \times$ reduction in inference compute operations in the best case) with accuracy comparable to the baseline network. By reducing the network size periodically during training, we achieve total training times that are shorter than those of previously proposed pruning methods. Furthermore, training compressed networks at different epochs with our proposed method yields considerable reduction in training compute complexity ($1.6\times$ to $3.2\times$ lower) at near iso-accuracy as compared to a baseline network trained entirely from scratch.

Learning to Prune in Training via Dynamic Channel Propagation

Shibo Shen, Rongpeng Li, Zhifeng Zhao, Honggang Zhang, Yugeng Zhou

Auto-TLDR; Dynamic Channel Propagation for Neural Network Pruning

Abstract Slides Poster Similar

In this paper, we propose a novel network training mechanism called "dynamic channel propagation" to prune the model during the training period. In particular, we pick up a specific group of channels in each convolutional layer to participate in the forward propagation in training time according to the significance level of channel, which is defined as channel utility. The utility values with respect to all selected channels are updated simultaneously with the error back-propagation process and will constantly change. Furthermore, when the training ends, channels with high utility values are retained whereas those with low utility values are discarded. Hence, our proposed method trains and prunes neural networks simultaneously. We empirically evaluate our novel training method on various representative benchmark datasets and advanced convolutional neural network (CNN) architectures, including VGGNet and ResNet. The experiment results verify superior performance and robust effectiveness of our approach.

A Discriminant Information Approach to Deep Neural Network Pruning

Zejiang Hou, Sy Kung

Auto-TLDR; Channel Pruning Using Discriminant Information and Reinforcement Learning

Abstract Slides Poster Similar

Network pruning has become the de facto tool to accelerate and compress deep convolutional neural networks for mobile and edge applications. Previous works tend to perform channel selection in layer-wise manner based on predefined heuristics, without considering layer importance or systematically optimizing the pruned structure. In this work, we propose a novel channel pruning method that jointly harnesses two strategies: (1) a channel importance ranking heuristics based on the feature-maps discriminant power, (2) a searching method for optimal pruning budget allocation. For the former, we propose a Discriminant Information (DI) based channel selection algorithm. We use a small batch of training samples to compute the DI score for each channel and rank the channel importance so that channels really contributing to the feature-maps discriminant power are retained. For the latter, in order to search the optimal pruning budget allocation, we formulate a reward maximization problem to discover the layer importance and generating the pruning budget accordingly. Such reward maximization can be efficiently solved by the policy gradient algorithm in reinforcement learning, yielding our final pruned network which achieves the best accuracy-efficiency trade-off. Experiments on a variety of CNN architectures and benchmark datasets show that our proposed channel pruning methods compare favorably with previous state-of-the-art methods. On ImageNet, our pruned MobileNetV2 outperforms the previous layer-wise state-of-the-art pruning method CPLI \cite{guo2020channel} by 2\% Top-1 accuracy while reducing the FLOPs by 50\%.

Progressive Gradient Pruning for Classification, Detection and Domain Adaptation

Le Thanh Nguyen-Meidine, Eric Granger, Marco Pedersoli, Madhu Kiran, Louis-Antoine Blais-Morin

Auto-TLDR; Progressive Gradient Pruning for Iterative Filter Pruning of Convolutional Neural Networks

Abstract Slides Poster Similar

Although deep neural networks (NNs) have achieved state-of-the-art accuracy in many visual recognition tasks, the growing computational complexity and energy consumption of networks remains an issue, especially for applications on plat-forms with limited resources and requiring real-time processing.Filter pruning techniques have recently shown promising results for the compression and acceleration of convolutional NNs(CNNs). However, these techniques involve numerous steps and complex optimisations because some only prune after training CNNs, while others prune from scratch during training by integrating sparsity constraints or modifying the loss function.In this paper we propose a new Progressive Gradient Pruning(PGP) technique for iterative filter pruning during training. In contrast to previous progressive pruning techniques, it relies on a novel filter selection criterion that measures the change in filter weights, uses a new hard and soft pruning strategy and effectively adapts momentum tensors during the backward propagation pass. Experimental results obtained after training various CNNs on image data for classification, object detection and domain adaptation benchmarks indicate that the PGP technique can achieve a better trade-off between classification accuracy and network (time and memory) complexity than PSFP and other state-of-the-art filter pruning techniques.

Softer Pruning, Incremental Regularization

Linhang Cai, Zhulin An, Yongjun Xu

Auto-TLDR; Asymptotic SofteR Filter Pruning for Deep Neural Network Pruning

Abstract Slides Poster Similar

Network pruning is widely used to compress Deep Neural Networks (DNNs). The Soft Filter Pruning (SFP) method zeroizes the pruned filters during training while updating them in the next training epoch. Thus the trained information of the pruned filters is completely dropped. To utilize the trained pruned filters, we proposed a SofteR Filter Pruning (SRFP) method and its variant, Asymptotic SofteR Filter Pruning (ASRFP), simply decaying the pruned weights with a monotonic decreasing parameter. Our methods perform well across various netowrks, datasets and pruning rates, also transferable to weight pruning. On ILSVRC-2012, ASRFP prunes 40% of the parameters on ResNet-34 with 1.63% top-1 and 0.68% top-5 accuracy improvement. In theory, SRFP and ASRFP are an incremental regularization of the pruned filters. Besides, We note that SRFP and ASRFP pursue better results while slowing down the speed of convergence.

MINT: Deep Network Compression Via Mutual Information-Based Neuron Trimming

Madan Ravi Ganesh, Jason Corso, Salimeh Yasaei Sekeh

Auto-TLDR; Mutual Information-based Neuron Trimming for Deep Compression via Pruning

Abstract Slides Poster Similar

Most approaches to deep neural network compression via pruning either evaluate a filter’s importance using its weights or optimize an alternative objective function with sparsity constraints. While these methods offer a useful way to approximate contributions from similar filters, they often either ignore the dependency between layers or solve a more difficult optimization objective than standard cross-entropy. Our method, Mutual Information-based Neuron Trimming (MINT), approaches deep compression via pruning by enforcing sparsity based on the strength of the relationship between filters of adjacent layers, across every pair of layers. The relationship is calculated using conditional geometric mutual information which evaluates the amount of similar information exchanged between the filters using a graph-based criterion. When pruning a network, we ensure that retained filters contribute the majority of the information towards succeeding layers which ensures high performance. Our novel approach outperforms existing state-of-the-art compression-via-pruning methods on the standard benchmarks for this task: MNIST, CIFAR-10, and ILSVRC2012, across a variety of network architectures. In addition, we discuss our observations of a common denominator between our pruning methodology’s response to adversarial attacks and calibration statistics when compared to the original network.

