The malicious use and widespread dissemination of deepfake pose a significant crisis of trust. Current deepfake detection models can generally recognize forgery images by training on a large dataset. However, the accuracy of detection models degrades significantly on images generated by new deepfake methods due to the difference in data distribution. To tackle this issue, we present a novel incremental learning framework that improves the generalization of deepfake detection models by continual learning from a small number of new samples. To cope with different data distributions, we propose to learn a domain-invariant representation based on supervised contrastive learning, preventing overfit to the insufficient new data. To mitigate catastrophic forgetting, we regularize our model in both feature-level and label-level based on a multi-perspective knowledge distillation approach. Finally, we propose to select both central and hard representative samples to update the replay set, which is beneficial for both domain-invariant representation learning and rehearsal-based knowledge preserving. We conduct extensive experiments on four benchmark datasets, obtaining the new state-of-the-art average forgetting rate of 7.01 and average accuracy of 85.49 on FF++, DFDC-P, DFD, and CDF2. Our code is released at https://github.com/DeepFakeIL/DFIL.
In this work, we introduce a framework for cross-lingual speech synthesis, which involves an upstream Voice Conversion (VC) model and a downstream Text-To-Speech (TTS) model. The proposed framework consists of 4 stages. In the first two stages, we use a VC model to convert utterances in the target locale to the voice of the target speaker. In the third stage, the converted data is combined with the linguistic features and durations from recordings in the target language, which are then used to train a single-speaker acoustic model. Finally, the last stage entails the training of a locale-independent vocoder. Our evaluations show that the proposed paradigm outperforms state-of-the-art approaches which are based on training a large multilingual TTS model. In addition, our experiments demonstrate the robustness of our approach with different model architectures, languages, speakers and amounts of data. Moreover, our solution is especially beneficial in low-resource settings.
Tiny, causal models are crucial for embedded audio machine learning applications. Model compression can be achieved via distilling knowledge from a large teacher into a smaller student model. In this work, we propose a novel two-step approach for tiny speech enhancement model distillation. In contrast to the standard approach of a weighted mixture of distillation and supervised losses, we firstly pre-train the student using only the knowledge distillation (KD) objective, after which we switch to a fully supervised training regime. We also propose a novel fine-grained similarity-preserving KD loss, which aims to match the student's intra-activation Gram matrices to that of the teacher. Our method demonstrates broad improvements, but particularly shines in adverse conditions including high compression and low signal to noise ratios (SNR), yielding signal to distortion ratio gains of 0.9 dB and 1.1 dB, respectively, at -5 dB input SNR and 63x compression compared to baseline.
Large-scale self-supervised pre-trained speech encoders outperform conventional approaches in speech recognition and translation tasks. Due to the high cost of developing these large models, building new encoders for new tasks and deploying them to on-device applications are infeasible. Prior studies propose model compression methods to address this issue, but those works focus on smaller models and less realistic tasks. Thus, we propose Contrastive Layer-to-layer Distillation (CoLLD), a novel knowledge distillation method to compress pre-trained speech encoders by leveraging masked prediction and contrastive learning to train student models to copy the behavior of a large teacher model. CoLLD outperforms prior methods and closes the gap between small and large models on multilingual speech-to-text translation and recognition benchmarks.
