Out-of-distribution (OOD) detection is essential in identifying test samples that deviate from the in-distribution (ID) data upon which a standard network is trained, ensuring network robustness and reliability. This paper introduces OOD knowledge distillation, a pioneering learning framework applicable whether or not training ID data is available, given a standard network. This framework harnesses unknown OOD-sensitive knowledge from the standard network to craft a certain binary classifier adept at distinguishing between ID and OOD samples. To accomplish this, we introduce Confidence Amendment (CA), an innovative methodology that transforms an OOD sample into an ID one while progressively amending prediction confidence derived from the standard network. This approach enables the simultaneous synthesis of both ID and OOD samples, each accompanied by an adjusted prediction confidence, thereby facilitating the training of a binary classifier sensitive to OOD. Theoretical analysis provides bounds on the generalization error of the binary classifier, demonstrating the pivotal role of confidence amendment in enhancing OOD sensitivity. Extensive experiments spanning various datasets and network architectures confirm the efficacy of the proposed method in detecting OOD samples.
Large-scale pretrained language models have achieved compelling performance in a wide range of language understanding and information retrieval tasks. Knowledge distillation offers an opportunity to compress a large language model to a small one, in order to reach a reasonable latency-performance tradeoff. However, for scenarios where the number of requests (e.g., queries submitted to a search engine) is highly variant, the static tradeoff attained by the compressed language model might not always fit. Once a model is assigned with a static tradeoff, it could be inadequate in that the latency is too high when the number of requests is large or the performance is too low when the number of requests is small. To this end, we propose an elastic language model (ElasticLM) that elastically adjusts the tradeoff according to the request stream. The basic idea is to introduce a compute elasticity to the compressed language model, so that the tradeoff could vary on-the-fly along scalable and controllable compute. Specifically, we impose an elastic structure to enable ElasticLM with compute elasticity and design an elastic optimization to learn ElasticLM under compute elasticity. To serve ElasticLM, we apply an elastic schedule. Considering the specificity of information retrieval, we adapt ElasticLM to dense retrieval and reranking and present ElasticDenser and ElasticRanker respectively. Offline evaluation is conducted on a language understanding benchmark GLUE; and several information retrieval tasks including Natural Question, Trivia QA, and MS MARCO. The results show that ElasticLM along with ElasticDenser and ElasticRanker can perform correctly and competitively compared with an array of static baselines. Furthermore, online simulation with concurrency is also carried out. The results demonstrate that ElasticLM can provide elastic tradeoffs with respect to varying request stream.
Speech data has rich acoustic and paralinguistic information with important cues for understanding a speaker's tone, emotion, and intent, yet traditional large language models such as BERT do not incorporate this information. There has been an increased interest in multi-modal language models leveraging audio and/or visual information and text. However, current multi-modal language models require both text and audio/visual data streams during inference/test time. In this work, we propose a methodology for training language models leveraging spoken language audio data but without requiring the audio stream during prediction time. This leads to an improved language model for analyzing spoken transcripts while avoiding an audio processing overhead at test time. We achieve this via an audio-language knowledge distillation framework, where we transfer acoustic and paralinguistic information from a pre-trained speech embedding (OpenAI Whisper) teacher model to help train a student language model on an audio-text dataset. In our experiments, the student model achieves consistent improvement over traditional language models on tasks analyzing spoken transcripts.
Local features detection and description are widely used in many vision applications with high industrial and commercial demands. With large-scale applications, these tasks raise high expectations for both the accuracy and speed of local features. Most existing studies on local features learning focus on the local descriptions of individual keypoints, which neglect their relationships established from global spatial awareness. In this paper, we present AWDesc with a consistent attention mechanism (CoAM) that opens up the possibility for local descriptors to embrace image-level spatial awareness in both the training and matching stages. For local features detection, we adopt local features detection with feature pyramid to obtain more stable and accurate keypoints localization. For local features description, we provide two versions of AWDesc to cope with different accuracy and speed requirements. On the one hand, we introduceContext Augmentation to address the inherent locality of convolutional neural networks by injecting non-local context information, so that local descriptors can “look wider to describe better”. Specifically, well-designed Adaptive Global Context Augmented Module (AGCA) and Diverse Surrounding Context Augmented Module (DSCA) are proposed to construct robust local descriptors with context information from global to surrounding. On the other hand, we design an extremely lightweight backbone network coupled with the proposed special knowledge distillation strategy to achieve the best trade-off in accuracy and speed. What is more, we perform thorough experiments on image matching, homography estimation, visual localization, and 3D reconstruction tasks, and the results demonstrate that our method surpasses the current state-of-the-art local descriptors. Code is available at: https://github.com/vignywang/AWDesc.