Slimming ResNet by Slimming Shortcut

Donggyu Joo, Doyeon Kim, Junmo Kim

Auto-TLDR; SSPruning: Slimming Shortcut Pruning on ResNet Based Networks

Abstract Slides Poster Similar

Conventional network pruning methods on convolutional neural networks (CNNs) reduce the number of input or output channels of convolution layers. With these approaches, the channels in the plain network can be pruned without any restrictions. However, in case of the ResNet based networks which have shortcuts (skip connections), the channel slimming of existing pruning methods is limited to the inside of each residual block. Since the number of Flops and parameters are also highly related to the number of channels in the shortcuts, more investigation on pruning channels in shortcuts is required. In this paper, we propose a novel pruning method, Slimming Shortcut Pruning (SSPruning), for pruning channels in shortcuts on ResNet based networks. First, we separate the long shortcut in individual regions that can be pruned independently without considering its long connections. Then, by applying our Importance Learning Gate (ILG) which learns the importance of channels globally regardless of channel type and location (i.e., in the shortcut or inside of the block), we can finally achieve an optimally pruned model. Through various experiments, we have confirmed that our method yields outstanding results when we prune the shortcuts and inside of the block together.

Exploiting Elasticity in Tensor Ranks for Compressing Neural Networks

Jie Ran, Rui Lin, Hayden Kwok-Hay So, Graziano Chesi, Ngai Wong

Auto-TLDR; Nuclear-Norm Rank Minimization Factorization for Deep Neural Networks

Abstract Slides Poster Similar

Elasticities in depth, width, kernel size and resolution have been explored in compressing deep neural networks (DNNs). Recognizing that the kernels in a convolutional neural network (CNN) are 4-way tensors, we further exploit a new elasticity dimension along the input-output channels. Specifically, a novel nuclear-norm rank minimization factorization (NRMF) approach is proposed to dynamically and globally search for the reduced tensor ranks during training. Correlation between tensor ranks across multiple layers is revealed, and a graceful tradeoff between model size and accuracy is obtained. Experiments then show the superiority of NRMF over the previous non-elastic variational Bayesian matrix factorization (VBMF) scheme.

Speeding-Up Pruning for Artificial Neural Networks: Introducing Accelerated Iterative Magnitude Pruning

Marco Zullich, Eric Medvet, Felice Andrea Pellegrino, Alessio Ansuini

Auto-TLDR; Iterative Pruning of Artificial Neural Networks with Overparametrization

Abstract Slides Poster Similar

In recent years, Artificial Neural Networks (ANNs) pruning has become the focal point of many researches, due to the extreme overparametrization of such models. This has urged the scientific world to investigate methods for the simplification of the structure of weights in ANNs, mainly in an effort to reduce time for both training and inference. Frankle and Carbin and later Renda, Frankle, and Carbin introduced and refined an iterative pruning method which is able to effectively prune the network of a great portion of its parameters with little to no loss in performance. On the downside, this method requires a large amount of time for its application, since, for each iteration, the network has to be trained for (almost) the same amount of epochs of the unpruned network. In this work, we show that, for a limited setting, if targeting high overall sparsity rates, this time can be effectively reduced for each iteration, save for the last one, by more than 50%, while yielding a final product (i.e., final pruned network) whose performance is comparable to the ANN obtained using the existing method.

Compact CNN Structure Learning by Knowledge Distillation

Waqar Ahmed, Andrea Zunino, Pietro Morerio, Vittorio Murino

Auto-TLDR; Knowledge Distillation for Compressing Deep Convolutional Neural Networks

Abstract Slides Poster Similar

The concept of compressing deep Convolutional Neural Networks (CNNs) is essential to use limited computation, power, and memory resources on embedded devices. However, existing methods achieve this objective at the cost of a drop in inference accuracy in computer vision tasks. To address such a drawback, we propose a framework that leverages knowledge distillation along with customizable block-wise optimization to learn a lightweight CNN structure while preserving better control over the compression-performance tradeoff. Considering specific resource constraints, e.g., floating-point operations per second (FLOPs) or model-parameters, our method results in a state of the art network compression while being capable of achieving better inference accuracy. In a comprehensive evaluation, we demonstrate that our method is effective, robust, and consistent with results over a variety of network architectures and datasets, at negligible training overhead. In particular, for the already compact network MobileNet_v2, our method offers up to 2x and 5.2x better model compression in terms of FLOPs and model-parameters, respectively, while getting 1.05% better model performance than the baseline network.

Filter Pruning Using Hierarchical Group Sparse Regularization for Deep Convolutional Neural Networks

Kakeru Mitsuno, Takio Kurita

Auto-TLDR; Hierarchical Group Sparse Regularization for Sparse Convolutional Neural Networks

Abstract Slides Poster Similar

Since the convolutional neural networks are often trained with redundant parameters, it is possible to reduce redundant kernels or filters to obtain a compact network without dropping the classification accuracy. In this paper, we propose a filter pruning method using the hierarchical group sparse regularization. It is shown in our previous work that the hierarchical group sparse regularization is effective in obtaining sparse networks in which filters connected to unnecessary channels are automatically close to zero. After training the convolutional neural network with the hierarchical group sparse regularization, the unnecessary filters are selected based on the increase of the classification loss of the randomly selected training samples to obtain a compact network. It is shown that the proposed method can reduce more than 50% parameters of ResNet for CIFAR-10 with only 0.3% decrease in the accuracy of test samples. Also, 34% parameters of ResNet are reduced for TinyImageNet-200 with higher accuracy than the baseline network.