Colorization is a well-explored problem in the domains of image and video processing. However, extending colorization to 3D scenes presents significant challenges. Recent Neural Radiance Field (NeRF) and Gaussian-Splatting(3DGS) methods enable high-quality novel-view synthesis for multi-view images. However, the question arises: How can we colorize these 3D representations? This work presents a method for synthesizing colorized novel views from input grayscale multi-view images. Using image or video colorization methods to colorize novel views from these 3D representations naively will yield output with severe inconsistencies. We introduce a novel method to use powerful image colorization models for colorizing 3D representations. We propose a distillation-based method that transfers color from these networks trained on natural images to the target 3D representation. Notably, this strategy does not add any additional weights or computational overhead to the original representation during inference. Extensive experiments demonstrate that our method produces high-quality colorized views for indoor and outdoor scenes, showcasing significant cross-view consistency advantages over baseline approaches. Our method is agnostic to the underlying 3D representation and easily generalizable to NeRF and 3DGS methods. Further, we validate the efficacy of our approach in several diverse applications: 1.) Infra-Red (IR) multi-view images and 2.) Legacy grayscale multi-view image sequences. Project Webpage: https://val.cds.iisc.ac.in/chroma-distill.github.io/
Implicit discourse relation recognition (IDRR) aims at recognizing the discourse relation between two text segments without an explicit connective. Recently, the prompt learning has just been applied to the IDRR task with great performance improvements over various neural network-based approaches. However, the discrete nature of the state-art-of-art prompting approach requires manual design of templates and answers, a big hurdle for its practical applications. In this paper, we propose a continuous version of prompt learning together with connective knowledge distillation, called AdaptPrompt, to reduce manual design efforts via continuous prompting while further improving performance via knowledge transfer. In particular, we design and train a few virtual tokens to form continuous templates and automatically select the most suitable one by gradient search in the embedding space. We also design an answer-relation mapping rule to generate a few virtual answers as the answer space. Furthermore, we notice the importance of annotated connectives in the training dataset and design a teacher-student architecture for knowledge transfer. Experiments on the up-to-date PDTB Corpus V3.0 validate our design objectives in terms of the better relation recognition performance over the state-of-the-art competitors.
We present a novel local-global feature fusion framework for body-weight exercise recognition with floor-based dynamic pressure maps. One step further from the existing studies using deep neural networks mainly focusing on global feature extraction, the proposed framework aims to combine local and global features using image processing techniques and the YOLO object detection to localize pressure profiles from different body parts and consider physical constraints. The proposed local feature extraction method generates two sets of high-level local features consisting of cropped pressure mapping and numerical features such as angular orientation, location on the mat, and pressure area. In addition, we adopt a knowledge distillation for regularization to preserve the knowledge of the global feature extraction and improve the performance of the exercise recognition. Our experimental results demonstrate a notable 11 percent improvement in F1 score for exercise recognition while preserving label-specific features.
Decoding visual representations from human brain activity has emerged as a thriving research domain, particularly in the context of brain-computer interfaces. Our study presents an innovative method that employs to classify and reconstruct images from the ImageNet dataset using electroencephalography (EEG) data from subjects that had viewed the images themselves (i.e. "brain decoding"). We analyzed EEG recordings from 6 participants, each exposed to 50 images spanning 40 unique semantic categories. These EEG readings were converted into spectrograms, which were then used to train a convolutional neural network (CNN), integrated with a knowledge distillation procedure based on a pre-trained Contrastive Language-Image Pre-Training (CLIP)-based image classification teacher network. This strategy allowed our model to attain a top-5 accuracy of 80%, significantly outperforming a standard CNN and various RNN-based benchmarks. Additionally, we incorporated an image reconstruction mechanism based on pre-trained latent diffusion models, which allowed us to generate an estimate of the images which had elicited EEG activity. Therefore, our architecture not only decodes images from neural activity but also offers a credible image reconstruction from EEG only, paving the way for e.g. swift, individualized feedback experiments. Our research represents a significant step forward in connecting neural signals with visual cognition.
Recent advancements in data-driven task-oriented dialogue systems (ToDs) struggle with incremental learning due to computational constraints and time-consuming issues. Continual Learning (CL) attempts to solve this by avoiding intensive pre-training, but it faces the problem of catastrophic forgetting (CF). While generative-based rehearsal CL methods have made significant strides, generating pseudo samples that accurately reflect the underlying task-specific distribution is still a challenge. In this paper, we present Dirichlet Continual Learning (DCL), a novel generative-based rehearsal strategy for CL. Unlike the traditionally used Gaussian latent variable in the Conditional Variational Autoencoder (CVAE), DCL leverages the flexibility and versatility of the Dirichlet distribution to model the latent prior variable. This enables it to efficiently capture sentence-level features of previous tasks and effectively guide the generation of pseudo samples. In addition, we introduce Jensen-Shannon Knowledge Distillation (JSKD), a robust logit-based knowledge distillation method that enhances knowledge transfer during pseudo sample generation. Our experiments confirm the efficacy of our approach in both intent detection and slot-filling tasks, outperforming state-of-the-art methods.
