We propose a novel framework and a solution to tackle the continual learning (CL) problem with changing network architectures. Most CL methods focus on adapting a single architecture to a new task/class by modifying its weights. However, with rapid progress in architecture design, the problem of adapting existing solutions to novel architectures becomes relevant. To address this limitation, we propose Heterogeneous Continual Learning (HCL), where a wide range of evolving network architectures emerge continually together with novel data/tasks. As a solution, we build on top of the distillation family of techniques and modify it to a new setting where a weaker model takes the role of a teacher; meanwhile, a new stronger architecture acts as a student. Furthermore, we consider a setup of limited access to previous data and propose Quick Deep Inversion (QDI) to recover prior task visual features to support knowledge transfer. QDI significantly reduces computational costs compared to previous solutions and improves overall performance. In summary, we propose a new setup for CL with a modified knowledge distillation paradigm and design a quick data inversion method to enhance distillation. Our evaluation of various benchmarks shows a significant improvement on accuracy in comparison to state-of-the-art methods over various networks architectures.
Purpose: Advances in surgical phase recognition are generally led by training deeper networks. Rather than going further with a more complex solution, we believe that current models can be exploited better. We propose a self-knowledge distillation framework that can be integrated into current state-of-the-art (SOTA) models without requiring any extra complexity to the models or annotations. Methods: Knowledge distillation is a framework for network regularization where knowledge is distilled from a teacher network to a student network. In self-knowledge distillation, the student model becomes the teacher such that the network learns from itself. Most phase recognition models follow an encoder-decoder framework. Our framework utilizes self-knowledge distillation in both stages. The teacher model guides the training process of the student model to extract enhanced feature representations from the encoder and build a more robust temporal decoder to tackle the over-segmentation problem. Results: We validate our proposed framework on the public dataset Cholec80. Our framework is embedded on top of four popular SOTA approaches and consistently improves their performance. Specifically, our best GRU model boosts performance by +3.33% accuracy and +3.95% F1-score over the same baseline model. Conclusion: We embed a self-knowledge distillation framework for the first time in the surgical phase recognition training pipeline. Experimental results demonstrate that our simple yet powerful framework can improve performance of existing phase recognition models. Moreover, our extensive experiments show that even with 75% of the training set we still achieve performance on par with the same baseline model trained on the full set.
End-to-end (E2E) systems have shown comparable performance to hybrid systems for automatic speech recognition (ASR). Word timings, as a by-product of ASR, are essential in many applications, especially for subtitling and computer-aided pronunciation training. In this paper, we improve the frame-level classifier for word timings in E2E system by introducing label priors in connectionist temporal classification (CTC) loss, which is adopted from prior works, and combining low-level Mel-scale filter banks with high-level ASR encoder output as input feature. On the internal Chinese corpus, the proposed method achieves 95.68%/94.18% compared to the hybrid system 93.0%/90.22% on the word timing accuracy metrics. It also surpass a previous E2E approach with an absolute increase of 4.80%/8.02% on the metrics on 7 languages. In addition, we further improve word timing accuracy by delaying CTC peaks with frame-wise knowledge distillation, though only experimenting on LibriSpeech.
This paper explores the tasks of leveraging auxiliary modalities which are only available at training to enhance multimodal representation learning through cross-modal Knowledge Distillation (KD). The widely adopted mutual information maximization-based objective leads to a short-cut solution of the weak teacher, i.e., achieving the maximum mutual information by simply making the teacher model as weak as the student model. To prevent such a weak solution, we introduce an additional objective term, i.e., the mutual information between the teacher and the auxiliary modality model. Besides, to narrow down the information gap between the student and teacher, we further propose to minimize the conditional entropy of the teacher given the student. Novel training schemes based on contrastive learning and adversarial learning are designed to optimize the mutual information and the conditional entropy, respectively. Experimental results on three popular multimodal benchmark datasets have shown that the proposed method outperforms a range of state-of-the-art approaches for video recognition, video retrieval and emotion classification.
Contacting customer service via chat is a common practice. Because employing customer service agents is expensive, many companies are turning to NLP that assists human agents by auto-generating responses that can be used directly or with modifications. Large Language Models (LLMs) are a natural fit for this use case; however, their efficacy must be balanced with the cost of training and serving them. This paper assesses the practical cost and impact of LLMs for the enterprise as a function of the usefulness of the responses that they generate. We present a cost framework for evaluating an NLP model's utility for this use case and apply it to a single brand as a case study in the context of an existing agent assistance product. We compare three strategies for specializing an LLM - prompt engineering, fine-tuning, and knowledge distillation - using feedback from the brand's customer service agents. We find that the usability of a model's responses can make up for a large difference in inference cost for our case study brand, and we extrapolate our findings to the broader enterprise space.
