Knowledge Injection

Recent advancements in deploying deep neural networks (DNNs) on resource-constrained devices have generated interest in input-adaptive dynamic neural networks (DyNNs). DyNNs offer more efficient inferences and enable the deployment of DNNs on devices with limited resources, such as mobile devices. However, we have discovered a new vulnerability in DyNNs that could potentially compromise their efficiency. Specifically, we investigate whether adversaries can manipulate DyNNs' computational costs to create a false sense of efficiency. To address this question, we propose EfficFrog, an adversarial attack that injects universal efficiency backdoors in DyNNs. To inject a backdoor trigger into DyNNs, EfficFrog poisons only a minimal percentage of the DyNNs' training data. During the inference phase, EfficFrog can slow down the backdoored DyNNs and abuse the computational resources of systems running DyNNs by adding the trigger to any input. To evaluate EfficFrog, we tested it on three DNN backbone architectures (based on VGG16, MobileNet, and ResNet56) using two popular datasets (CIFAR-10 and Tiny ImageNet). Our results demonstrate that EfficFrog reduces the efficiency of DyNNs on triggered input samples while keeping the efficiency of clean samples almost the same.

Modern Generative Adversarial Networks (GANs) generate realistic images remarkably well. Previous work has demonstrated the feasibility of "GAN-classifiers" that are distinct from the co-trained discriminator, and operate on images generated from a frozen GAN. That such classifiers work at all affirms the existence of "knowledge gaps" (out-of-distribution artifacts across samples) present in GAN training. We iteratively train GAN-classifiers and train GANs that "fool" the classifiers (in an attempt to fill the knowledge gaps), and examine the effect on GAN training dynamics, output quality, and GAN-classifier generalization. We investigate two settings, a small DCGAN architecture trained on low dimensional images (MNIST), and StyleGAN2, a SOTA GAN architecture trained on high dimensional images (FFHQ). We find that the DCGAN is unable to effectively fool a held-out GAN-classifier without compromising the output quality. However, StyleGAN2 can fool held-out classifiers with no change in output quality, and this effect persists over multiple rounds of GAN/classifier training which appears to reveal an ordering over optima in the generator parameter space. Finally, we study different classifier architectures and show that the architecture of the GAN-classifier has a strong influence on the set of its learned artifacts.

Mean ensemble (i.e. averaging predictions from multiple models) is a commonly-used technique in machine learning that improves the performance of each individual model. We formalize it as feature alignment for ensemble in open-set face recognition and generalize it into Bayesian Ensemble Averaging (BEA) through the lens of probabilistic modeling. This generalization brings up two practical benefits that existing methods could not provide: (1) the uncertainty of a face image can be evaluated and further decomposed into aleatoric uncertainty and epistemic uncertainty, the latter of which can be used as a measure for out-of-distribution detection of faceness; (2) a BEA statistic provably reflects the aleatoric uncertainty of a face image, acting as a measure for face image quality to improve recognition performance. To inherit the uncertainty estimation capability from BEA without the loss of inference efficiency, we propose BEA-KD, a student model to distill knowledge from BEA. BEA-KD mimics the overall behavior of ensemble members and consistently outperforms SOTA knowledge distillation methods on various challenging benchmarks.

Pre-training on large-scale video data has become a common recipe for learning transferable spatiotemporal representations in recent years. Despite some progress, existing methods are mostly limited to highly curated datasets (e.g., K400) and exhibit unsatisfactory out-of-the-box representations. We argue that it is due to the fact that they only capture pixel-level knowledge rather than spatiotemporal semantics, which hinders further progress in video understanding. Inspired by the great success of image-text pre-training (e.g., CLIP), we take the first step to exploit language semantics to boost transferable spatiotemporal representation learning. We introduce a new pretext task, Turning to Video for Transcript Sorting (TVTS), which sorts shuffled ASR scripts by attending to learned video representations. We do not rely on descriptive captions and learn purely from video, i.e., leveraging the natural transcribed speech knowledge to provide noisy but useful semantics over time. Our method enforces the vision model to contextualize what is happening over time so that it can re-organize the narrative transcripts, and can seamlessly apply to large-scale uncurated video data in the real world. Our method demonstrates strong out-of-the-box spatiotemporal representations on diverse benchmarks, e.g., +13.6% gains over VideoMAE on SSV2 via linear probing. The code is available at https://github.com/TencentARC/TVTS.