Deep learning-based models are at the forefront of most driver observation benchmarks due to their remarkable accuracies but are also associated with high computational costs. This is challenging, as resources are often limited in real-world driving scenarios. This paper introduces a lightweight framework for resource-efficient driver activity recognition. The framework enhances 3D MobileNet, a neural architecture optimized for speed in video classification, by incorporating knowledge distillation and model quantization to balance model accuracy and computational efficiency. Knowledge distillation helps maintain accuracy while reducing the model size by leveraging soft labels from a larger teacher model (I3D), instead of relying solely on original ground truth data. Model quantization significantly lowers memory and computation demands by using lower precision integers for model weights and activations. Extensive testing on a public dataset for in-vehicle monitoring during autonomous driving demonstrates that this new framework achieves a threefold reduction in model size and a 1.4-fold improvement in inference time, compared to an already optimized architecture. The code for this study is available at https://github.com/calvintanama/qd-driver-activity-reco.
This paper introduces DONUT-hole, a sparse OCR-free visual document understanding (VDU) model that addresses the limitations of its predecessor model, dubbed DONUT. The DONUT model, leveraging a transformer architecture, overcoming the challenges of separate optical character recognition (OCR) and visual semantic understanding (VSU) components. However, its deployment in production environments and edge devices is hindered by high memory and computational demands, particularly in large-scale request services. To overcome these challenges, we propose an optimization strategy based on knowledge distillation and model pruning. Our paradigm to produce DONUT-hole, reduces the model denisty by 54\% while preserving performance. We also achieve a global representational similarity index between DONUT and DONUT-hole based on centered kernel alignment (CKA) metric of 0.79. Moreover, we evaluate the effectiveness of DONUT-hole in the document image key information extraction (KIE) task, highlighting its potential for developing more efficient VDU systems for logistic companies.
Event cameras are gaining popularity due to their unique properties, such as their low latency and high dynamic range. One task where these benefits can be crucial is real-time object detection. However, RGB detectors still outperform event-based detectors due to the sparsity of the event data and missing visual details. In this paper, we develop a novel knowledge distillation approach to shrink the performance gap between these two modalities. To this end, we propose a cross-modality object detection distillation method that by design can focus on regions where the knowledge distillation works best. We achieve this by using an object-centric slot attention mechanism that can iteratively decouple features maps into object-centric features and corresponding pixel-features used for distillation. We evaluate our novel distillation approach on a synthetic and a real event dataset with aligned grayscale images as a teacher modality. We show that object-centric distillation allows to significantly improve the performance of the event-based student object detector, nearly halving the performance gap with respect to the teacher.
Pre-trained language models (PLMs) have recently shown great success in text representation field. However, the high computational cost and high-dimensional representation of PLMs pose significant challenges for practical applications. To make models more accessible, an effective method is to distill large models into smaller representation models. In order to relieve the issue of performance degradation after distillation, we propose a novel Knowledge Distillation method called IBKD. This approach is motivated by the Information Bottleneck principle and aims to maximize the mutual information between the final representation of the teacher and student model, while simultaneously reducing the mutual information between the student model's representation and the input data. This enables the student model to preserve important learned information while avoiding unnecessary information, thus reducing the risk of over-fitting. Empirical studies on two main downstream applications of text representation (Semantic Textual Similarity and Dense Retrieval tasks) demonstrate the effectiveness of our proposed approach.
