Leveraging the capabilities of Knowledge Distillation (KD) strategies, we devise a strategy to fight the recent retraction of face recognition datasets. Given a pretrained Teacher model trained on a real dataset, we show that carefully utilising synthetic datasets, or a mix between real and synthetic datasets to distil knowledge from this teacher to smaller students can yield surprising results. In this sense, we trained 33 different models with and without KD, on different datasets, with different architectures and losses. And our findings are consistent, using KD leads to performance gains across all ethnicities and decreased bias. In addition, it helps to mitigate the performance gap between real and synthetic datasets. This approach addresses the limitations of synthetic data training, improving both the accuracy and fairness of face recognition models.
In this paper, we propose HiTSR, a hierarchical transformer model for reference-based image super-resolution, which enhances low-resolution input images by learning matching correspondences from high-resolution reference images. Diverging from existing multi-network, multi-stage approaches, we streamline the architecture and training pipeline by incorporating the double attention block from GAN literature. Processing two visual streams independently, we fuse self-attention and cross-attention blocks through a gating attention strategy. The model integrates a squeeze-and-excitation module to capture global context from the input images, facilitating long-range spatial interactions within window-based attention blocks. Long skip connections between shallow and deep layers further enhance information flow. Our model demonstrates superior performance across three datasets including SUN80, Urban100, and Manga109. Specifically, on the SUN80 dataset, our model achieves PSNR/SSIM values of 30.24/0.821. These results underscore the effectiveness of attention mechanisms in reference-based image super-resolution. The transformer-based model attains state-of-the-art results without the need for purpose-built subnetworks, knowledge distillation, or multi-stage training, emphasizing the potency of attention in meeting reference-based image super-resolution requirements.
For visual recognition, knowledge distillation typically involves transferring knowledge from a large, well-trained teacher model to a smaller student model. In this paper, we introduce an effective method to distill knowledge from an off-the-shelf vision-language model (VLM), demonstrating that it provides novel supervision in addition to those from a conventional vision-only teacher model. Our key technical contribution is the development of a framework that generates novel text supervision and distills free-form text into a vision encoder. We showcase the effectiveness of our approach, termed VLM-KD, across various benchmark datasets, showing that it surpasses several state-of-the-art long-tail visual classifiers. To our knowledge, this work is the first to utilize knowledge distillation with text supervision generated by an off-the-shelf VLM and apply it to vanilla randomly initialized vision encoders.
This survey paper delves into the emerging and critical area of symbolic knowledge distillation in Large Language Models (LLMs). As LLMs like Generative Pre-trained Transformer-3 (GPT-3) and Bidirectional Encoder Representations from Transformers (BERT) continue to expand in scale and complexity, the challenge of effectively harnessing their extensive knowledge becomes paramount. This survey concentrates on the process of distilling the intricate, often implicit knowledge contained within these models into a more symbolic, explicit form. This transformation is crucial for enhancing the interpretability, efficiency, and applicability of LLMs. We categorize the existing research based on methodologies and applications, focusing on how symbolic knowledge distillation can be used to improve the transparency and functionality of smaller, more efficient Artificial Intelligence (AI) models. The survey discusses the core challenges, including maintaining the depth of knowledge in a comprehensible format, and explores the various approaches and techniques that have been developed in this field. We identify gaps in current research and potential opportunities for future advancements. This survey aims to provide a comprehensive overview of symbolic knowledge distillation in LLMs, spotlighting its significance in the progression towards more accessible and efficient AI systems.
As in school, one teacher to cover all subjects is insufficient to distill equally robust information to a student. Hence, each subject is taught by a highly specialised teacher. Following a similar philosophy, we propose a multiple specialized teacher framework to distill knowledge to a student network. In our approach, directed at face recognition use cases, we train four teachers on one specific ethnicity, leading to four highly specialized and biased teachers. Our strategy learns a project of these four teachers into a common space and distill that information to a student network. Our results highlighted increased performance and reduced bias for all our experiments. In addition, we further show that having biased/specialized teachers is crucial by showing that our approach achieves better results than when knowledge is distilled from four teachers trained on balanced datasets. Our approach represents a step forward to the understanding of the importance of ethnicity-specific features.