Efficient Online Subclass Knowledge Distillation for Image Classification

Maria Tzelepi, Nikolaos Passalis, Anastasios Tefas

Auto-TLDR; OSKD: Online Subclass Knowledge Distillation

Abstract Slides Poster Similar

Deploying state-of-the-art deep learning models on embedded systems dictates certain storage and computation limitations. During the recent few years Knowledge Distillation (KD) has been recognized as a prominent approach to address this issue. That is, KD has been effectively proposed for training fast and compact deep learning models by transferring knowledge from more complex and powerful models. However, knowledge distillation, in its conventional form, involves multiple stages of training, rendering it a computationally and memory demanding procedure. In this paper, a novel single-stage self knowledge distillation method is proposed, namely Online Subclass Knowledge Distillation (OSKD), that aims at revealing the similarities inside classes, improving the performance of any deep neural model in an online manner. Hence, as opposed to existing online distillation methods, we are able to acquire further knowledge from the model itself, without building multiple identical models or using multiple models to teach each other, rendering the OSKD approach more efficient. The experimental evaluation on two datasets validates that the proposed method improves the classification performance.

Compression Strategies and Space-Conscious Representations for Deep Neural Networks

Giosuè Marinò, Gregorio Ghidoli, Marco Frasca, Dario Malchiodi

Auto-TLDR; Compression of Large Convolutional Neural Networks by Weight Pruning and Quantization

Abstract Slides Poster Similar

Recent advances in deep learning have made available large, powerful convolutional neural networks (CNN) with state-of-the-art performance in several real-world applications. Unfortunately, these large-sized models have millions of parameters, thus they are not deployable on resource-limited platforms (e.g. where RAM is limited). Compression of CNNs thereby becomes a critical problem to achieve memory-efficient and possibly computationally faster model representations. In this paper, we investigate the impact of lossy compression of CNNs by weight pruning and quantization, and lossless weight matrix representations based on source coding. We tested several combinations of these techniques on four benchmark datasets for classification and regression problems, achieving compression rates up to 165 times, while preserving or improving the model performance.

Attention Based Pruning for Shift Networks

Ghouthi Hacene, Carlos Lassance, Vincent Gripon, Matthieu Courbariaux, Yoshua Bengio

Auto-TLDR; Shift Attention Layers for Efficient Convolutional Layers

Abstract Slides Poster Similar

In many application domains such as computer vision, Convolutional Layers (CLs) are key to the accuracy of deep learning methods. However, it is often required to assemble a large number of CLs, each containing thousands of parameters, in order to reach state-of-the-art accuracy, thus resulting in complex and demanding systems that are poorly fitted to resource-limited devices. Recently, methods have been proposed to replace the generic convolution operator by the combination of a shift operation and a simpler 1x1 convolution. The resulting block, called Shift Layer (SL), is an efficient alternative to CLs in the sense it allows to reach similar accuracies on various tasks with faster computations and fewer parameters. In this contribution, we introduce Shift Attention Layers (SALs), which extend SLs by using an attention mechanism that learns which shifts are the best at the same time the network function is trained. We demonstrate SALs are able to outperform vanilla SLs (and CLs) on various object recognition benchmarks while significantly reducing the number of float operations and parameters for the inference.

RNN Training along Locally Optimal Trajectories via Frank-Wolfe Algorithm

Yun Yue, Ming Li, Venkatesh Saligrama, Ziming Zhang

Auto-TLDR; Frank-Wolfe Algorithm for Efficient Training of RNNs

Abstract Slides Poster Similar

We propose a novel and efficient training method for RNNs by iteratively seeking a local minima on the loss surface within a small region, and leverage this directional vector for the update, in an outer-loop. We propose to utilize the Frank-Wolfe (FW) algorithm in this context. Although, FW implicitly involves normalized gradients, which can lead to a slow convergence rate, we develop a novel RNN training method that, surprisingly, even with the additional cost, the overall training cost is empirically observed to be lower than back-propagation. Our method leads to a new Frank-Wolfe method, that is in essence an SGD algorithm with a restart scheme. We prove that under certain conditions our algorithm has a sublinear convergence rate of $O(1/\epsilon)$ for $\epsilon$ error. We then conduct empirical experiments on several benchmark datasets including those that exhibit long-term dependencies, and show significant performance improvement. We also experiment with deep RNN architectures and show efficient training performance. Finally, we demonstrate that our training method is robust to noisy data.

Channel Planting for Deep Neural Networks Using Knowledge Distillation

Kakeru Mitsuno, Yuichiro Nomura, Takio Kurita

Auto-TLDR; Incremental Training for Deep Neural Networks with Knowledge Distillation

Abstract Slides Poster Similar

In recent years, deeper and wider neural networks have shown excellent performance in computer vision tasks, while their enormous amount of parameters results in increased computational cost and overfitting. Several methods have been proposed to compress the size of the networks without reducing network performance. Network pruning can reduce redundant and unnecessary parameters from a network. Knowledge distillation can transfer the knowledge of deeper and wider networks to smaller networks. The performance of the smaller network obtained by these methods is bounded by the predefined network. Neural architecture search has been proposed, which can search automatically the architecture of the networks to break the structure limitation. Also, there is a dynamic configuration method to train networks incrementally as sub-networks. In this paper, we present a novel incremental training algorithm for deep neural networks called planting. Our planting can search the optimal network architecture with smaller number of parameters for improving the network performance by augmenting channels incrementally to layers of the initial networks while keeping the earlier trained parameters fixed. Also, we propose using the knowledge distillation method for training the channels planted. By transferring the knowledge of deeper and wider networks, we can grow the networks effectively and efficiently. We evaluate the effectiveness of the proposed method on different datasets such as CIFAR-10/100 and STL-10. For the STL-10 dataset, we show that we are able to achieve comparable performance with only 7% parameters compared to the larger network and reduce the overfitting caused by a small amount of the data.

Compression of YOLOv3 Via Block-Wise and Channel-Wise Pruning for Real-Time and Complicated Autonomous Driving Environment Sensing Applications

Jiaqi Li, Yanan Zhao, Li Gao, Feng Cui

Auto-TLDR; Pruning YOLOv3 with Batch Normalization for Autonomous Driving

Abstract Slides Poster Similar

Nowadays, in the area of autonomous driving, the computational power of the object detectors is limited by the embedded devices and the public datasets for autonomous driving are over-idealistic. In this paper, we propose a pipeline combining both block-wise pruning and channel-wise pruning to compress the object detection model iteratively. We enforce the introduced factor of the residual blocks and the scale parameters in Batch Normalization (BN) layers to sparsity to select the less important residual blocks and channels. Moreover, a modified loss function has been proposed to remedy the class-imbalance problem. After removing the unimportant structures iteratively, we get the pruned YOLOv3 trained on our datasets which have more abundant and elaborate classes. Evaluated by our validation sets on the server, the pruned YOLOv3 saves 79.7% floating point operations (FLOPs), 93.8% parameter size, 93.8% model volume and 45.4% inference times with only 4.16% mean of average precision (mAP) loss. Evaluated on the embedded device, the pruned model operates about 13 frames per second with 4.53% mAP loss. These results show that the real-time property and accuracy of the pruned YOLOv3 can meet the needs of the embedded devices in complicated autonomous driving environments.

