Knowledge Distillation Beyond Model Compression

Fahad Sarfraz, Elahe Arani, Bahram Zonooz

Responsive image

Auto-TLDR; Knowledge Distillation from Teacher to Student

Slides Poster

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.

Similar papers

Efficient Online Subclass Knowledge Distillation for Image Classification

Maria Tzelepi, Nikolaos Passalis, Anastasios Tefas

Responsive image

Auto-TLDR; OSKD: Online Subclass Knowledge Distillation

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.

Feature Fusion for Online Mutual Knowledge Distillation

Jangho Kim, Minsung Hyun, Inseop Chung, Nojun Kwak

Responsive image

Auto-TLDR; Feature Fusion Learning Using Fusion of Sub-Networks

Slides Poster Similar

We propose a learning framework named Feature Fusion Learning (FFL) that efficiently trains a powerful classifier through a fusion module which combines the feature maps generated from parallel neural networks and generates meaningful feature maps. Specifically, we train a number of parallel neural networks as sub-networks, then we combine the feature maps from each sub-network using a fusion module to create a more meaningful feature map. The fused feature map is passed into the fused classifier for overall classification. Unlike existing feature fusion methods, in our framework, an ensemble of sub-network classifiers transfers its knowledge to the fused classifier and then the fused classifier delivers its knowledge back to each sub-network, mutually teaching one another in an online-knowledge distillation manner. This mutually teaching system not only improves the performance of the fused classifier but also obtains performance gain in each sub-network. Moreover, our model is more beneficial than other alternative methods because different types of network can be used for each sub-network. We have performed a variety of experiments on multiple datasets such as CIFAR-10, CIFAR-100 and ImageNet and proved that our method is more effective than other alternative methods in terms of performances of both sub-networks and the fused classifier, and the aspect of generating meaningful feature maps.

Adaptive Noise Injection for Training Stochastic Student Networks from Deterministic Teachers

Yi Xiang Marcus Tan, Yuval Elovici, Alexander Binder

Responsive image

Auto-TLDR; Adaptive Stochastic Networks for Adversarial Attacks

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.

Knowledge Distillation with a Precise Teacher and Prediction with Abstention

Xu Yi, Jian Pu, Hui Zhao

Responsive image

Auto-TLDR; Knowledge Distillation using Deep gambler loss and selective classification framework

Slides Poster Similar

Knowledge distillation, which aims to train model under the supervision from another large model (teacher model) to the original model (student model), has achieved remarkable results in supervised learning. However, there are two major problems with existing knowledge distillation methods. One is the teacher's supervision is sometimes misleading, and the other is the student's prediction is not accurate enough. To address the first issue, instead of learning a combination of both teachers and ground truth, we apply knowledge adjustment to correct teachers' supervision using ground truth. For the second problem, we use the selective classification framework to train the student model. In particular, the deep gambler loss is adopted to predict with reservation by explicitly introducing the $(m+1)$-th class. We consider two settings of knowledge distillation: (1) distillation across different network structures ({\it AlexNet, ResNet}), and (2) distillation across networks with different depths ({\it ResNet18, ResNet50}) to evaluate the effectiveness of our method. The experimental results on benchmark datasets (i.e., {\it Fashion-MNIST, SVHN, CIFAR10, CIFAR100}) are reported with higher prediction accuracies and lower coverage errors.

Distilling Spikes: Knowledge Distillation in Spiking Neural Networks

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

Responsive image

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

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.

Compact CNN Structure Learning by Knowledge Distillation

Waqar Ahmed, Andrea Zunino, Pietro Morerio, Vittorio Murino

Responsive image

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

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.

Local Clustering with Mean Teacher for Semi-Supervised Learning

Zexi Chen, Benjamin Dutton, Bharathkumar Ramachandra, Tianfu Wu, Ranga Raju Vatsavai

Responsive image

Auto-TLDR; Local Clustering for Semi-supervised Learning

Slides Similar

The Mean Teacher (MT) model of Tarvainen and Valpola has shown favorable performance on several semi-supervised benchmark datasets. MT maintains a teacher model's weights as the exponential moving average of a student model's weights and minimizes the divergence between their probability predictions under diverse perturbations of the inputs. However, MT is known to suffer from confirmation bias, that is, reinforcing incorrect teacher model predictions. In this work, we propose a simple yet effective method called Local Clustering (LC) to mitigate the effect of confirmation bias. In MT, each data point is considered independent of other points during training; however, data points are likely to be close to each other in feature space if they share similar features. Motivated by this, we cluster data points locally by minimizing the pairwise distance between neighboring data points in feature space. Combined with a standard classification cross-entropy objective on labeled data points, the misclassified unlabeled data points are pulled towards high-density regions of their correct class with the help of their neighbors, thus improving model performance. We demonstrate on semi-supervised benchmark datasets SVHN and CIFAR-10 that adding our LC loss to MT yields significant improvements compared to MT and performance comparable to the state of the art in semi-supervised learning.

Stochastic Label Refinery: Toward Better Target Label Distribution

Xi Fang, Jiancheng Yang, Bingbing Ni

Responsive image

Auto-TLDR; Stochastic Label Refinery for Deep Supervised Learning

Slides Poster Similar

This paper proposes a simple yet effective strategy for improving deep supervised learning, named Stochastic Label Refinery (SLR), by refining training labels to more informative labels. When training a neural network, target distributions (or ground-truth) are typically "hard", which means the target label of each category consists of only 0 and 1. However, the fixed "hard" target distributions do not capture association between categories or that between objects. In this study, instead of using the hard target distributions, we iteratively generate "soft" target label distributions for training the neural networks, which leads to better performances. The soft target distributions are obtained via an Expectation-Maximization (EM) iteration, where the "true" target distributions and the learned models are regarded as hidden variables. In E step, the models are optimized to approximate the target distributions on stochastic splits of training data; In M step, the target distributions are updated with predicted pseudo-label on leave-out splits. Extensive experiments on classification and ordinal regression tasks, empirically prove that the refined target distribution consistently leads to considerable performance improvements even applied on competitive baselines. Notably, in DeepDR 2020 Diabetic Retinopathy Grading (DeepDRiD) challenge, our method improves the quadratic weighted kappa on official validation set from 0.8247 to 0.8348 and achieves a state-of-the-art score on online test set. The proposed SLR technique is easy to implement and practically applicable. The code will be open sourced soon.