Training on high-quality synthetic data from strong language models (LMs) is a common strategy to improve the reasoning performance of LMs. In this work, we revisit whether this strategy is compute-optimal under a fixed inference budget (e.g., FLOPs). To do so, we investigate the trade-offs between generating synthetic data using a stronger but more expensive (SE) model versus a weaker but cheaper (WC) model. We evaluate the generated data across three key metrics: coverage, diversity, and false positive rate, and show that the data from WC models may have higher coverage and diversity, but also exhibit higher false positive rates. We then finetune LMs on data from SE and WC models in different settings: knowledge distillation, self-improvement, and a novel weak-to-strong improvement setup where a weaker LM teaches reasoning to a stronger LM. Our findings reveal that models finetuned on WC-generated data consistently outperform those trained on SE-generated data across multiple benchmarks and multiple choices of WC and SE models. These results challenge the prevailing practice of relying on SE models for synthetic data generation, suggesting that WC may be the compute-optimal approach for training advanced LM reasoners.
Lossless speculative decoding accelerates target large language model (LLM) inference by employing a lightweight draft model for generating tree-structured candidates, which are subsequently verified in parallel by the target LLM. Currently, effective approaches leverage feature-level rather than token-level autoregression within the draft model to facilitate more straightforward predictions and enhanced knowledge distillation. In this paper, we reassess these approaches and propose FSPAD (Feature Sampling and Partial Alignment Distillation for Lossless Speculative Decoding), which introduces two straightforward and effective components within the existing framework to boost lossless speculative decoding. Firstly, FSPAD utilizes token embeddings to sample features of the target LLM in high-dimensional space before feeding them into the draft model, due to the inherent uncertainty of the features preventing the draft model from obtaining the specific token output by the target LLM. Secondly, FSPAD introduces partial alignment distillation to weaken the draft model's connection between features and logits, aiming to reduce the conflict between feature alignment and logit confidence during training. Our experiments include both greedy and non-greedy decoding on the largest and smallest models from the Vicuna and LLaMA3-Instruct series, as well as tasks in multi-turn conversation, translation, summarization, question answering, mathematical reasoning, and retrieval-augmented generation. The results show that FSPAD outperforms the state-of-the-art method across all the aforementioned tasks and target LLMs.
We introduce LLaVA-MoD, a novel framework designed to enable the efficient training of small-scale Multimodal Language Models (s-MLLM) by distilling knowledge from large-scale MLLM (l-MLLM). Our approach tackles two fundamental challenges in MLLM distillation. First, we optimize the network structure of s-MLLM by integrating a sparse Mixture of Experts (MoE) architecture into the language model, striking a balance between computational efficiency and model expressiveness. Second, we propose a progressive knowledge transfer strategy to ensure comprehensive knowledge migration. This strategy begins with mimic distillation, where we minimize the Kullback-Leibler (KL) divergence between output distributions to enable the student model to emulate the teacher network's understanding. Following this, we introduce preference distillation via Direct Preference Optimization (DPO), where the key lies in treating l-MLLM as the reference model. During this phase, the s-MLLM's ability to discriminate between superior and inferior examples is significantly enhanced beyond l-MLLM, leading to a better student that surpasses its teacher, particularly in hallucination benchmarks. Extensive experiments demonstrate that LLaVA-MoD outperforms existing models across various multimodal benchmarks while maintaining a minimal number of activated parameters and low computational costs. Remarkably, LLaVA-MoD, with only 2B activated parameters, surpasses Qwen-VL-Chat-7B by an average of 8.8% across benchmarks, using merely 0.3% of the training data and 23% trainable parameters. These results underscore LLaVA-MoD's ability to effectively distill comprehensive knowledge from its teacher model, paving the way for the development of more efficient MLLMs. The code will be available on: https://github.com/shufangxun/LLaVA-MoD.