Dynamic Multi-Path Neural Network

Yingcheng Su, Yichao Wu, Ken Chen, Ding Liang, Xiaolin Hu

Auto-TLDR; Dynamic Multi-path Neural Network

Abstract Slides Similar

Although deeper and larger neural networks have achieved better performance, due to overwhelming burden on computation, they cannot meet the demands of deployment on resource-limited devices. An effective strategy to address this problem is to make use of dynamic inference mechanism, which changes the inference path for different samples at runtime. Existing methods only reduce the depth by skipping an entire specific layer, which may lose important information in this layer. In this paper, we propose a novel method called Dynamic Multi-path Neural Network (DMNN), which provides more topology choices in terms of both width and depth on the fly. For better modelling the inference path selection, we further introduce previous state and object category information to guide the training process. Compared to previous dynamic inference techniques, the proposed method is more flexible and easier to incorporate into most modern network architectures. Experimental results on ImageNet and CIFAR-100 demonstrate the superiority of our method on both efficiency and classification accuracy.

Revisiting the Training of Very Deep Neural Networks without Skip Connections

Oyebade Kayode Oyedotun, Abd El Rahman Shabayek, Djamila Aouada, Bjorn Ottersten

Auto-TLDR; Optimization of Very Deep PlainNets without shortcut connections with 'vanishing and exploding units' activations'

Abstract Slides Poster Similar

Deep neural networks (DNNs) with many layers of feature representations yield state-of-the-art results on several difficult learning tasks. However, optimizing very deep DNNs without shortcut connections known as PlainNets, is a notoriously hard problem. Considering the growing interest in this area, this paper investigates holistically two scenarios that plague the training of very deep PlainNets: (1) the relatively popular challenge of 'vanishing and exploding units' activations', and (2) the less investigated 'singularity' problem, which is studied in details in this paper. In contrast to earlier works that study only the saturation and explosion of units' activations in isolation, this paper harmonizes the inconspicuous coexistence of the aforementioned problems for very deep PlainNets. Particularly, we argue that the aforementioned problems would have to be tackled simultaneously for the successful training of very deep PlainNets. Finally, different techniques that can be employed for tackling the optimization problem are discussed, and a specific combination of simple techniques that allows the successful training of PlainNets having up to 100 layers is demonstrated.

Norm Loss: An Efficient yet Effective Regularization Method for Deep Neural Networks

Theodoros Georgiou, Sebastian Schmitt, Thomas Baeck, Wei Chen, Michael Lew

Auto-TLDR; Weight Soft-Regularization with Oblique Manifold for Convolutional Neural Network Training

Abstract Slides Poster Similar

Convolutional neural network training can suffer from diverse issues like exploding or vanishing gradients, scaling-based weight space symmetry and covariant-shift. In order to address these issues, researchers develop weight regularization methods and activation normalization methods. In this work we propose a weight soft-regularization method based on the Oblique manifold. The proposed method uses a loss function which pushes each weight vector to have a norm close to one, i.e. the weight matrix is smoothly steered toward the so-called Oblique manifold. We evaluate our method on the very popular CIFAR-10, CIFAR-100 and ImageNet 2012 datasets using two state-of-the-art architectures, namely the ResNet and wide-ResNet. Our method introduces negligible computational overhead and the results show that it is competitive to the state-of-the-art and in some cases superior to it. Additionally, the results are less sensitive to hyperparameter settings such as batch size and regularization factor.

Towards Low-Bit Quantization of Deep Neural Networks with Limited Data

Yong Yuan, Chen Chen, Xiyuan Hu, Silong Peng

Auto-TLDR; Low-Precision Quantization of Deep Neural Networks with Limited Data

Abstract Slides Poster Similar

Recent machine learning methods use increasingly large deep neural networks to achieve state-of-the-art results in various tasks. Network quantization can effectively reduce computation and memory costs without modifying network structures, facilitating the deployment of deep neural networks (DNNs) on cloud and edge devices. However, most of the existing methods usually need time-consuming training or fine-tuning and access to the original training dataset that may be unavailable due to privacy or security concerns. In this paper, we present a novel method to achieve low-precision quantization of deep neural networks with limited data. Firstly, to reduce the complexity of per-channel quantization and the degeneration of per-layer quantization, we introduce group-wise quantization which separates the output channels into groups that each group is quantized separately. Secondly, to better distill knowledge from the pre-trained FP32 model with limited data, we introduce a two-stage knowledge distillation method that divides the optimization process into independent optimization stage and joint optimization stage to address the limitation of layer-wise supervision and global supervision. Extensive experiments on ImageNet2012 (ResNet18/50, ShuffleNetV2, and MobileNetV2) demonstrate that the proposed approach can significantly improve the quantization model's accuracy when only a few training samples are available. We further show that the method also extends to other computer vision architectures and tasks such as object detection and semantic segmentation.

Knowledge Distillation Beyond Model Compression

Fahad Sarfraz, Elahe Arani, Bahram Zonooz

Auto-TLDR; Knowledge Distillation from Teacher to Student

Abstract Slides Poster Similar

Knowledge distillation (KD) is commonly deemed as an effective model compression technique in which a compact model (student) is trained under the supervision of a larger pretrained model or an ensemble of models (teacher). Various techniques have been proposed since the original formulation, which mimics different aspects of the teacher such as the representation space, decision boundary or intra-data relationship. Some methods replace the one way knowledge distillation from a static teacher with collaborative learning between a cohort of students. Despite the recent advances, a clear understanding of where knowledge resides in a deep neural network and optimal method for capturing knowledge from teacher and transferring it to student still remains an open question. In this study we provide an extensive study on 9 different knowledge distillation methods which covers a broad spectrum of approaches to capture and transfer knowledge. We demonstrate the versatility of the KD framework on different datasets and network architectures under varying capacity gaps between the teacher and student. The study provides intuition for the effects of mimicking different aspects of the teacher and derives insights from the performance of the different distillation approaches to guide the the design of more effective KD methods . Furthermore, our study shows the effectiveness of the KD framework in learning efficiently under varying severity levels of label noise and class imbalance, consistently providing significant generalization gains over standard training. We emphasize that the efficacy of KD goes much beyond a model compression technique and should be considered as a general purpose training paradigm which offers more robustness to common challenges in the real-world datasets compared to the standard training procedure.