Towards Robust Learning with Different Label Noise Distributions

Diego Ortego, Eric Arazo, Paul Albert, Noel E O'Connor, Kevin Mcguinness

Responsive image

Auto-TLDR; Distribution Robust Pseudo-Labeling with Semi-supervised Learning

Slides Similar

Noisy labels are an unavoidable consequence of labeling processes and detecting them is an important step towards preventing performance degradations in Convolutional Neural Networks. Discarding noisy labels avoids a harmful memorization, while the associated image content can still be exploited in a semi-supervised learning (SSL) setup. Clean samples are usually identified using the small loss trick, i.e. they exhibit a low loss. However, we show that different noise distributions make the application of this trick less straightforward and propose to continuously relabel all images to reveal a discriminative loss against multiple distributions. SSL is then applied twice, once to improve the clean-noisy detection and again for training the final model. We design an experimental setup based on ImageNet32/64 for better understanding the consequences of representation learning with differing label noise distributions and find that non-uniform out-of-distribution noise better resembles real-world noise and that in most cases intermediate features are not affected by label noise corruption. Experiments in CIFAR-10/100, ImageNet32/64 and WebVision (real-world noise) demonstrate that the proposed label noise Distribution Robust Pseudo-Labeling (DRPL) approach gives substantial improvements over recent state-of-the-art. Code will be made available.

On-Manifold Adversarial Data Augmentation Improves Uncertainty Calibration

Kanil Patel, William Beluch, Dan Zhang, Michael Pfeiffer, Bin Yang

Responsive image

Auto-TLDR; On-Manifold Adversarial Data Augmentation for Uncertainty Estimation

Slides Similar

Uncertainty estimates help to identify ambiguous, novel, or anomalous inputs, but the reliable quantification of uncertainty has proven to be challenging for modern deep networks. To improve uncertainty estimation, we propose On-Manifold Adversarial Data Augmentation or OMADA, which specifically attempts to generate challenging examples by following an on-manifold adversarial attack path in the latent space of an autoencoder that closely approximates the decision boundaries between classes. On a variety of datasets and for multiple network architectures, OMADA consistently yields more accurate and better calibrated classifiers than baseline models, and outperforms competing approaches such as Mixup, as well as achieving similar performance to (at times better than) post-processing calibration methods such as temperature scaling. Variants of OMADA can employ different sampling schemes for ambiguous on-manifold examples based on the entropy of their estimated soft labels, which exhibit specific strengths for generalization, calibration of predicted uncertainty, or detection of out-of-distribution inputs.

Meta Soft Label Generation for Noisy Labels

Görkem Algan, Ilkay Ulusoy

Responsive image

Auto-TLDR; MSLG: Meta-Learning for Noisy Label Generation

Slides Poster Similar

The existence of noisy labels in the dataset causes significant performance degradation for deep neural networks (DNNs). To address this problem, we propose a Meta Soft Label Generation algorithm called MSLG, which can jointly generate soft labels using meta-learning techniques and learn DNN parameters in an end-to-end fashion. Our approach adapts the meta-learning paradigm to estimate optimal label distribution by checking gradient directions on both noisy training data and noise-free meta-data. In order to iteratively update soft labels, meta-gradient descent step is performed on estimated labels, which would minimize the loss of noise-free meta samples. In each iteration, the base classifier is trained on estimated meta labels. MSLG is model-agnostic and can be added on top of any existing model at hand with ease. We performed extensive experiments on CIFAR10, Clothing1M and Food101N datasets. Results show that our approach outperforms other state-of-the-art methods by a large margin. Our code is available at \url{https://github.com/gorkemalgan/MSLG_noisy_label}.

Channel Planting for Deep Neural Networks Using Knowledge Distillation

Kakeru Mitsuno, Yuichiro Nomura, Takio Kurita

Responsive image

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

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.

Is the Meta-Learning Idea Able to Improve the Generalization of Deep Neural Networks on the Standard Supervised Learning?

Xiang Deng, Zhongfei Zhang

Responsive image

Auto-TLDR; Meta-learning Based Training of Deep Neural Networks for Few-Shot Learning

Slides Poster Similar

Substantial efforts have been made on improving the generalization abilities of deep neural networks (DNNs) in order to obtain better performances without introducing more parameters. On the other hand, meta-learning approaches exhibit powerful generalization on new tasks in few-shot learning. Intuitively, few-shot learning is more challenging than the standard supervised learning as each target class only has a very few or no training samples. The natural question that arises is whether the meta-learning idea can be used for improving the generalization of DNNs on the standard supervised learning. In this paper, we propose a novel meta-learning based training procedure (MLTP) for DNNs and demonstrate that the meta-learning idea can indeed improve the generalization abilities of DNNs. MLTP simulates the meta-training process by considering a batch of training samples as a task. The key idea is that the gradient descent step for improving the current task performance should also improve a new task performance, which is ignored by the current standard procedure for training neural networks. MLTP also benefits from all the existing training techniques such as dropout, weight decay, and batch normalization. We evaluate MLTP by training a variety of small and large neural networks on three benchmark datasets, i.e., CIFAR-10, CIFAR-100, and Tiny ImageNet. The experimental results show a consistently improved generalization performance on all the DNNs with different sizes, which verifies the promise of MLTP and demonstrates that the meta-learning idea is indeed able to improve the generalization of DNNs on the standard supervised learning.