In this study, we investigate the impact of online pre-training with continuous video clips. We will examine three methods for pre-training (masked image modeling, contrastive learning, and knowledge distillation), and assess the performance on downstream action recognition tasks. As a result, online pre-training with contrast learning showed the highest performance in downstream tasks. Our findings suggest that learning from long-form videos can be helpful for action recognition with short videos.
Multimodal Federated Learning frequently encounters challenges of client modality heterogeneity, leading to undesired performances for secondary modality in multimodal learning. It is particularly prevalent in audiovisual learning, with audio is often assumed to be the weaker modality in recognition tasks. To address this challenge, we introduce ModalityMirror to improve audio model performance by leveraging knowledge distillation from an audiovisual federated learning model. ModalityMirror involves two phases: a modality-wise FL stage to aggregate uni-modal encoders; and a federated knowledge distillation stage on multi-modality clients to train an unimodal student model. Our results demonstrate that ModalityMirror significantly improves the audio classification compared to the state-of-the-art FL methods such as Harmony, particularly in audiovisual FL facing video missing. Our approach unlocks the potential for exploiting the diverse modality spectrum inherent in multi-modal FL.
Knowledge Distillation (KD) is a powerful approach for compressing a large model into a smaller, more efficient model, particularly beneficial for latency-sensitive applications like recommender systems. However, current KD research predominantly focuses on Computer Vision (CV) and NLP tasks, overlooking unique data characteristics and challenges inherent to recommender systems. This paper addresses these overlooked challenges, specifically: (1) mitigating data distribution shifts between teacher and student models, (2) efficiently identifying optimal teacher configurations within time and budgetary constraints, and (3) enabling computationally efficient and rapid sharing of teacher labels to support multiple students. We present a robust KD system developed and rigorously evaluated on multiple large-scale personalized video recommendation systems within Google. Our live experiment results demonstrate significant improvements in student model performance while ensuring consistent and reliable generation of high quality teacher labels from a continuous data stream of data.
Smaller machine learning models, with less complex architectures and sensor inputs, can benefit wearable sensor-based human activity recognition (HAR) systems in many ways, from complexity and cost to battery life. In the specific case of smart factories, optimizing human-robot collaboration hinges on the implementation of cutting-edge, human-centric AI systems. To this end, workers' activity recognition enables accurate quantification of performance metrics, improving efficiency holistically. We present a two-stage semantic-aware knowledge distillation (KD) approach, TSAK, for efficient, privacy-aware, and wearable HAR in manufacturing lines, which reduces the input sensor modalities as well as the machine learning model size, while reaching similar recognition performance as a larger multi-modal and multi-positional teacher model. The first stage incorporates a teacher classifier model encoding attention, causal, and combined representations. The second stage encompasses a semantic classifier merging the three representations from the first stage. To evaluate TSAK, we recorded a multi-modal dataset at a smart factory testbed with wearable and privacy-aware sensors (IMU and capacitive) located on both workers' hands. In addition, we evaluated our approach on OpenPack, the only available open dataset mimicking the wearable sensor placements on both hands in the manufacturing HAR scenario. We compared several KD strategies with different representations to regulate the training process of a smaller student model. Compared to the larger teacher model, the student model takes fewer sensor channels from a single hand, has 79% fewer parameters, runs 8.88 times faster, and requires 96.6% less computing power (FLOPS).