Self-training allows a network to learn from the predictions of a more complicated model, thus often requires well-trained teacher models and mixture of teacher-student data while multi-task learning jointly optimizes different targets to learn salient interrelationship and requires multi-task annotations for each training example. These frameworks, despite being particularly data demanding have potentials for data exploitation if such assumptions can be relaxed. In this paper, we compare self-training object detection under the deficiency of teacher training data where students are trained on unseen examples by the teacher, and multi-task learning with partially annotated data, i.e. single-task annotation per training example. Both scenarios have their own limitation but potentially helpful with limited annotated data. Experimental results show the improvement of performance when using a weak teacher with unseen data for training a multi-task student. Despite the limited setup we believe the experimental results show the potential of multi-task knowledge distillation and self-training, which could be beneficial for future study. Source code is at https://lhoangan.github.io/multas.
Semi-supervised learning (SSL) has become a crucial approach in deep learning as a way to address the challenge of limited labeled data. The success of deep neural networks heavily relies on the availability of large-scale high-quality labeled data. However, the process of data labeling is time-consuming and unscalable, leading to shortages in labeled data. SSL aims to tackle this problem by leveraging additional unlabeled data in the training process. One of the popular SSL algorithms, FixMatch, trains identical weight-sharing teacher and student networks simultaneously using a siamese neural network (SNN). However, it is prone to performance degradation when the pseudo labels are heavily noisy in the early training stage. We present KD-FixMatch, a novel SSL algorithm that addresses the limitations of FixMatch by incorporating knowledge distillation. The algorithm utilizes a combination of sequential and simultaneous training of SNNs to enhance performance and reduce performance degradation. Firstly, an outer SNN is trained using labeled and unlabeled data. After that, the network of the well-trained outer SNN generates pseudo labels for the unlabeled data, from which a subset of unlabeled data with trusted pseudo labels is then carefully created through high-confidence sampling and deep embedding clustering. Finally, an inner SNN is trained with the labeled data, the unlabeled data, and the subset of unlabeled data with trusted pseudo labels. Experiments on four public data sets demonstrate that KD-FixMatch outperforms FixMatch in all cases. Our results indicate that KD-FixMatch has a better training starting point that leads to improved model performance compared to FixMatch.
This paper explores the use of ASR-pretrained Conformers for speaker verification, leveraging their strengths in modeling speech signals. We introduce three strategies: (1) Transfer learning to initialize the speaker embedding network, improving generalization and reducing overfitting. (2) Knowledge distillation to train a more flexible speaker verification model, incorporating frame-level ASR loss as an auxiliary task. (3) A lightweight speaker adaptor for efficient feature conversion without altering the original ASR Conformer, allowing parallel ASR and speaker verification. Experiments on VoxCeleb show significant improvements: transfer learning yields a 0.48% EER, knowledge distillation results in a 0.43% EER, and the speaker adaptor approach, with just an added 4.92M parameters to a 130.94M-parameter model, achieves a 0.57% EER. Overall, our methods effectively transfer ASR capabilities to speaker verification tasks.
Deep learning-based image compression has made great progresses recently. However, many leading schemes use serial context-adaptive entropy model to improve the rate-distortion (R-D) performance, which is very slow. In addition, the complexities of the encoding and decoding networks are quite high and not suitable for many practical applications. In this paper, we introduce four techniques to balance the trade-off between the complexity and performance. We are the first to introduce deformable convolutional module in compression framework, which can remove more redundancies in the input image, thereby enhancing compression performance. Second, we design a checkerboard context model with two separate distribution parameter estimation networks and different probability models, which enables parallel decoding without sacrificing the performance compared to the sequential context-adaptive model. Third, we develop an improved three-step knowledge distillation and training scheme to achieve different trade-offs between the complexity and the performance of the decoder network, which transfers both the final and intermediate results of the teacher network to the student network to help its training. Fourth, we introduce $L_{1}$ regularization to make the numerical values of the latent representation more sparse. Then we only encode non-zero channels in the encoding and decoding process, which can greatly reduce the encoding and decoding time. Experiments show that compared to the state-of-the-art learned image coding scheme, our method can be about 20 times faster in encoding and 70-90 times faster in decoding, and our R-D performance is also $2.3 \%$ higher. Our method outperforms the traditional approach in H.266/VVC-intra (4:4:4) and some leading learned schemes in terms of PSNR and MS-SSIM metrics when testing on Kodak and Tecnick-40 datasets.