FastSal: A Computationally Efficient Network for Visual Saliency Prediction

Feiyan Hu, Kevin Mcguinness

Auto-TLDR; MobileNetV2: A Convolutional Neural Network for Saliency Prediction

Abstract Slides Poster Similar

This paper focuses on the problem of visual saliency prediction, predicting regions of an image that tend to attract human visual attention, under a constrained computational budget. We modify and test various recent efficient convolutional neural network architectures like EfficientNet and MobileNetV2 and compare them with existing state-of-the-art saliency models such as SalGAN and DeepGaze II both in terms of standard accuracy metrics like AUC and NSS, and in terms of the computational complexity and model size. We find that MobileNetV2 makes an excellent backbone for a visual saliency model and can be effective even without a complex decoder. We also show that knowledge transfer from a more computationally expensive model like DeepGaze II can be achieved via pseudo-labelling an unlabelled dataset, and that this approach gives result on-par with many state-of-the-art algorithms with a fraction of the computational cost and model size.

Hcore-Init: Neural Network Initialization Based on Graph Degeneracy

Stratis Limnios, George Dasoulas, Dimitrios Thilikos, Michalis Vazirgiannis

Auto-TLDR; K-hypercore: Graph Mining for Deep Neural Networks

Abstract Slides Poster Similar

Neural networks are the pinnacle of Artificial Intelligence, as in recent years we witnessed many novel architectures, learning and optimization techniques for deep learning. Capitalizing on the fact that neural networks inherently constitute multipartite graphs among neuron layers, we aim to analyze directly their structure to extract meaningful information that can improve the learning process. To our knowledge graph mining techniques for enhancing learning in neural networks have not been thoroughly investigated. In this paper we propose an adapted version of the k-core structure for the complete weighted multipartite graph extracted from a deep learning architecture. As a multipartite graph is a combination of bipartite graphs, that are in turn the incidence graphs of hypergraphs, we design k-hypercore decomposition, the hypergraph analogue of k-core degeneracy. We applied k-hypercore to several neural network architectures, more specifically to convolutional neural networks and multilayer perceptrons for image recognition tasks after a very short pretraining. Then we used the information provided by the hypercore numbers of the neurons to re-initialize the weights of the neural network, thus biasing the gradient optimization scheme. Extensive experiments proved that k-hypercore outperforms the state-of-the-art initialization methods.

Distilling Spikes: Knowledge Distillation in Spiking Neural Networks

Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee, Rajbabu Velmurugan

Auto-TLDR; Knowledge Distillation in Spiking Neural Networks for Image Classification

Abstract Slides Poster Similar

Spiking Neural Networks (SNN) are energy-efficient computing architectures that exchange spikes for processing information, unlike classical Artificial Neural Networks (ANN). Due to this, SNNs are better suited for real-life deployments. However, similar to ANNs, SNNs also benefit from deeper architectures to obtain improved performance. Furthermore, like the deep ANNs, the memory, compute and power requirements of SNNs also increase with model size, and model compression becomes a necessity. Knowledge distillation is a model com- pression technique that enables transferring the learning of a large machine learning model to a smaller model with minimal loss in performance. In this paper, we propose techniques for knowledge distillation in spiking neural networks for the task of image classification. We present ways to distill spikes from a larger SNN, also called the teacher network, to a smaller one, also called the student network, while minimally impacting the classification accuracy. We demonstrate the effectiveness of the proposed method with detailed experiments on three standard datasets while proposing novel distillation methodologies and loss functions. We also present a multi-stage knowledge distillation technique for SNNs using an intermediate network to obtain higher performance from the student network. Our approach is expected to open up new avenues for deploying high performing large SNN models on resource-constrained hardware platforms.

CQNN: Convolutional Quadratic Neural Networks

Pranav Mantini, Shishir Shah

Auto-TLDR; Quadratic Neural Network for Image Classification

Abstract Slides Poster Similar

Image classification is a fundamental task in computer vision. A variety of deep learning models based on the Convolutional Neural Network (CNN) architecture have proven to be an efficient solution. Numerous improvements have been proposed over the years, where broader, deeper, and denser networks have been constructed. However, the atomic operation for these models has remained a linear unit (single neuron). In this work, we pursue an alternative dimension by hypothesizing the atomic operation to be performed by a quadratic unit. We construct convolutional layers using quadratic neurons for feature extraction and subsequently use dense layers for classification. We perform analysis to quantify the implication of replacing linear neurons with quadratic units. Results show a keen improvement in classification accuracy with quadratic neurons over linear neurons.

Generalization Comparison of Deep Neural Networks Via Output Sensitivity

Mahsa Forouzesh, Farnood Salehi, Patrick Thiran

Auto-TLDR; Generalization of Deep Neural Networks using Sensitivity

Abstract Slides Similar

Although recent works have brought some insights into the performance improvement of techniques used in state-of-the-art deep-learning models, more work is needed to understand their generalization properties. We shed light on this matter by linking the loss function to the output's sensitivity to its input. We find a rather strong empirical relation between the output sensitivity and the variance in the bias-variance decomposition of the loss function, which hints on using sensitivity as a metric for comparing the generalization performance of networks, without requiring labeled data. We find that sensitivity is decreased by applying popular methods which improve the generalization performance of the model, such as (1) using a deep network rather than a wide one, (2) adding convolutional layers to baseline classifiers instead of adding fully-connected layers, (3) using batch normalization, dropout and max-pooling, and (4) applying parameter initialization techniques.

How Does DCNN Make Decisions?