A Delayed Elastic-Net Approach for Performing Adversarial Attacks

Brais Cancela, Veronica Bolon-Canedo, Amparo Alonso-Betanzos

Responsive image

Auto-TLDR; Robustness of ImageNet Pretrained Models against Adversarial Attacks

Slides Poster Similar

With the rise of the so-called Adversarial Attacks, there is an increased concern on model security. In this paper we present two different contributions: novel measures of robustness (based on adversarial attacks) and a novel adversarial attack. The key idea behind these metrics is to obtain a measure that could compare different architectures, with independence of how the input is preprocessed (robustness against different input sizes and value ranges). To do so, a novel adversarial attack is presented, performing a delayed elastic-net adversarial attack (constraints are only used whenever a successful adversarial attack is obtained). Experimental results show that our approach obtains state-of-the-art adversarial samples, in terms of minimal perturbation distance. Finally, a benchmark of ImageNet pretrained models is used to conduct experiments aiming to shed some light about which model should be selected whenever security is a role factor.

Automatic Student Network Search for Knowledge Distillation

Zhexi Zhang, Wei Zhu, Junchi Yan, Peng Gao, Guotong Xie

Responsive image

Auto-TLDR; NAS-KD: Knowledge Distillation for BERT

Slides Poster Similar

Pre-trained language models (PLMs), such as BERT, have achieved outstanding performance on multiple natural language processing (NLP) tasks. However, such pre-trained models usually contain a huge number of parameters and are computationally expensive. The high resource demand hinders their application on resource-restricted devices like mobile phones. Knowledge distillation (KD) is an effective compression approach, aiming at encouraging a light-weight student network to imitate the teacher network, and accordingly latent knowledge is transferred from the teacher to student. However, the great majority of student networks in previous KD methods are manually designed, normally a subnetwork of the teacher network. Transformer is generally utilized as the student for compressing BERT but still contains masses of parameters. Motivated by this, we propose a novel approach named NAS-KD, which automatically generates an optimal student network using neural architecture search (NAS) to enhance the distillation for BERT. Experiment on 7 classification tasks in NLP domain demonstrates that NAS-KD can substantially reduce the size of BERT without much performance sacrifice.

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

Wei Hu, Qihao Zhao, Yangyu Huang, Fan Zhang

Responsive image

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

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.

Improving Model Accuracy for Imbalanced Image Classification Tasks by Adding a Final Batch Normalization Layer: An Empirical Study

Veysel Kocaman, Ofer M. Shir, Thomas Baeck

Responsive image

Auto-TLDR; Exploiting Batch Normalization before the Output Layer in Deep Learning for Minority Class Detection in Imbalanced Data Sets

Slides Poster Similar

Some real-world domains, such as Agriculture and Healthcare, comprise early-stage disease indications whose recording constitutes a rare event, and yet, whose precise detection at that stage is critical. In this type of highly imbalanced classification problems, which encompass complex features, deep learning (DL) is much needed because of its strong detection capabilities. At the same time, DL is observed in practice to favor majority over minority classes and consequently suffer from inaccurate detection of the targeted early-stage indications. To simulate such scenarios, we artificially generate skewness (99% vs. 1%) for certain plant types out of the PlantVillage dataset as a basis for classification of scarce visual cues through transfer learning. By randomly and unevenly picking healthy and unhealthy samples from certain plant types to form a training set, we consider a base experiment as fine-tuning ResNet34 and VGG19 architectures and then testing the model performance on a balanced dataset of healthy and unhealthy images. We empirically observe that the initial F1 test score jumps from 0.29 to 0.95 for the minority class upon adding a final Batch Normalization (BN) layer just before the output layer in VGG19. We demonstrate that utilizing an additional BN layer before the output layer in modern CNN architectures has a considerable impact in terms of minimizing the training time and testing error for minority classes in highly imbalanced data sets. Moreover, when the final BN is employed, trying to minimize validation and training losses may not be an optimal way for getting a high F1 test score for minority classes in anomaly detection problems. That is, the network might perform better even if it is not ‘confident’ enough while making a prediction; leading to another discussion about why softmax output is not a good uncertainty measure for DL models.

FastSal: A Computationally Efficient Network for Visual Saliency Prediction

Feiyan Hu, Kevin Mcguinness

Responsive image

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

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.

Adversarial Knowledge Distillation for a Compact Generator

Hideki Tsunashima, Shigeo Morishima, Junji Yamato, Qiu Chen, Hirokatsu Kataoka

Responsive image

Auto-TLDR; Adversarial Knowledge Distillation for Generative Adversarial Nets

Slides Poster Similar

In this paper, we propose memory-efficient Generative Adversarial Nets (GANs) in line with knowledge distillation. Most existing GANs have a shortcoming in terms of the number of model parameters and low processing speed. Here, to tackle the problem, we propose Adversarial Knowledge Distillation for Generative models (AKDG) for highly efficient GANs, in terms of unconditional generation. Using AKDG, model size and processing speed are substantively reduced. Through an adversarial training exercise with a distillation discriminator, a student generator successfully mimics a teacher generator in fewer model layers and fewer parameters and at a higher processing speed. Moreover, our AKDG is network architecture-agnostic. Comparison of AKDG-applied models to vanilla models suggests that it achieves closer scores to a teacher generator and more efficient performance than a baseline method with respect to Inception Score (IS) and Frechet Inception Distance (FID). In CIFAR-10 experiments, improving IS/FID 1.17pt/55.19pt and in LSUN bedroom experiments, improving FID 71.1pt in comparison to the conventional distillation method for GANs.