Cyber-physical systems (CPSs), like train control and management systems (TCMS), are becoming ubiquitous in critical infrastructures. As safety-critical systems, ensuring their dependability during operation is crucial. Digital twins (DTs) have been increasingly studied for this purpose owing to their capability of runtime monitoring and warning, prediction and detection of anomalies, etc. However, constructing a DT for anomaly detection in TCMS necessitates sufficient training data and extracting both chronological and context features with high quality. Hence, in this paper, we propose a novel method named KDDT for TCMS anomaly detection. KDDT harnesses a language model (LM) and a long short-term memory (LSTM) network to extract contexts and chronological features, respectively. To enrich data volume, KDDT benefits from out-of-domain data with knowledge distillation (KD). We evaluated KDDT with two datasets from our industry partner Alstom and obtained the F1 scores of 0.931 and 0.915, respectively, demonstrating the effectiveness of KDDT. We also explored individual contributions of the DT model, LM, and KD to the overall performance of KDDT, via a comprehensive empirical study, and observed average F1 score improvements of 12.4%, 3%, and 6.05%, respectively.
With the increasing deployment of deep neural networks in safety-critical applications such as self-driving cars, medical imaging, anomaly detection, etc., adversarial robustness has become a crucial concern in the reliability of these networks in real-world scenarios. A plethora of works based on adversarial training and regularization-based techniques have been proposed to make these deep networks robust against adversarial attacks. However, these methods require either retraining models or training them from scratch, making them infeasible to defend pre-trained models when access to training data is restricted. To address this problem, we propose a test time Data-free Adversarial Defense (DAD) containing detection and correction frameworks. Moreover, to further improve the efficacy of the correction framework in cases when the detector is under-confident, we propose a soft-detection scheme (dubbed as "DAD++"). We conduct a wide range of experiments and ablations on several datasets and network architectures to show the efficacy of our proposed approach. Furthermore, we demonstrate the applicability of our approach in imparting adversarial defense at test time under data-free (or data-efficient) applications/setups, such as Data-free Knowledge Distillation and Source-free Unsupervised Domain Adaptation, as well as Semi-supervised classification frameworks. We observe that in all the experiments and applications, our DAD++ gives an impressive performance against various adversarial attacks with a minimal drop in clean accuracy. The source code is available at: https://github.com/vcl-iisc/Improved-Data-free-Test-Time-Adversarial-Defense
In this paper, we investigate the task of zero-shot human-object interaction (HOI) detection, a novel paradigm for identifying HOIs without the need for task-specific annotations. To address this challenging task, we employ CLIP, a large-scale pre-trained vision-language model (VLM), for knowledge distillation on multiple levels. Specifically, we design a multi-branch neural network that leverages CLIP for learning HOI representations at various levels, including global images, local union regions encompassing human-object pairs, and individual instances of humans or objects. To train our model, CLIP is utilized to generate HOI scores for both global images and local union regions that serve as supervision signals. The extensive experiments demonstrate the effectiveness of our novel multi-level CLIP knowledge integration strategy. Notably, the model achieves strong performance, which is even comparable with some fully-supervised and weakly-supervised methods on the public HICO-DET benchmark.