Autonomous systems are increasingly implemented using end-to-end learning-based controllers. Such controllers make decisions that are executed on the real system, with images as one of the primary sensing modalities. Deep neural networks form a fundamental building block of such controllers. Unfortunately, the existing neural-network verification tools do not scale to inputs with thousands of dimensions -- especially when the individual inputs (such as pixels) are devoid of clear physical meaning. This paper takes a step towards connecting exhaustive closed-loop verification with high-dimensional controllers. Our key insight is that the behavior of a high-dimensional controller can be approximated with several low-dimensional controllers. To balance the approximation accuracy and verifiability of our low-dimensional controllers, we leverage the latest verification-aware knowledge distillation. Then, we inflate low-dimensional reachability results with statistical approximation errors, yielding a high-confidence reachability guarantee for the high-dimensional controller. We investigate two inflation techniques -- based on trajectories and control actions -- both of which show convincing performance in three OpenAI gym benchmarks.
Feature-based knowledge distillation has been applied to compress modern recommendation models, usually with projectors that align student (small) recommendation models' dimensions with teacher dimensions. However, existing studies have only focused on making the projected features (i.e., student features after projectors) similar to teacher features, overlooking investigating whether the user preference can be transferred to student features (i.e., student features before projectors) in this manner. In this paper, we find that due to the lack of restrictions on projectors, the process of transferring user preferences will likely be interfered with. We refer to this phenomenon as preference inconsistency. It greatly wastes the power of feature-based knowledge distillation. To mitigate preference inconsistency, we propose PCKD, which consists of two regularization terms for projectors. We also propose a hybrid method that combines the two regularization terms. We focus on items with high preference scores and significantly mitigate preference inconsistency, improving the performance of feature-based knowledge distillation. Extensive experiments on three public datasets and three backbones demonstrate the effectiveness of PCKD. The code of our method is provided in https://github.com/woriazzc/KDs.
Deep neural networks (DNNs) have improved NLP tasks significantly, but training and maintaining such networks could be costly. Model compression techniques, such as, knowledge distillation (KD), have been proposed to address the issue; however, the compression process could be lossy. Motivated by this, our work investigates how a distilled student model differs from its teacher, if the distillation process causes any information losses, and if the loss follows a specific pattern. Our experiments aim to shed light on the type of tasks might be less or more sensitive to KD by reporting data points on the contribution of different factors, such as the number of layers or attention heads. Results such as ours could be utilized when determining effective and efficient configurations to achieve optimal information transfers between larger (teacher) and smaller (student) models.
Multi-task partially annotated data where each data point is annotated for only a single task are potentially helpful for data scarcity if a network can leverage the inter-task relationship. In this paper, we study the joint learning of object detection and semantic segmentation, the two most popular vision problems, from multi-task data with partial annotations. Extensive experiments are performed to evaluate each task performance and explore their complementarity when a multi-task network cannot optimize both tasks simultaneously. We propose employing knowledge distillation to leverage joint-task optimization. The experimental results show favorable results for multi-task learning and knowledge distillation over single-task learning and even full supervision scenario. All code and data splits are available at https://github.com/lhoangan/multas
Building extraction $-$ needed for inventory management and planning of urban environment $-$ is affected by the misalignment between labels and off-nadir source imagery in training data. Teacher-Student learning of noise-tolerant convolutional neural networks (CNNs) is the existing solution, but the Student networks typically have lower accuracy and cannot surpass the Teacher's performance. This paper proposes a supervised domain adaptation (SDA) of encoder-decoder networks (EDNs) between noisy and clean datasets to tackle the problem. EDNs are configured with high-performing lightweight encoders such as EfficientNet, ResNeSt, and MobileViT. The proposed method is compared against the existing Teacher-Student learning methods like knowledge distillation (KD) and deep mutual learning (DML) with three newly developed datasets. The methods are evaluated for different urban buildings (low-rise, mid-rise, high-rise, and skyscrapers), where misalignment increases with the increase in building height and spatial resolution. For a robust experimental design, 43 lightweight CNNs, five optimisers, nine loss functions, and seven EDNs are benchmarked to obtain the best-performing EDN for SDA. The SDA of the best-performing EDN from our study significantly outperformed KD and DML with up to 0.943, 0.868, 0.912, and 0.697 F1 scores in the low-rise, mid-rise, high-rise, and skyscrapers respectively. The proposed method and the experimental findings will be beneficial in training robust CNNs for building extraction.