Exploiting Non-Linear Redundancy for Neural Model Compression

Muhammad Ahmed Shah, Raphael Olivier, Bhiksha Raj

Responsive image

Auto-TLDR; Compressing Deep Neural Networks with Linear Dependency

Slides Poster Similar

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.

Beyond Cross-Entropy: Learning Highly Separable Feature Distributions for Robust and Accurate Classification

Arslan Ali, Andrea Migliorati, Tiziano Bianchi, Enrico Magli

Responsive image

Auto-TLDR; Gaussian class-conditional simplex loss for adversarial robust multiclass classifiers

Slides Poster Similar

Deep learning has shown outstanding performance in several applications including image classification. However, deep classifiers are known to be highly vulnerable to adversarial attacks, in that a minor perturbation of the input can easily lead to an error. Providing robustness to adversarial attacks is a very challenging task especially in problems involving a large number of classes, as it typically comes at the expense of an accuracy decrease. In this work, we propose the Gaussian class-conditional simplex (GCCS) loss: a novel approach for training deep robust multiclass classifiers that provides adversarial robustness while at the same time achieving or even surpassing the classification accuracy of state-of-the-art methods. Differently from other frameworks, the proposed method learns a mapping of the input classes onto target distributions in a latent space such that the classes are linearly separable. Instead of maximizing the likelihood of target labels for individual samples, our objective function pushes the network to produce feature distributions yielding high inter-class separation. The mean values of the distributions are centered on the vertices of a simplex such that each class is at the same distance from every other class. We show that the regularization of the latent space based on our approach yields excellent classification accuracy and inherently provides robustness to multiple adversarial attacks, both targeted and untargeted, outperforming state-of-the-art approaches over challenging datasets.

Teacher-Student Training and Triplet Loss for Facial Expression Recognition under Occlusion

Mariana-Iuliana Georgescu, Radu Ionescu

Responsive image

Auto-TLDR; Knowledge Distillation for Facial Expression Recognition under Occlusion

Slides Similar

In this paper, we study the task of facial expression recognition under strong occlusion. We are particularly interested in cases where 50% of the face is occluded, e.g. when the subject wears a Virtual Reality (VR) headset. While previous studies show that pre-training convolutional neural networks (CNNs) on fully-visible (non-occluded) faces improves the accuracy, we propose to employ knowledge distillation to achieve further improvements. First of all, we employ the classic teacher-student training strategy, in which the teacher is a CNN trained on fully-visible faces and the student is a CNN trained on occluded faces. Second of all, we propose a new approach for knowledge distillation based on triplet loss. During training, the goal is to reduce the distance between an anchor embedding, produced by a student CNN that takes occluded faces as input, and a positive embedding (from the same class as the anchor), produced by a teacher CNN trained on fully-visible faces, so that it becomes smaller than the distance between the anchor and a negative embedding (from a different class than the anchor), produced by the student CNN. Third of all, we propose to combine the distilled embeddings obtained through the classic teacher-student strategy and our novel teacher-student strategy based on triplet loss into a single embedding vector. We conduct experiments on two benchmarks, FER+ and AffectNet, with two CNN architectures, VGG-f and VGG-face, showing that knowledge distillation can bring significant improvements over the state-of-the-art methods designed for occluded faces in the VR setting. Furthermore, we obtain accuracy rates that are quite close to the state-of-the-art models that take as input fully-visible faces. For example, on the FER+ data set, our VGG-face based on concatenated distilled embeddings attains an accuracy rate of 82.75% on lower-half-visible faces, which is only 2.24% below the accuracy rate of a state-of-the-art VGG-13 that is evaluated on fully-visible faces. Given that our model sees only the lower-half of the face, we consider this to be a remarkable achievement. In conclusion, we consider that our distilled CNN models can provide useful feedback for the task of recognizing the facial expressions of a person wearing a VR headset.

Contextual Classification Using Self-Supervised Auxiliary Models for Deep Neural Networks

Sebastian Palacio, Philipp Engler, Jörn Hees, Andreas Dengel

Responsive image

Auto-TLDR; Self-Supervised Autogenous Learning for Deep Neural Networks

Slides Poster Similar

Classification problems solved with deep neural networks (DNNs) typically rely on a closed world paradigm, and optimize over a single objective (e.g., minimization of the cross- entropy loss). This setup dismisses all kinds of supporting signals that can be used to reinforce the existence or absence of particular patterns. The increasing need for models that are interpretable by design makes the inclusion of said contextual signals a crucial necessity. To this end, we introduce the notion of Self-Supervised Autogenous Learning (SSAL). A SSAL objective is realized through one or more additional targets that are derived from the original supervised classification task, following architectural principles found in multi-task learning. SSAL branches impose low-level priors into the optimization process (e.g., grouping). The ability of using SSAL branches during inference, allow models to converge faster, focusing on a richer set of class-relevant features. We equip state-of-the-art DNNs with SSAL objectives and report consistent improvements for all of them on CIFAR100 and Imagenet. We show that SSAL models outperform similar state-of-the-art methods focused on contextual loss functions, auxiliary branches and hierarchical priors.

MaxDropout: Deep Neural Network Regularization Based on Maximum Output Values

Claudio Filipi Gonçalves Santos, Danilo Colombo, Mateus Roder, Joao Paulo Papa

Responsive image

Auto-TLDR; MaxDropout: A Regularizer for Deep Neural Networks

Slides Poster Similar

Different techniques have emerged in the deep learning scenario, such as Convolutional Neural Networks, Deep Belief Networks, and Long Short-Term Memory Networks, to cite a few. In lockstep, regularization methods, which aim to prevent overfitting by penalizing the weight connections, or turning off some units, have been widely studied either. In this paper, we present a novel approach called MaxDropout, a regularizer for deep neural network models that works in a supervised fashion by removing (shutting off) the prominent neurons (i.e., most active) in each hidden layer. The model forces fewer activated units to learn more representative information, thus providing sparsity. Regarding the experiments, we show that it is possible to improve existing neural networks and provide better results in neural networks when Dropout is replaced by MaxDropout. The proposed method was evaluated in image classification, achieving comparable results to existing regularizers, such as Cutout and RandomErasing, also improving the accuracy of neural networks that uses Dropout by replacing the existing layer by MaxDropout.