Yi Lin, Namin Wang, Xiaoqing Ma, Ziwei Li, Gang Bai

Auto-TLDR; Exploring Deep Convolutional Neural Network's Decision-Making Interpretability

Abstract Slides Poster Similar

Deep Convolutional Neural Networks (DCNN), despite imitating the human visual system, present no such decision credibility as human observers. This phenomenon, therefore, leads to the limitations of DCNN's applications in the security and trusted computing, such as self-driving cars and medical diagnosis. Focusing on this issue, our work aims to explore the way DCNN makes decisions. In this paper, the major contributions we made are: firstly, provide the hypothesis, “point-wise activation” of convolution function, according to the analysis of DCNN’s architectures and training process; secondly, point out the effect of “point-wise activation” on DCNN’s uninterpretable classification and pool robustness, and then suggest, in particular, the contradiction between the traditional and DCNN’s convolution kernel functions; finally, distinguish decision-making interpretability from semantic interpretability, and indicate that DCNN’s decision-making mechanism need to evolve towards the direction of semantics in the future. Besides, the “point-wise activation” hypothesis and conclusions proposed in our paper are supported by extensive experimental results.

WeightAlign: Normalizing Activations by Weight Alignment

Xiangwei Shi, Yunqiang Li, Xin Liu, Jan Van Gemert

Auto-TLDR; WeightAlign: Normalization of Activations without Sample Statistics

Abstract Slides Poster Similar

Batch normalization (BN) allows training very deep networks by normalizing activations by mini-batch sample statistics which renders BN unstable for small batch sizes. Current small-batch solutions such as Instance Norm, Layer Norm, and Group Norm use channel statistics which can be computed even for a single sample. Such methods are less stable than BN as they critically depend on the statistics of a single input sample. To address this problem, we propose a normalization of activation without sample statistics. We present WeightAlign: a method that normalizes the weights by the mean and scaled standard derivation computed within a filter, which normalizes activations without computing any sample statistics. Our proposed method is independent of batch size and stable over a wide range of batch sizes. Because weight statistics are orthogonal to sample statistics, we can directly combine WeightAlign with any method for activation normalization. We experimentally demonstrate these benefits for classification on CIFAR-10, CIFAR-100, ImageNet, for semantic segmentation on PASCAL VOC 2012 and for domain adaptation on Office-31.

Nearest Neighbor Classification Based on Activation Space of Convolutional Neural Network

Xinbo Ju, Shuo Shao, Huan Long, Weizhe Wang

Auto-TLDR; Convolutional Neural Network with Convex Hull Based Classifier

Abstract Poster Similar

In this paper, we propose a new image classifier based on the incorporation of the nearest neighbor algorithm and the activation space of convolutional neural network. The classifier has been successfully used on some state-of-the-art models and further improve their performance. Main technique tools we used are convex hull based classification and its acceleration. We find that 1) in several cases, the classifier can reach higher accuracy than original CNN; 2) by sampling, the classifier can work more efficiently; 3) centroid of each convex hull shows surprising ability in classification. Most of the work has strong geometry meanings, which helps us have a new understanding about convolutional layers.

Regularized Flexible Activation Function Combinations for Deep Neural Networks

Renlong Jie, Junbin Gao, Andrey Vasnev, Minh-Ngoc Tran

Auto-TLDR; Flexible Activation in Deep Neural Networks using ReLU and ELUs

Abstract Slides Poster Similar

Activation in deep neural networks is fundamental to achieving non-linear mappings. Traditional studies mainly focus on ﬁnding ﬁxed activations for a particular set of learning tasks or model architectures. The research on ﬂexible activation is quite limited in both designing philosophy and application scenarios. In this study, three principles of choosing ﬂexible activation components are proposed and a general combined form of ﬂexible activation functions is implemented. Based on this, a novel family of ﬂexible activation functions that can replace sigmoid or tanh in LSTM cells are implemented, as well as a new family by combining ReLU and ELUs. Also, two new regularisation terms based on assumptions as prior knowledge are introduced. It has been shown that LSTM models with proposed ﬂexible activations P-Sig-Ramp provide signiﬁcant improvements in time series forecasting, while the proposed P-E2-ReLU achieves better and more stable performance on lossy image compression tasks with convolutional auto-encoders. In addition, the proposed regularization terms improve the convergence,performance and stability of the models with ﬂexible activation functions. The code for this paper is available at https://github.com/9NXJRDDRQK/Flexible Activation.

Not All Domains Are Equally Complex: Adaptive Multi-Domain Learning

Ali Senhaji, Jenni Karoliina Raitoharju, Moncef Gabbouj, Alexandros Iosifidis

Auto-TLDR; Adaptive Parameterization for Multi-Domain Learning

Abstract Slides Poster Similar

Deep learning approaches are highly specialized and require training separate models for different tasks. Multi-domain learning looks at ways to learn a multitude of different tasks, each coming from a different domain, at once. The most common approach in multi-domain learning is to form a domain agnostic model, the parameters of which are shared among all domains, and learn a small number of extra domain-specific parameters for each individual new domain. However, different domains come with different levels of difficulty; parameterizing the models of all domains using an augmented version of the domain agnostic model leads to unnecessarily inefficient solutions, especially for easy to solve tasks. We propose an adaptive parameterization approach to deep neural networks for multi-domain learning. The proposed approach performs on par with the original approach while reducing by far the number of parameters, leading to efficient multi-domain learning solutions.

Self-Supervised Learning for Astronomical Image Classification

Ana Martinazzo, Mateus Espadoto, Nina S. T. Hirata

Auto-TLDR; Unlabeled Astronomical Images for Deep Neural Network Pre-training

Abstract Slides Poster Similar

In Astronomy, a huge amount of image data is generated daily by photometric surveys, which scan the sky to collect data from stars, galaxies and other celestial objects. In this paper, we propose a technique to leverage unlabeled astronomical images to pre-train deep convolutional neural networks, in order to learn a domain-specific feature extractor which improves the results of machine learning techniques in setups with small amounts of labeled data available. We show that our technique produces results which are in many cases better than using ImageNet pre-training.