Deep learning has excelled in image recognition tasks through neural networks inspired by the human brain. However, the necessity for large models to improve prediction accuracy introduces significant computational demands and extended training times.Conventional methods such as fine-tuning, knowledge distillation, and pruning have the limitations like potential accuracy drops. Drawing inspiration from human neurogenesis, where neuron formation continues into adulthood, we explore a novel approach of progressively increasing neuron numbers in compact models during training phases, thereby managing computational costs effectively. We propose a method that reduces feature extraction biases and neuronal redundancy by introducing constraints based on neuron similarity distributions. This approach not only fosters efficient learning in new neurons but also enhances feature extraction relevancy for given tasks. Results on CIFAR-10 and CIFAR-100 datasets demonstrated accuracy improvement, and our method pays more attention to whole object to be classified in comparison with conventional method through Grad-CAM visualizations. These results suggest that our method's potential to decision-making processes.
Recent studies showcase the competitive accuracy of Vision Transformers (ViTs) in relation to Convolutional Neural Networks (CNNs), along with their remarkable robustness. However, ViTs demand a large amount of data to achieve adequate performance, which makes their application to small datasets challenging, falling behind CNNs. To overcome this, we propose GenFormer, a data augmentation strategy utilizing generated images, thereby improving transformer accuracy and robustness on small-scale image classification tasks. In our comprehensive evaluation we propose Tiny ImageNetV2, -R, and -A as new test set variants of Tiny ImageNet by transferring established ImageNet generalization and robustness benchmarks to the small-scale data domain. Similarly, we introduce MedMNIST-C and EuroSAT-C as corrupted test set variants of established fine-grained datasets in the medical and aerial domain. Through a series of experiments conducted on small datasets of various domains, including Tiny ImageNet, CIFAR, EuroSAT and MedMNIST datasets, we demonstrate the synergistic power of our method, in particular when combined with common train and test time augmentations, knowledge distillation, and architectural design choices. Additionally, we prove the effectiveness of our approach under challenging conditions with limited training data, demonstrating significant improvements in both accuracy and robustness, bridging the gap between CNNs and ViTs in the small-scale dataset domain.
AI-assisted lesion detection models play a crucial role in the early screening of cancer. However, previous image-based models ignore the inter-frame contextual information present in videos. On the other hand, video-based models capture the inter-frame context but are computationally expensive. To mitigate this contradiction, we delve into Video-to-Image knowledge distillation leveraging DEtection TRansformer (V2I-DETR) for the task of medical video lesion detection. V2I-DETR adopts a teacher-student network paradigm. The teacher network aims at extracting temporal contexts from multiple frames and transferring them to the student network, and the student network is an image-based model dedicated to fast prediction in inference. By distilling multi-frame contexts into a single frame, the proposed V2I-DETR combines the advantages of utilizing temporal contexts from video-based models and the inference speed of image-based models. Through extensive experiments, V2I-DETR outperforms previous state-of-the-art methods by a large margin while achieving the real-time inference speed (30 FPS) as the image-based model.
Due to the high complexity and technical requirements of industrial production processes, surface defects will inevitably appear, which seriously affects the quality of products. Although existing lightweight detection networks are highly efficient, they are susceptible to false or missed detection of non-salient defects due to the lack of semantic information. In contrast, the diffusion model can generate higher-order semantic representations in the denoising process. Therefore, the aim of this paper is to incorporate the higher-order modelling capability of the diffusion model into the detection model, so as to better assist in the classification and localization of difficult targets. First, the denoising diffusion probabilistic model (DDPM) is pre-trained to extract the features of denoising process to construct as a feature repository. In particular, to avoid the potential bottleneck of memory caused by the dataloader loading high-dimensional features, a residual convolutional variational auto-encoder (ResVAE) is designed to further compress the feature repository. The image is fed into both image backbone and feature repository for feature extraction and querying respectively. The queried latent features are reconstructed and filtered to obtain high-dimensional DDPM features. A dynamic cross-fusion method is proposed to fully refine the contextual features of DDPM to optimize the detection model. Finally, we employ knowledge distillation to migrate the higher-order modelling capabilities back into the lightweight baseline model without additional efficiency cost. Experiment results demonstrate that our method achieves competitive results on several industrial datasets.