Iterative Label Improvement: Robust Training by Confidence Based Filtering and Dataset Partitioning

Christian Haase-Schütz, Rainer Stal, Heinz Hertlein, Bernhard Sick

Responsive image

Auto-TLDR; Meta Training and Labelling for Unlabelled Data

Slides Poster Similar

State-of-the-art, high capacity deep neural networks not only require large amounts of labelled training data, they are also highly susceptible to labelling errors in this data, typically resulting in large efforts and costs and therefore limiting the applicability of deep learning. To alleviate this issue, we propose a novel meta training and labelling scheme that is able to use inexpensive unlabelled data by taking advantage of the generalization power of deep neural networks. We show experimentally that by solely relying on one network architecture and our proposed scheme of combining self-training with pseudolabels, both label quality and resulting model accuracy, can be improved significantly. Our method achieves state-of-the-art results, while being architecture agnostic and therefore broadly applicable. Compared to other methods dealing with erroneous labels, our approach does neither require another network to be trained, nor does it necessarily need an additional, highly accurate reference label set. Instead of removing samples from a labelled set, our technique uses additional sensor data without the need for manual labelling. Furthermore, our approach can be used for semi-supervised learning.

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

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

Responsive image

Auto-TLDR; Compressing Deep Neural Models Using Mutual Information

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%.

Verifying the Causes of Adversarial Examples

Honglin Li, Yifei Fan, Frieder Ganz, Tony Yezzi, Payam Barnaghi

Responsive image

Auto-TLDR; Exploring the Causes of Adversarial Examples in Neural Networks

Slides Poster Similar

The robustness of neural networks is challenged by adversarial examples that contain almost imperceptible perturbations to inputs which mislead a classifier to incorrect outputs in high confidence. Limited by the extreme difficulty in examining a high-dimensional image space thoroughly, research on explaining and justifying the causes of adversarial examples falls behind studies on attacks and defenses. In this paper, we present a collection of potential causes of adversarial examples and verify (or partially verify) them through carefully-designed controlled experiments. The major causes of adversarial examples include model linearity, one-sum constraint, and geometry of the categories. To control the effect of those causes, multiple techniques are applied such as $L_2$ normalization, replacement of loss functions, construction of reference datasets, and novel models using multi-layer perceptron probabilistic neural networks (MLP-PNN) and density estimation (DE). Our experiment results show that geometric factors tend to be more direct causes and statistical factors magnify the phenomenon, especially for assigning high prediction confidence. We hope this paper will inspire more studies to rigorously investigate the root causes of adversarial examples, which in turn provide useful guidance on designing more robust models.

Learning Embeddings for Image Clustering: An Empirical Study of Triplet Loss Approaches

Kalun Ho, Janis Keuper, Franz-Josef Pfreundt, Margret Keuper

Responsive image

Auto-TLDR; Clustering Objectives for K-means and Correlation Clustering Using Triplet Loss

Slides Poster Similar

In this work, we evaluate two different image clustering objectives, k-means clustering and correlation clustering, in the context of Triplet Loss induced feature space embeddings. Specifically, we train a convolutional neural network to learn discriminative features by optimizing two popular versions of the Triplet Loss in order to study their clustering properties under the assumption of noisy labels. Additionally, we propose a new, simple Triplet Loss formulation, which shows desirable properties with respect to formal clustering objectives and outperforms the existing methods. We evaluate all three Triplet loss formulations for K-means and correlation clustering on the CIFAR-10 image classification dataset.

Can Data Placement Be Effective for Neural Networks Classification Tasks? Introducing the Orthogonal Loss

Brais Cancela, Veronica Bolon-Canedo, Amparo Alonso-Betanzos

Responsive image

Auto-TLDR; Spatial Placement for Neural Network Training Loss Functions

Slides Poster Similar

Traditionally, a Neural Network classification training loss function follows the same principle: minimizing the distance between samples that belong to the same class, while maximizing the distance to the other classes. There are no restrictions on the spatial placement of deep features (last layer input). This paper addresses this issue when dealing with Neural Networks, providing a set of loss functions that are able to train a classifier by forcing the deep features to be projected over a predefined orthogonal basis. Experimental results shows that these `data placement' functions can overcome the training accuracy provided by the classic cross-entropy loss function.

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

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

Responsive image

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

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

Responsive image

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

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.

Making Every Label Count: Handling Semantic Imprecision by Integrating Domain Knowledge

Clemens-Alexander Brust, Björn Barz, Joachim Denzler

Responsive image

Auto-TLDR; Class Hierarchies for Imprecise Label Learning and Annotation eXtrapolation

Slides Poster Similar

Noisy data, crawled from the web or supplied by volunteers such as Mechanical Turkers or citizen scientists, is considered an alternative to professionally labeled data. There has been research focused on mitigating the effects of label noise. It is typically modeled as inaccuracy, where the correct label is replaced by an incorrect label from the same set. We consider an additional dimension of label noise: imprecision. For example, a non-breeding snow bunting is labeled as a bird. This label is correct, but not as precise as the task requires. Standard softmax classifiers cannot learn from such a weak label because they consider all classes mutually exclusive, which non-breeding snow bunting and bird are not. We propose CHILLAX (Class Hierarchies for Imprecise Label Learning and Annotation eXtrapolation), a method based on hierarchical classification, to fully utilize labels of any precision. Experiments on noisy variants of NABirds and ILSVRC2012 show that our method outperforms strong baselines by as much as 16.4 percentage points, and the current state of the art by up to 3.9 percentage points.