Delving in the Loss Landscape to Embed Robust Watermarks into Neural Networks

Enzo Tartaglione, Marco Grangetto, Davide Cavagnino, Marco Botta

Auto-TLDR; Watermark Aware Training of Neural Networks

Abstract Slides Poster Similar

In the last decade the use of artificial neural networks (ANNs) in many fields like image processing or speech recognition has become a common practice because of their effectiveness to solve complex tasks. However, in such a rush, very little attention has been paid to security aspects. In this work we explore the possibility to embed a watermark into the ANN parameters. We exploit model redundancy and adaptation capacity to lock a subset of its parameters to carry the watermark sequence. The watermark can be extracted in a simple way to claim copyright on models but can be very easily attacked with model fine-tuning. To tackle this culprit we devise a novel watermark aware training strategy. We aim at delving into the loss landscape to find an optimal configuration of the parameters such that we are robust to fine-tuning attacks towards the watermarked parameters. Our experimental results on classical ANN models trained on well-known MNIST and CIFAR-10 datasets show that the proposed approach makes the embedded watermark robust to fine-tuning and compression attacks.

E-DNAS: Differentiable Neural Architecture Search for Embedded Systems

Javier García López, Antonio Agudo, Francesc Moreno-Noguer

Auto-TLDR; E-DNAS: Differentiable Architecture Search for Light-Weight Networks for Image Classification

Abstract Slides Poster Similar

Designing optimal and light weight networks to fit in resource-limited platforms like mobiles, DSPs or GPUs is a challenging problem with a wide range of interesting applications, {\em e.g.} in embedded systems for autonomous driving. While most approaches are based on manual hyperparameter tuning, there exist a new line of research, the so-called NAS (Neural Architecture Search) methods, that aim to optimize several metrics during the design process, including memory requirements of the network, number of FLOPs, number of MACs (Multiply-ACcumulate operations) or inference latency. However, while NAS methods have shown very promising results, they are still significantly time and cost consuming. In this work we introduce E-DNAS, a differentiable architecture search method, which improves the efficiency of NAS methods in designing light-weight networks for the task of image classification. Concretely, E-DNAS computes, in a differentiable manner, the optimal size of a number of meta-kernels that capture patterns of the input data at different resolutions. We also leverage on the additive property of convolution operations to merge several kernels with different compatible sizes into a single one, reducing thus the number of operations and the time required to estimate the optimal configuration. We evaluate our approach on several datasets to perform classification. We report results in terms of the SoC (System on Chips) metric, typically used in the Texas Instruments TDA2x families for autonomous driving applications. The results show that our approach allows designing low latency architectures significantly faster than state-of-the-art.

Resource-efficient DNNs for Keyword Spotting using Neural Architecture Search and Quantization

David Peter, Wolfgang Roth, Franz Pernkopf

Auto-TLDR; Neural Architecture Search for Keyword Spotting in Limited Resource Environments

Abstract Slides Poster Similar

This paper introduces neural architecture search (NAS) for the automatic discovery of small models for keyword spotting (KWS) in limited resource environments. We employ a differentiable NAS approach to optimize the structure of convolutional neural networks (CNNs) to meet certain memory constraints for storing weights as well as constraints on the number of operations per inference. Using NAS only, we were able to obtain a highly efficient model with 95.6% accuracy on the Google speech commands dataset with 494.8 kB of memory usage and 19.6 million operations. Additionally, weight quantization is used to reduce the memory consumption even further. We show that weight quantization to low bit-widths (e.g. 1 bit) can be used without substantial loss in accuracy. By increasing the number of input features from 10 MFCC to 20 MFCC we were able to increase the accuracy to 96.6% at 340.1 kB of memory usage and 27.1 million operations.

Adaptive Noise Injection for Training Stochastic Student Networks from Deterministic Teachers

Yi Xiang Marcus Tan, Yuval Elovici, Alexander Binder

Auto-TLDR; Adaptive Stochastic Networks for Adversarial Attacks

Abstract Slides Similar

Adversarial attacks have been a prevalent problem causing misclassification in machine learning models, with stochasticity being a promising direction towards greater robustness. However, stochastic networks frequently underperform compared to deterministic deep networks. In this work, we present a conceptually clear adaptive noise injection mechanism in combination with teacher-initialisation, which adjusts its degree of randomness dynamically through the computation of mini-batch statistics. This mechanism is embedded within a simple framework to obtain stochastic networks from existing deterministic networks. Our experiments show that our method is able to outperform prior baselines under white-box settings, exemplified through CIFAR-10 and CIFAR-100. Following which, we perform in-depth analysis on varying different components of training with our approach on the effects of robustness and accuracy, through the study of the evolution of decision boundary and trend curves of clean accuracy/attack success over differing degrees of stochasticity. We also shed light on the effects of adversarial training on a pre-trained network, through the lens of decision boundaries.

P-DIFF: Learning Classifier with Noisy Labels Based on Probability Difference Distributions

Wei Hu, Qihao Zhao, Yangyu Huang, Fan Zhang

Auto-TLDR; P-DIFF: A Simple and Effective Training Paradigm for Deep Neural Network Classifier with Noisy Labels

Abstract Slides Poster Similar

Learning deep neural network (DNN) classifier with noisy labels is a challenging task because the DNN can easily over- fit on these noisy labels due to its high capability. In this paper, we present a very simple but effective training paradigm called P-DIFF, which can train DNN classifiers but obviously alleviate the adverse impact of noisy labels. Our proposed probability difference distribution implicitly reflects the probability of a training sample to be clean, then this probability is employed to re-weight the corresponding sample during the training process. P-DIFF can also achieve good performance even without prior- knowledge on the noise rate of training samples. Experiments on benchmark datasets also demonstrate that P-DIFF is superior to the state-of-the-art sample selection methods.

VPU Specific CNNs through Neural Architecture Search

Ciarán Donegan, Hamza Yous, Saksham Sinha, Jonathan Byrne

Auto-TLDR; Efficient Convolutional Neural Networks for Edge Devices using Neural Architecture Search

Abstract Slides Poster Similar

The success of deep learning at computer vision tasks has led to an ever-increasing number of applications on edge devices. Often with the use of edge AI hardware accelerators like the Intel Movidius Vision Processing Unit (VPU). Performing computer vision tasks on edge devices is challenging. Many Convolutional Neural Networks (CNNs) are too complex to run on edge devices with limited computing power. This has created large interest in designing efficient CNNs and one promising way of doing this is through Neural Architecture Search (NAS). NAS aims to automate the design of neural networks. NAS can also optimize multiple different objectives together, like accuracy and efficiency, which is difficult for humans. In this paper, we use a differentiable NAS method to find efficient CNNs for VPU that achieves state-of-the-art classification accuracy on ImageNet. Our NAS designed model outperforms MobileNetV2, having almost 1\% higher top-1 accuracy while being 13\% faster on MyriadX VPU. To the best of our knowledge, this is the first time a VPU specific CNN has been designed using a NAS algorithm. Our results also reiterate the fact that efficient networks must be designed for each specific hardware. We show that efficient networks targeted at different devices do not perform as well on the VPU.