Knowledge distillation (KD) is a key element in neural network compression that allows knowledge transfer from a pre-trained teacher model to a more compact student model. KD relies on access to the training dataset, which may not always be fully available due to privacy concerns or logistical issues related to the size of the data. To address this, "data-free" KD methods use synthetic data, generated through model inversion, to mimic the target data distribution. However, conventional model inversion methods are not designed to utilize supplementary information from the target dataset, and thus, cannot leverage it to improve performance, even when it is available. In this paper, we consider condensed samples, as a form of supplementary information, and introduce a method for using them to better approximate the target data distribution, thereby enhancing the KD performance. Our approach is versatile, evidenced by improvements of up to 11.4% in KD accuracy across various datasets and model inversion-based methods. Importantly, it remains effective even when using as few as one condensed sample per class, and can also enhance performance in few-shot scenarios where only limited real data samples are available.
Semiconductor imaging and analysis are critical yet understudied in deep learning, limiting our ability for precise control and optimization in semiconductor manufacturing. We introduce a small-scale multimodal framework for analyzing semiconductor electron microscopy images (MAEMI) through vision-language instruction tuning. We generate a customized instruction-following dataset using large multimodal models on microscopic image analysis. We perform knowledge transfer from larger to smaller models through knowledge distillation, resulting in improved accuracy of smaller models on visual question answering (VQA) tasks. This approach eliminates the need for expensive, human expert-annotated datasets for microscopic image analysis tasks. Enterprises can further finetune MAEMI on their intellectual data, enhancing privacy and performance on low-cost consumer hardware. Our experiments show that MAEMI outperforms traditional methods, adapts to data distribution shifts, and supports high-throughput screening.
Multi-label class-incremental learning (MLCIL) is essential for real-world multi-label applications, allowing models to learn new labels while retaining previously learned knowledge continuously. However, recent MLCIL approaches can only achieve suboptimal performance due to the oversight of the positive-negative imbalance problem, which manifests at both the label and loss levels because of the task-level partial label issue. The imbalance at the label level arises from the substantial absence of negative labels, while the imbalance at the loss level stems from the asymmetric contributions of the positive and negative loss parts to the optimization. To address the issue above, we propose a Rebalance framework for both the Loss and Label levels (RebLL), which integrates two key modules: asymmetric knowledge distillation (AKD) and online relabeling (OR). AKD is proposed to rebalance at the loss level by emphasizing the negative label learning in classification loss and down-weighting the contribution of overconfident predictions in distillation loss. OR is designed for label rebalance, which restores the original class distribution in memory by online relabeling the missing classes. Our comprehensive experiments on the PASCAL VOC and MS-COCO datasets demonstrate that this rebalancing strategy significantly improves performance, achieving new state-of-the-art results even with a vanilla CNN backbone.
Space debris and inactive satellites pose a threat to the safety and integrity of operational spacecraft and motivate the need for space situational awareness techniques. These uncooperative targets create a challenging tracking and detection problem due to a lack of prior knowledge of their features, trajectories, or even existence. Recent advancements in computer vision models can be used to improve upon existing methods for tracking such uncooperative targets to make them more robust and reliable to the wide-ranging nature of the target. This paper introduces an autonomous detection model designed to identify and monitor these objects using learning and computer vision. The autonomous detection method aims to identify and accurately track the uncooperative targets in varied circumstances, including different camera spectral sensitivities, lighting, and backgrounds. Our method adapts to the relative distance between the observing spacecraft and the target, and different detection strategies are adjusted based on distance. At larger distances, we utilize You Only Look Once (YOLOv8), a multitask Convolutional Neural Network (CNN), for zero-shot and domain-specific single-shot real time detection of the target. At shorter distances, we use knowledge distillation to combine visual foundation models with a lightweight fast segmentation CNN (Fast-SCNN) to segment the spacecraft components with low storage requirements and fast inference times, and to enable weight updates from earth and possible onboard training. Lastly, we test our method on a custom dataset simulating the unique conditions encountered in space, as well as a publicly-available dataset.