Optimal Transport As a Defense against Adversarial Attacks

Quentin Bouniot, Romaric Audigier, Angélique Loesch

Responsive image

Auto-TLDR; Sinkhorn Adversarial Training with Optimal Transport Theory

Slides Poster Similar

Deep learning classifiers are now known to have flaws in the representations of their class. Adversarial attacks can find a human-imperceptible perturbation for a given image that will mislead a trained model. The most effective methods to defend against such attacks trains on generated adversarial examples to learn their distribution. Previous work aimed to align original and adversarial image representations in the same way as domain adaptation to improve robustness. Yet, they partially align the representations using approaches that do not reflect the geometry of space and distribution. In addition, it is difficult to accurately compare robustness between defended models. Until now, they have been evaluated using a fixed perturbation size. However, defended models may react differently to variations of this perturbation size. In this paper, the analogy of domain adaptation is taken a step further by exploiting optimal transport theory. We propose to use a loss between distributions that faithfully reflect the ground distance. This leads to SAT (Sinkhorn Adversarial Training), a more robust defense against adversarial attacks. Then, we propose to quantify more precisely the robustness of a model to adversarial attacks over a wide range of perturbation sizes using a different metric, the Area Under the Accuracy Curve (AUAC). We perform extensive experiments on both CIFAR-10 and CIFAR-100 datasets and show that our defense is globally more robust than the state-of-the-art.

Defense Mechanism against Adversarial Attacks Using Density-Based Representation of Images

Yen-Ting Huang, Wen-Hung Liao, Chen-Wei Huang

Responsive image

Auto-TLDR; Adversarial Attacks Reduction Using Input Recharacterization

Slides Poster Similar

Adversarial examples are slightly modified inputs devised to cause erroneous inference of deep learning models. Protection against the intervention of adversarial examples is a fundamental issue that needs to be addressed before the wide adoption of deep-learning based intelligent systems. In this research, we utilize the method known as input recharacterization to effectively eliminate the perturbations found in the adversarial examples. By converting images from the intensity domain into density-based representation using halftoning operation, performance of the classifier can be properly maintained. With adversarial attacks generated using FGSM, I-FGSM, and PGD, the top-5 accuracy of the hybrid model can still achieve 80.97%, 78.77%, 81.56%, respectively. Although the accuracy has been slightly affected, the influence of adversarial examples is significantly discounted. The average improvement over existing input transform defense mechanisms is approximately 10%.

Neural Compression and Filtering for Edge-assisted Real-time Object Detection in Challenged Networks

Yoshitomo Matsubara, Marco Levorato

Responsive image

Auto-TLDR; Deep Neural Networks for Remote Object Detection Using Edge Computing

Slides Poster Similar

The edge computing paradigm places compute-capable devices - edge servers - at the network edge to assist mobile devices in executing data analysis tasks. Intuitively, offloading compute-intense tasks to edge servers can reduce their execution time. However, poor conditions of the wireless channel connecting the mobile devices to the edge servers may degrade the overall capture-to-output delay achieved by edge offloading. Herein, we focus on edge computing supporting remote object detection by means of Deep Neural Networks (DNN), and develop a framework to reduce the amount of data transmitted over the wireless link. The core idea we propose builds on recent approaches splitting DNNs into sections - namely head and tail models - executed by the mobile device and edge server, respectively. The wireless link, then, is used to transport the output of the last layer of the head model to the edge server, instead of the DNN input. Most prior work focuses on classification tasks and leaves the DNN structure unaltered. Herein, we focus on DNNs for three different object detection tasks, which present a much more convoluted structure, and modify the architecture of the network to: (i) achieve in-network compression by introducing a bottleneck layer in the early layers on the head model, and (ii) prefilter pictures that do not contain objects of interest using a convolutional neural network. Results show that the proposed technique represents an effective intermediate option between local and edge computing in a parameter region where these extreme point solutions fail to provide satisfactory performance.

Multi-Order Feature Statistical Model for Fine-Grained Visual Categorization

Qingtao Wang, Ke Zhang, Shaoli Huang, Lianbo Zhang, Jin Fan

Responsive image

Auto-TLDR; Multi-Order Feature Statistical Method for Fine-Grained Visual Categorization

Slides Poster Similar

Fine-grained visual categorization aims to learn a robust image representation modeling subtle differences from similar categories. Existing methods in this field tackle the problem by designing complex frameworks, which produce high-level features by performing first-order or second-order pooling. Despite the impressive performance achieved by these strategies, the single-order networks only carry linear or non-linear information of the last convolutional layer, neglecting the fact that feature from different orders are mutually complementary. In this paper, we propose a Multi-Order Feature Statistical Method (MOFS), which learns fine-grained features characterizing multiple orders. Specifically, the MOFS consists of two sub-modules: (i) a first-order module modeling both mid-level and high-level features. (ii) a covariance feature statistical module capturing high-order features. By deploying these two sub-modules on the top of existing backbone networks, MOFS simultaneously captures multi-level of discrimative patters including local, global and co-related patters. We evaluate the proposed method on three challenging benchmarks, namely CUB-200-2011, Stanford Cars, and FGVC-Aircraft. Compared with state-of-the-art methods, experiments results exhibit superior performance in recognizing fine-grained objects

Exploiting Distilled Learning for Deep Siamese Tracking

Chengxin Liu, Zhiguo Cao, Wei Li, Yang Xiao, Shuaiyuan Du, Angfan Zhu

Responsive image

Auto-TLDR; Distilled Learning Framework for Siamese Tracking

Slides Poster Similar

Existing deep siamese trackers are typically built on off-the-shelf CNN models for feature learning, with the demand for huge power consumption and memory storage. This limits current deep siamese trackers to be carried on resource-constrained devices like mobile phones, given factor that such a deployment normally requires cost-effective considerations. In this work, we address this issue by presenting a novel Distilled Learning Framework(DLF) for siamese tracking, which aims at learning tracking model with efficiency and high accuracy. Specifically, we propose two simple yet effective knowledge distillation strategies, denote as point-wise distillation and pair-wise distillation, which are designed for transferring knowledge from a more discriminative teacher tracker into a compact student tracker. In this way, cost-effective and high performance tracking could be achieved. Extensive experiments on several tracking benchmarks demonstrate the effectiveness of our proposed method.