Trainable Spectrally Initializable Matrix Transformations in Convolutional Neural Networks

Michele Alberti, Angela Botros, Schuetz Narayan, Rolf Ingold, Marcus Liwicki, Mathias Seuret

Auto-TLDR; Trainable and Spectrally Initializable Matrix Transformations for Neural Networks

Abstract Slides Poster Similar

In this work, we introduce a new architectural component to Neural Networks (NN), i.e., trainable and spectrally initializable matrix transformations on feature maps. While previous literature has already demonstrated the possibility of adding static spectral transformations as feature processors, our focus is on more general trainable transforms. We study the transforms in various architectural configurations on four datasets of different nature: from medical (ColorectalHist, HAM10000) and natural (Flowers) images to historical documents (CB55). With rigorous experiments that control for the number of parameters and randomness, we show that networks utilizing the introduced matrix transformations outperform vanilla neural networks. The observed accuracy increases appreciably across all datasets. In addition, we show that the benefit of spectral initialization leads to significantly faster convergence, as opposed to randomly initialized matrix transformations. The transformations are implemented as auto-differentiable PyTorch modules that can be incorporated into any neural network architecture. The entire code base is open-source.

Learning Stable Deep Predictive Coding Networks with Weight Norm Supervision

Guo Ruohao

Auto-TLDR; Stability of Predictive Coding Network with Weight Norm Supervision

Abstract Slides Poster Similar

Predictive Coding Network (PCN) is an important neural network inspired by visual processing models in neuroscience. It combines the feedforward and feedback processing and has the architecture of recurrent neural networks (RNNs). This type of network is usually trained with backpropagation through time (BPTT). With infinite recurrent steps, PCN is a dynamic system. However, as one of the most important properties, stability is rarely studied in this type of network. Inspired by reservoir computing, we investigate the stability of hierarchical RNNs from the perspective of dynamic systems, and propose a sufficient condition for their echo state property (ESP). Our study shows the global stability is determined by stability of the local layers and the feedback between neighboring layers. Based on it, we further propose Weight Norm Supervision, a new algorithm that controls the stability of PCN dynamics by imposing different weight norm constraints on different parts of the network. We compare our approach with other training methods in terms of stability and prediction capability. The experiments show that our algorithm learns stable PCNs with a reliable prediction precision in the most effective and controllable way.

Stage-Wise Neural Architecture Search

Artur Jordão, Fernando Akio Yamada, Maiko Lie, William Schwartz

Auto-TLDR; Efficient Neural Architecture Search for Deep Convolutional Networks

Abstract Slides Poster Similar

Modern convolutional networks such as ResNet and NASNet have achieved state-of-the-art results in many computer vision applications. These architectures consist of stages, which are sets of layers that operate on representations in the same resolution. It has been demonstrated that increasing the number of layers in each stage improves the prediction ability of the network. However, the resulting architecture becomes computationally expensive in terms of floating point operations, memory requirements and inference time. Thus, significant human effort is necessary to evaluate different trade-offs between depth and performance. To handle this problem, recent works have proposed to automatically design high-performance architectures, mainly by means of neural architecture search (NAS). Current NAS strategies analyze a large set of possible candidate architectures and, hence, require vast computational resources and take many GPUs days. Motivated by this, we propose a NAS approach to efficiently design accurate and low-cost convolutional architectures and demonstrate that an efficient strategy for designing these architectures is to learn the depth stage-by-stage. For this purpose, our approach increases depth incrementally in each stage taking into account its importance, such that stages with low importance are kept shallow while stages with high importance become deeper. We conduct experiments on the CIFAR and different versions of ImageNet datasets, where we show that architectures discovered by our approach achieve better accuracy and efficiency than human-designed architectures. Additionally, we show that architectures discovered on CIFAR-10 can be successfully transferred to large datasets. Compared to previous NAS approaches, our method is substantially more efficient, as it evaluates one order of magnitude fewer models and yields architectures on par with the state-of-the-art.

Locality-Promoting Representation Learning

Johannes Schneider

Auto-TLDR; Locality-promoting Regularization for Neural Networks

Abstract Slides Poster Similar

This work investigates questions related to learning features in convolutional neural networks (CNN). Empirical findings across multiple architectures such as VGG, ResNet, Inception and MobileNet indicate that weights near the center of a filter are larger than weights on the outside. Current regularization schemes violate this principle. Thus, we introduce Locality-promoting Regularization, which yields accuracy gains across multiple architectures and datasets. We also show theoretically that the empirical finding could be explained by maximizing feature cohesion under the assumption of spatial locality.

ResNet-Like Architecture with Low Hardware Requirements

Elena Limonova, Daniil Alfonso, Dmitry Nikolaev, Vladimir V. Arlazarov

Auto-TLDR; BM-ResNet: Bipolar Morphological ResNet for Image Classification

Abstract Slides Poster Similar

One of the most computationally intensive parts in modern recognition systems is an inference of deep neural networks that are used for image classification, segmentation, enhancement, and recognition. The growing popularity of edge computing makes us look for ways to reduce its time for mobile and embedded devices. One way to decrease the neural network inference time is to modify a neuron model to make it more efficient for computations on a specific device. The example of such a model is a bipolar morphological neuron model. The bipolar morphological neuron is based on the idea of replacing multiplication with addition and maximum operations. This model has been demonstrated for simple image classification with LeNet-like architectures [1]. In the paper, we introduce a bipolar morphological ResNet (BM-ResNet) model obtained from a much more complex ResNet architecture by converting its layers to bipolar morphological ones. We apply BM-ResNet to image classification on MNIST and CIFAR-10 datasets with only a moderate accuracy decrease from 99.3% to 99.1% and from 85.3% to 85.1%. We also estimate the computational complexity of the resulting model. We show that for the majority of ResNet layers, the considered model requires 2.1-2.9 times fewer logic gates for implementation and 15-30% lower latency.