Data-free Knowledge Distillation (DFKD) has gained popularity recently, with the fundamental idea of carrying out knowledge transfer from a Teacher neural network to a Student neural network in the absence of training data. However, in the Adversarial DFKD framework, the student network's accuracy, suffers due to the non-stationary distribution of the pseudo-samples under multiple generator updates. To this end, at every generator update, we aim to maintain the student's performance on previously encountered examples while acquiring knowledge from samples of the current distribution. Thus, we propose a meta-learning inspired framework by treating the task of Knowledge-Acquisition (learning from newly generated samples) and Knowledge-Retention (retaining knowledge on previously met samples) as meta-train and meta-test, respectively. Hence, we dub our method as Learning to Retain while Acquiring. Moreover, we identify an implicit aligning factor between the Knowledge-Retention and Knowledge-Acquisition tasks indicating that the proposed student update strategy enforces a common gradient direction for both tasks, alleviating interference between the two objectives. Finally, we support our hypothesis by exhibiting extensive evaluation and comparison of our method with prior arts on multiple datasets.

Knowledge distillation facilitates the training of a compact student network by using a deep teacher one. While this has achieved great success in many tasks, it remains completely unstudied for image-based 6D object pose estimation. In this work, we introduce the first knowledge distillation method driven by the 6D pose estimation task. To this end, we observe that most modern 6D pose estimation frameworks output local predictions, such as sparse 2D keypoints or dense representations, and that the compact student network typically struggles to predict such local quantities precisely. Therefore, instead of imposing prediction-to-prediction supervision from the teacher to the student, we propose to distill the teacher's distribution of local predictions into the student network, facilitating its training. Our experiments on several benchmarks show that our distillation method yields state-of-the-art results with different compact student models and for both keypoint-based and dense prediction-based architectures.

Rotated bounding boxes drastically reduce output ambiguity of elongated objects, making it superior to axis-aligned bounding boxes. Despite the effectiveness, rotated detectors are not widely employed. Annotating rotated bounding boxes is such a laborious process that they are not provided in many detection datasets where axis-aligned annotations are used instead. In this paper, we propose a framework that allows the model to predict precise rotated boxes only requiring cheaper axis-aligned annotation of the target dataset. To achieve this, we leverage the fact that neural networks are capable of learning richer representation of the target domain than what is utilized by the task. The under-utilized representation can be exploited to address a more detailed task. Our framework combines task knowledge of an out-of-domain source dataset with stronger annotation and domain knowledge of the target dataset with weaker annotation. A novel assignment process and projection loss are used to enable the co-training on the source and target datasets. As a result, the model is able to solve the more detailed task in the target domain, without additional computation overhead during inference. We extensively evaluate the method on various target datasets including fresh-produce dataset, HRSC2016 and SSDD. Results show that the proposed method consistently performs on par with the fully supervised approach.

Machine unlearning can fortify the privacy and security of machine learning applications. Unfortunately, the exact unlearning approaches are inefficient, and the approximate unlearning approaches are unsuitable for complicated CNNs. Moreover, the approximate approaches have serious security flaws because even unlearning completely different data points can produce the same contribution estimation as unlearning the target data points. To address the above problems, we try to define machine unlearning from the knowledge perspective, and we propose a knowledge-level machine unlearning method, namely ERM-KTP. Specifically, we propose an entanglement-reduced mask (ERM) structure to reduce the knowledge entanglement among classes during the training phase. When receiving the unlearning requests, we transfer the knowledge of the non-target data points from the original model to the unlearned model and meanwhile prohibit the knowledge of the target data points via our proposed knowledge transfer and prohibition (KTP) method. Finally, we will get the unlearned model as the result and delete the original model to accomplish the unlearning process. Especially, our proposed ERM-KTP is an interpretable unlearning method because the ERM structure and the crafted masks in KTP can explicitly explain the operation and the effect of unlearning data points. Extensive experiments demonstrate the effectiveness, efficiency, high fidelity, and scalability of the ERM-KTP unlearning method.