Revisiting ImprovedGAN with Metric Learning for Semi-Supervised Learning

Jaewoo Park, Yoon Gyo Jung, Andrew Teoh

Responsive image

Auto-TLDR; Improving ImprovedGAN with Metric Learning for Semi-supervised Learning

Slides Poster Similar

Semi-supervised Learning (SSL) is a classical problem where a model needs to solve classification as it is trained on a partially labeled train data. After the introduction of generative adversarial network (GAN) and its success, the model has been modified to be applicable to SSL. ImprovedGAN as a representative model for GAN-based SSL, it showed promising performance on the SSL problem. However, the inner mechanism of this model has been only partially revealed. In this work, we revisit ImprovedGAN with a fresh perspective based on metric learning. In particular, we interpret ImprovedGAN by general pair weighting, a recent framework in metric learning. Based on this interpretation, we derive two theoretical properties of ImprovedGAN: (i) its discriminator learns to make confident predictions over real samples, (ii) the adversarial interaction in ImprovedGAN along with semi-supervision results in cluster separation by reducing intra-class variance and increasing the inter-class variance, thereby improving the model generalization. These theoretical implications are experimentally supported. Motivated by the findings, we propose a variant of ImprovedGAN, called Intensified ImprovedGAN (I2GAN), where its cluster separation characteristic is enhanced by two proposed techniques: (a) the unsupervised discriminator loss is scaled up and (b) the generated batch size is enlarged. As a result, I2GAN produces better class-wise cluster separation and, hence, generalization. Extensive experiments on the widely known benchmark data sets verify the effectiveness of our proposed method, showing that its performance is better than or comparable to other GAN based SSL models.

Graph-Based Interpolation of Feature Vectors for Accurate Few-Shot Classification

Yuqing Hu, Vincent Gripon, Stéphane Pateux

Responsive image

Auto-TLDR; Transductive Learning for Few-Shot Classification using Graph Neural Networks

Slides Poster Similar

In few-shot classification, the aim is to learn models able to discriminate classes using only a small number of labeled examples. In this context, works have proposed to introduce Graph Neural Networks (GNNs) aiming at exploiting the information contained in other samples treated concurrently, what is commonly referred to as the transductive setting in the literature. These GNNs are trained all together with a backbone feature extractor. In this paper, we propose a new method that relies on graphs only to interpolate feature vectors instead, resulting in a transductive learning setting with no additional parameters to train. Our proposed method thus exploits two levels of information: a) transfer features obtained on generic datasets, b) transductive information obtained from other samples to be classified. Using standard few-shot vision classification datasets, we demonstrate its ability to bring significant gains compared to other works.

Adaptive Distillation for Decentralized Learning from Heterogeneous Clients

Jiaxin Ma, Ryo Yonetani, Zahid Iqbal

Responsive image

Auto-TLDR; Decentralized Learning via Adaptive Distillation

Slides Poster Similar

This paper addresses the problem of decentralized learning to achieve a high-performance global model by asking a group of clients to share local models pre-trained with their own data resources. We are particularly interested in a specific case where both the client model architectures and data distributions are diverse, which makes it nontrivial to adopt conventional approaches such as Federated Learning and network co-distillation. To this end, we propose a new decentralized learning method called Decentralized Learning via Adaptive Distillation (DLAD). Given a collection of client models and a large number of unlabeled distillation samples, the proposed DLAD 1) aggregates the outputs of the client models while adaptively emphasizing those with higher confidence in given distillation samples and 2) trains the global model to imitate the aggregated outputs. Our extensive experimental evaluation on multiple public datasets (MNIST, CIFAR-10, and CINIC-10) demonstrates the effectiveness of the proposed method.

Variational Inference with Latent Space Quantization for Adversarial Resilience

Vinay Kyatham, Deepak Mishra, Prathosh A.P.

Responsive image

Auto-TLDR; A Generalized Defense Mechanism for Adversarial Attacks on Data Manifolds

Slides Poster Similar

Despite their tremendous success in modelling highdimensional data manifolds, deep neural networks suffer from the threat of adversarial attacks - Existence of perceptually valid input-like samples obtained through careful perturbation that lead to degradation in the performance of the underlying model. Major concerns with existing defense mechanisms include non-generalizability across different attacks, models and large inference time. In this paper, we propose a generalized defense mechanism capitalizing on the expressive power of regularized latent space based generative models. We design an adversarial filter, devoid of access to classifier and adversaries, which makes it usable in tandem with any classifier. The basic idea is to learn a Lipschitz constrained mapping from the data manifold, incorporating adversarial perturbations, to a quantized latent space and re-map it to the true data manifold. Specifically, we simultaneously auto-encode the data manifold and its perturbations implicitly through the perturbations of the regularized and quantized generative latent space, realized using variational inference. We demonstrate the efficacy of the proposed formulation in providing resilience against multiple attack types (black and white box) and methods, while being almost real-time. Our experiments show that the proposed method surpasses the stateof-the-art techniques in several cases.