Denoising Diffusion models have exhibited remarkable capabilities in image generation. However, generating high-quality samples requires a large number of iterations. Knowledge distillation for diffusion models is an effective method to address this limitation with a shortened sampling process but causes degraded generative quality. Based on our analysis with bias-variance decomposition and experimental observations, we attribute the degradation to the spatial fitting error occurring in the training of both the teacher and student model. Accordingly, we propose $\textbf{S}$patial $\textbf{F}$itting-$\textbf{E}$rror $\textbf{R}$eduction $\textbf{D}$istillation model ($\textbf{SFERD}$). SFERD utilizes attention guidance from the teacher model and a designed semantic gradient predictor to reduce the student's fitting error. Empirically, our proposed model facilitates high-quality sample generation in a few function evaluations. We achieve an FID of 5.31 on CIFAR-10 and 9.39 on ImageNet 64$\times$64 with only one step, outperforming existing diffusion methods. Our study provides a new perspective on diffusion distillation by highlighting the intrinsic denoising ability of models. Project link: \url{https://github.com/Sainzerjj/SFERD}.
Self-supervised foundation models have shown great potential in computer vision thanks to the pre-training paradigm of masked autoencoding. Scale is a primary factor influencing the performance of these foundation models. However, these large foundation models often result in high computational cost. This paper focuses on pre-training relatively small vision transformer models that could be efficiently adapted to downstream tasks. Specifically, taking inspiration from knowledge distillation in model compression, we propose a new asymmetric masked distillation (AMD) framework for pre-training relatively small models with autoencoding. The core of AMD is to devise an asymmetric masking strategy, where the teacher model is enabled to see more context information with a lower masking ratio, while the student model is still equipped with a high masking ratio. We design customized multi-layer feature alignment between the teacher encoder and student encoder to regularize the pre-training of student MAE. To demonstrate the effectiveness and versatility of AMD, we apply it to both ImageMAE and VideoMAE for pre-training relatively small ViT models. AMD achieved 84.6% classification accuracy on IN1K using the ViT-B model. And AMD achieves 73.3% classification accuracy using the ViT-B model on the Something-in-Something V2 dataset, a 3.7% improvement over the original ViT-B model from VideoMAE. We also transfer AMD pre-trained models to downstream tasks and obtain consistent performance improvement over the original masked autoencoding. The code and models are available at https://github.com/MCG-NJU/AMD.
In the era of large scale pretrained models, Knowledge Distillation (KD) serves an important role in transferring the wisdom of computationally heavy teacher models to lightweight, efficient student models while preserving performance. Traditional KD paradigms, however, assume readily available access to teacher models for frequent inference -- a notion increasingly at odds with the realities of costly, often proprietary, large scale models. Addressing this gap, our paper considers how to minimize the dependency on teacher model inferences in KD in a setting we term Few Teacher Inference Knowledge Distillation (FTI KD). We observe that prevalent KD techniques and state of the art data augmentation strategies fall short in this constrained setting. Drawing inspiration from educational principles that emphasize learning through comparison, we propose Comparative Knowledge Distillation (CKD), which encourages student models to understand the nuanced differences in a teacher model's interpretations of samples. Critically, CKD provides additional learning signals to the student without making additional teacher calls. We also extend the principle of CKD to groups of samples, enabling even more efficient learning from limited teacher calls. Empirical evaluation across varied experimental settings indicates that CKD consistently outperforms state of the art data augmentation and KD techniques.