The success of multi-task learning (MTL) can largely be attributed to the shared representation of related tasks, allowing the models to better generalise. In deep learning, this is usually achieved by sharing a common neural network architecture and jointly training the weights. However, the joint training of weighting parameters on multiple related tasks may lead to performance degradation, known as negative transfer. To address this issue, this work proposes an evolutionary multi-tasking neural architecture search (EMT-NAS) algorithm to accelerate the search process by transferring architectural knowledge across multiple related tasks. In EMT-NAS, unlike the traditional MTL, the model for each task has a personalised network architecture and its own weights, thus offering the capability of effectively alleviating negative transfer. A fitness re-evaluation method is suggested to alleviate fluctuations in performance evaluations resulting from parameter sharing and the mini-batch gradient descent training method, thereby avoiding losing promising solutions during the search process. To rigorously verify the performance of EMT-NAS, the classification tasks used in the empirical assessments are derived from different datasets, including the CIFAR-10 and CIFAR-100, and four MedMNIST datasets. Extensive comparative experiments on different numbers of tasks demonstrate that EMT-NAS takes 8% and up to 40% on CIFAR and MedMNIST, respectively, less time to find competitive neural architectures than its single-task counterparts.

Spiking neural networks (SNNs) are well known as the brain-inspired models with high computing efficiency, due to a key component that they utilize spikes as information units, close to the biological neural systems. Although spiking based models are energy efficient by taking advantage of discrete spike signals, their performance is limited by current network structures and their training methods. As discrete signals, typical SNNs cannot apply the gradient descent rules directly into parameters adjustment as artificial neural networks (ANNs). Aiming at this limitation, here we propose a novel method of constructing deep SNN models with knowledge distillation (KD) that uses ANN as teacher model and SNN as student model. Through ANN-SNN joint training algorithm, the student SNN model can learn rich feature information from the teacher ANN model through the KD method, yet it avoids training SNN from scratch when communicating with non-differentiable spikes. Our method can not only build a more efficient deep spiking structure feasibly and reasonably, but use few time steps to train whole model compared to direct training or ANN to SNN methods. More importantly, it has a superb ability of noise immunity for various types of artificial noises and natural signals. The proposed novel method provides efficient ways to improve the performance of SNN through constructing deeper structures in a high-throughput fashion, with potential usage for light and efficient brain-inspired computing of practical scenarios.

Previous knowledge distillation methods have shown their impressive performance on model compression tasks, however, it is hard to explain how the knowledge they transferred helps to improve the performance of the student network. In this work, we focus on proposing a knowledge distillation method that has both high interpretability and competitive performance. We first revisit the structure of mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions of input is critical for CNN to perform classification. Furthermore, we demonstrate that this capacity can be obtained and enhanced by transferring class activation maps. Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD). Different from previous KD methods, we explore and present several properties of the knowledge transferred by our method, which not only improve the interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Code is available at: https://github.com/GzyAftermath/CAT-KD.

Humans possess the capacity to reason about the future based on a sparse collection of visual cues acquired over time. In order to emulate this ability, we introduce a novel task called Anticipation Captioning, which generates a caption for an unseen oracle image using a sparsely temporally-ordered set of images. To tackle this new task, we propose a model called A-CAP, which incorporates commonsense knowledge into a pre-trained vision-language model, allowing it to anticipate the caption. Through both qualitative and quantitative evaluations on a customized visual storytelling dataset, A-CAP outperforms other image captioning methods and establishes a strong baseline for anticipation captioning. We also address the challenges inherent in this task.

Knowledge distillation (KD) is a technique that transfers the knowledge from a large teacher network to a small student network. It has been widely applied to many different tasks, such as model compression and federated learning. However, existing KD methods fail to generalize to general deep directed graphical models (DGMs) with arbitrary layers of random variables. We refer by deep DGMs to DGMs whose conditional distributions are parameterized by deep neural networks. In this work, we propose a novel unified knowledge distillation framework for deep DGMs on various applications. Specifically, we leverage the reparameterization trick to hide the intermediate latent variables, resulting in a compact DGM. Then we develop a surrogate distillation loss to reduce error accumulation through multiple layers of random variables. Moreover, we present the connections between our method and some existing knowledge distillation approaches. The proposed framework is evaluated on four applications: data-free hierarchical variational autoencoder (VAE) compression, data-free variational recurrent neural networks (VRNN) compression, data-free Helmholtz Machine (HM) compression, and VAE continual learning. The results show that our distillation method outperforms the baselines in data-free model compression tasks. We further demonstrate that our method significantly improves the performance of KD-based continual learning for data generation. Our source code is available at https://github.com/YizhuoChen99/KD4DGM-CVPR.