A Close Look at Deep Learning with Small Data

Lorenzo Brigato, Luca Iocchi

Responsive image

Auto-TLDR; Low-Complex Neural Networks for Small Data Conditions

Slides Poster Similar

In this work, we perform a wide variety of experiments with different Deep Learning architectures in small data conditions. We show that model complexity is a critical factor when only a few samples per class are available. Differently from the literature, we improve the state of the art using low complexity models. We show that standard convolutional neural networks with relatively few parameters are effective in this scenario. In many of our experiments, low complexity models outperform state-of-the-art architectures. Moreover, we propose a novel network that uses an unsupervised loss to regularize its training. Such architecture either improves the results either performs comparably well to low capacity networks. Surprisingly, experiments show that the dynamic data augmentation pipeline is not beneficial in this particular domain. Statically augmenting the dataset might be a promising research direction while dropout maintains its role as a good regularizer.

Adversarially Training for Audio Classifiers

Raymel Alfonso Sallo, Mohammad Esmaeilpour, Patrick Cardinal

Responsive image

Auto-TLDR; Adversarially Training for Robust Neural Networks against Adversarial Attacks

Slides Poster Similar

In this paper, we investigate the potential effect of the adversarially training on the robustness of six advanced deep neural networks against a variety of targeted and non-targeted adversarial attacks. We firstly show that, the ResNet-56 model trained on the 2D representation of the discrete wavelet transform appended with the tonnetz chromagram outperforms other models in terms of recognition accuracy. Then we demonstrate the positive impact of adversarially training on this model as well as other deep architectures against six types of attack algorithms (white and black-box) with the cost of the reduced recognition accuracy and limited adversarial perturbation. We run our experiments on two benchmarking environmental sound datasets and show that without any imposed limitations on the budget allocations for the adversary, the fooling rate of the adversarially trained models can exceed 90%. In other words, adversarial attacks exist in any scales, but they might require higher adversarial perturbations compared to non-adversarially trained models.

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

Madan Ravi Ganesh, Jason Corso, Salimeh Yasaei Sekeh

Responsive image

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

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.

Smart Inference for Multidigit Convolutional Neural Network Based Barcode Decoding

Duy-Thao Do, Tolcha Yalew, Tae Joon Jun, Daeyoung Kim

Responsive image

Auto-TLDR; Smart Inference for Barcode Decoding using Deep Convolutional Neural Network

Slides Poster Similar

Barcodes are ubiquitous and have been used in most of critical daily activities for decades. However, most of traditional decoders require well-founded barcode under a relatively standard condition. While wilder conditioned barcodes such as underexposed, occluded, blurry, wrinkled and rotated are commonly captured in reality, those traditional decoders show weakness of recognizing. Several works attempted to solve those challenging barcodes, but many limitations still exist. This work aims to solve the decoding problem using deep convolutional neural network with the possibility of running on portable devices. Firstly, we proposed a special modification of inference based on the feature of having checksum and test-time augmentation, named as Smart Inference (SI) in prediction phase of a trained model. SI considerably boosts accuracy and reduces the false prediction for trained models. Secondly, we have created a large practical evaluation dataset of real captured 1D barcode under various challenging conditions to test our methods vigorously, which is publicly available for other researchers. The experiments' results demonstrated the SI effectiveness with the highest accuracy of 95.85% which outperformed many existing decoders on the evaluation set. Finally, we successfully minimized the best model by knowledge distillation to a shallow model which is shown to have high accuracy (90.85%) with good inference speed of 34.2 ms per image on a real edge device.

Dynamic Multi-Path Neural Network

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

Responsive image

Auto-TLDR; Dynamic Multi-path Neural Network

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.

Minority Class Oriented Active Learning for Imbalanced Datasets

Umang Aggarwal, Adrian Popescu, Celine Hudelot

Responsive image

Auto-TLDR; Active Learning for Imbalanced Datasets

Slides Poster Similar

Active learning aims to optimize the dataset annotation process when resources are constrained. Most existing methods are designed for balanced datasets. Their practical applicability is limited by the fact that a majority of real-life datasets are actually imbalanced. Here, we introduce a new active learning method which is designed for imbalanced datasets. It favors samples likely to be in minority classes so as to reduce the imbalance of the labeled subset and create a better representation for these classes. We also compare two training schemes for active learning: (1) the one commonly deployed in deep active learning using model fine tuning for each iteration and (2) a scheme which is inspired by transfer learning and exploits generic pre-trained models and train shallow classifiers for each iteration. Evaluation is run with three imbalanced datasets. Results show that the proposed active learning method outperforms competitive baselines. Equally interesting, they also indicate that the transfer learning training scheme outperforms model fine tuning if features are transferable from the generic dataset to the unlabeled one. This last result is surprising and should encourage the community to explore the design of deep active learning methods.

Rethinking of Deep Models Parameters with Respect to Data Distribution

Shitala Prasad, Dongyun Lin, Yiqun Li, Sheng Dong, Zaw Min Oo

Responsive image

Auto-TLDR; A progressive stepwise training strategy for deep neural networks

Slides Poster Similar

The performance of deep learning models are driven by various parameters but to tune all of them every time, for every dataset, is a heuristic practice. In this paper, unlike the common practice of decaying the learning rate, we propose a step-wise training strategy where the learning rate and the batch size are tuned based on the dataset size. Here, the given dataset size is progressively increased during the training to boost the network performance without saturating the learning curve, after certain epochs. We conducted extensive experiments on multiple networks and datasets to validate the proposed training strategy. The experimental results proves our hypothesis that the learning rate, the batch size and the data size are interrelated and can improve the network accuracy if an optimal progressive stepwise training strategy is applied. The proposed strategy also the overall training computational cost is reduced.