Recent years have witnessed the great success of vision transformer (ViT), which has achieved state-of-the-art performance on multiple computer vision benchmarks. However, ViT models suffer from vast amounts of parameters and high computation cost, leading to difficult deployment on resource-constrained edge devices. Existing solutions mostly compress ViT models to a compact model but still cannot achieve real-time inference. To tackle this issue, we propose to explore the divisibility of transformer structure, and decompose the large ViT into multiple small models for collaborative inference at edge devices. Our objective is to achieve fast and energy-efficient collaborative inference while maintaining comparable accuracy compared with large ViTs. To this end, we first propose a collaborative inference framework termed DeViT to facilitate edge deployment by decomposing large ViTs. Subsequently, we design a decomposition-and-ensemble algorithm based on knowledge distillation, termed DEKD, to fuse multiple small decomposed models while dramatically reducing communication overheads, and handle heterogeneous models by developing a feature matching module to promote the imitations of decomposed models from the large ViT. Extensive experiments for three representative ViT backbones on four widely-used datasets demonstrate our method achieves efficient collaborative inference for ViTs and outperforms existing lightweight ViTs, striking a good trade-off between efficiency and accuracy. For example, our DeViTs improves end-to-end latency by 2.89$\times$ with only 1.65% accuracy sacrifice using CIFAR-100 compared to the large ViT, ViT-L/16, on the GPU server. DeDeiTs surpasses the recent efficient ViT, MobileViT-S, by 3.54% in accuracy on ImageNet-1K, while running 1.72$\times$ faster and requiring 55.28% lower energy consumption on the edge device.
We propose EmoDistill, a novel speech emotion recognition (SER) framework that leverages cross-modal knowledge distillation during training to learn strong linguistic and prosodic representations of emotion from speech. During inference, our method only uses a stream of speech signals to perform unimodal SER thus reducing computation overhead and avoiding run-time transcription and prosodic feature extraction errors. During training, our method distills information at both embedding and logit levels from a pair of pre-trained Prosodic and Linguistic teachers that are fine-tuned for SER. Experiments on the IEMOCAP benchmark demonstrate that our method outperforms other unimodal and multimodal techniques by a considerable margin, and achieves state-of-the-art performance of 77.49% unweighted accuracy and 78.91% weighted accuracy. Detailed ablation studies demonstrate the impact of each component of our method.
In this paper, we propose a novel probabilistic self-supervised learning via Scoring Rule Minimization (ProSMIN), which leverages the power of probabilistic models to enhance representation quality and mitigate collapsing representations. Our proposed approach involves two neural networks; the online network and the target network, which collaborate and learn the diverse distribution of representations from each other through knowledge distillation. By presenting the input samples in two augmented formats, the online network is trained to predict the target network representation of the same sample under a different augmented view. The two networks are trained via our new loss function based on proper scoring rules. We provide a theoretical justification for ProSMIN's convergence, demonstrating the strict propriety of its modified scoring rule. This insight validates the method's optimization process and contributes to its robustness and effectiveness in improving representation quality. We evaluate our probabilistic model on various downstream tasks, such as in-distribution generalization, out-of-distribution detection, dataset corruption, low-shot learning, and transfer learning. Our method achieves superior accuracy and calibration, surpassing the self-supervised baseline in a wide range of experiments on large-scale datasets like ImageNet-O and ImageNet-C, ProSMIN demonstrates its scalability and real-world applicability.
Automatic Speech Recognition (ASR) models need to be optimized for specific hardware before they can be deployed on devices. This can be done by tuning the model's hyperparameters or exploring variations in its architecture. Re-training and re-validating models after making these changes can be a resource-intensive task. This paper presents TODM (Train Once Deploy Many), a new approach to efficiently train many sizes of hardware-friendly on-device ASR models with comparable GPU-hours to that of a single training job. TODM leverages insights from prior work on Supernet, where Recurrent Neural Network Transducer (RNN-T) models share weights within a Supernet. It reduces layer sizes and widths of the Supernet to obtain subnetworks, making them smaller models suitable for all hardware types. We introduce a novel combination of three techniques to improve the outcomes of the TODM Supernet: adaptive dropouts, an in-place Alpha-divergence knowledge distillation, and the use of ScaledAdam optimizer. We validate our approach by comparing Supernet-trained versus individually tuned Multi-Head State Space Model (MH-SSM) RNN-T using LibriSpeech. Results demonstrate that our TODM Supernet either matches or surpasses the performance of manually tuned models by up to a relative of 3% better in word error rate (WER), while efficiently keeping the cost of training many models at a small constant.