Knowledge Distillation (KD) compresses computationally expensive pre-trained language models (PLMs) by transferring their knowledge to smaller models, allowing their use in resource-constrained or real-time settings. However, most smaller models fail to surpass the performance of the original larger model, resulting in sacrificing performance to improve inference speed. To address this issue, we propose Co-Training and Co-Distillation (CTCD), a novel framework that improves performance and inference speed together by co-training two models while mutually distilling knowledge. The CTCD framework successfully achieves this based on two significant findings: 1) Distilling knowledge from the smaller model to the larger model during co-training improves the performance of the larger model. 2) The enhanced performance of the larger model further boosts the performance of the smaller model. The CTCD framework shows promise as it can be combined with existing techniques like architecture design or data augmentation, replacing one-way KD methods, to achieve further performance improvement. Extensive ablation studies demonstrate the effectiveness of CTCD, and the small model distilled by CTCD outperforms the original larger model by a significant margin of 1.66 on the GLUE benchmark.
The conventional few-shot classification aims at learning a model on a large labeled base dataset and rapidly adapting to a target dataset that is from the same distribution as the base dataset. However, in practice, the base and the target datasets of few-shot classification are usually from different domains, which is the problem of cross-domain few-shot classification. We tackle this problem by making a small proportion of unlabeled images in the target domain accessible in the training stage. In this setup, even though the base data are sufficient and labeled, the large domain shift still makes transferring the knowledge from the base dataset difficult. We meticulously design a cross-level knowledge distillation method, which can strengthen the ability of the model to extract more discriminative features in the target dataset by guiding the network's shallow layers to learn higher-level information. Furthermore, in order to alleviate the overfitting in the evaluation stage, we propose a feature denoising operation which can reduce the feature redundancy and mitigate overfitting. Our approach can surpass the previous state-of-the-art method, Dynamic-Distillation, by 5.44% on 1-shot and 1.37% on 5-shot classification tasks on average in the BSCD-FSL benchmark. The implementation code will be available at https://github.com/jarucezh/cldfd.
Underwater docking is critical to enable the persistent operation of Autonomous Underwater Vehicles (AUVs). For this, the AUV must be capable of detecting and localizing the docking station, which is complex due to the highly dynamic undersea environment. Image-based solutions offer a high acquisition rate and versatile alternative to adapt to this environment; however, the underwater environment presents challenges such as low visibility, high turbidity, and distortion. In addition to this, field experiments to validate underwater docking capabilities can be costly and dangerous due to the specialized equipment and safety considerations required to conduct the experiments. This work compares different deep-learning architectures to perform underwater docking detection and classification. The architecture with the best performance is then compressed using knowledge distillation under the teacher-student paradigm to reduce the network's memory footprint, allowing real-time implementation. To reduce the simulation-to-reality gap, a Generative Adversarial Network (GAN) is used to do image-to-image translation, converting the Gazebo simulation image into a realistic underwater-looking image. The obtained image is then processed using an underwater image formation model to simulate image attenuation over distance under different water types. The proposed method is finally evaluated according to the AUV docking success rate and compared with classical vision methods. The simulation results show an improvement of 20% in the high turbidity scenarios regardless of the underwater currents. Furthermore, we show the performance of the proposed approach by showing experimental results on the off-the-shelf AUV Iver3.
Data-Free Knowledge Distillation (DFKD) plays a vital role in compressing the model when original training data is unavailable. Previous works for DFKD in NLP mainly focus on distilling encoder-only structures like BERT on classification tasks, which overlook the notable progress of generative language modeling. In this work, we propose a novel DFKD framework, namely DFKD-T$^{3}$, where the pretrained generative language model can also serve as a controllable data generator for model compression. This novel framework DFKD-T$^{3}$ leads to an end-to-end learnable text-to-text framework to transform the general domain corpus to compression-friendly task data, targeting to improve both the \textit{specificity} and \textit{diversity}. Extensive experiments show that our method can boost the distillation performance in various downstream tasks such as sentiment analysis, linguistic acceptability, and information extraction. Furthermore, we show that the generated texts can be directly used for distilling other language models and outperform the SOTA methods, making our method more appealing in a general DFKD setting. Our code is available at https://gitee.com/mindspore/models/tree/master/research/nlp/DFKD\_T3.