Multi-Teacher knowledge distillation provides students with additional supervision from multiple pre-trained teachers with diverse information sources. Most existing methods explore different weighting strategies to obtain a powerful ensemble teacher, while ignoring the student with poor learning ability may not benefit from such specialized integrated knowledge. To address this problem, we propose Adaptive Multi-teacher Knowledge Distillation with Meta-Learning (MMKD) to supervise student with appropriate knowledge from a tailored ensemble teacher. With the help of a meta-weight network, the diverse yet compatible teacher knowledge in the output layer and intermediate layers is jointly leveraged to enhance the student performance. Extensive experiments on multiple benchmark datasets validate the effectiveness and flexibility of our methods. Code is available: https://github.com/Rorozhl/MMKD.
This paper presents our participation in the FinNLP-2023 shared task on multi-lingual environmental, social, and corporate governance issue identification (ML-ESG). The task's objective is to classify news articles based on the 35 ESG key issues defined by the MSCI ESG rating guidelines. Our approach focuses on the English and French subtasks, employing the CerebrasGPT, OPT, and Pythia models, along with the zero-shot and GPT3Mix Augmentation techniques. We utilize various encoder models, such as RoBERTa, DeBERTa, and FinBERT, subjecting them to knowledge distillation and additional training. Our approach yielded exceptional results, securing the first position in the English text subtask with F1-score 0.69 and the second position in the French text subtask with F1-score 0.78. These outcomes underscore the effectiveness of our methodology in identifying ESG issues in news articles across different languages. Our findings contribute to the exploration of ESG topics and highlight the potential of leveraging advanced language models for ESG issue identification.
Currently, the reduction in the parameter scale of large-scale pre-trained language models (PLMs) through knowledge distillation has greatly facilitated their widespread deployment on various devices. However, the deployment of knowledge distillation systems faces great challenges in real-world industrial-strength applications, which require the use of complex distillation methods on even larger-scale PLMs (over 10B), limited by memory on GPUs and the switching of methods. To overcome these challenges, we propose GKD, a general knowledge distillation framework that supports distillation on larger-scale PLMs using various distillation methods. With GKD, developers can build larger distillation models on memory-limited GPUs and easily switch and combine different distillation methods within a single framework. Experimental results show that GKD can support the distillation of at least 100B-scale PLMs and 25 mainstream methods on 8 NVIDIA A100 (40GB) GPUs.
The large scale of pre-trained language models poses a challenge for their deployment on various devices, with a growing emphasis on methods to compress these models, particularly knowledge distillation. However, current knowledge distillation methods rely on the model's intermediate layer features and the golden labels (also called hard labels), which usually require aligned model architecture and enough labeled data respectively. Moreover, the parameters of vocabulary are usually neglected in existing methods. To address these problems, we propose a general language model distillation (GLMD) method that performs two-stage word prediction distillation and vocabulary compression, which is simple and surprisingly shows extremely strong performance. Specifically, GLMD supports more general application scenarios by eliminating the constraints of dimension and structure between models and the need for labeled datasets through the absence of intermediate layers and golden labels. Meanwhile, based on the long-tailed distribution of word frequencies in the data, GLMD designs a strategy of vocabulary compression through decreasing vocabulary size instead of dimensionality. Experimental results show that our method outperforms 25 state-of-the-art methods on the SuperGLUE benchmark, achieving an average score that surpasses the best method by 3%.
Web applications where users are presented with a limited selection of items have long employed ranking models to put the most relevant results first. Any feedback received from users is typically assumed to reflect a relative judgement on the utility of items, e.g. a user clicking on an item only implies it is better than items not clicked in the same ranked list. Hence, the objectives optimized in Learning-to-Rank (LTR) tend to be pairwise or listwise. Yet, by only viewing feedback as relative, we neglect the user's absolute feedback on the list's overall quality, e.g. when no items in the selection are clicked. We thus reconsider the standard LTR paradigm and argue the benefits of learning from this listwide signal. To this end, we propose the RankFormer as an architecture that, with a Transformer at its core, can jointly optimize a novel listwide assessment objective and a traditional listwise LTR objective. We simulate implicit feedback on public datasets and observe that the RankFormer succeeds in benefitting from listwide signals. Additionally, we conduct experiments in e-commerce on Amazon Search data and find the RankFormer to be superior to all baselines offline. An online experiment shows that knowledge distillation can be used to find immediate practical use for the RankFormer.
Diffusion models have demonstrated excellent potential for generating diverse images. However, their performance often suffers from slow generation due to iterative denoising. Knowledge distillation has been recently proposed as a remedy that can reduce the number of inference steps to one or a few without significant quality degradation. However, existing distillation methods either require significant amounts of offline computation for generating synthetic training data from the teacher model or need to perform expensive online learning with the help of real data. In this work, we present a novel technique called BOOT, that overcomes these limitations with an efficient data-free distillation algorithm. The core idea is to learn a time-conditioned model that predicts the output of a pre-trained diffusion model teacher given any time step. Such a model can be efficiently trained based on bootstrapping from two consecutive sampled steps. Furthermore, our method can be easily adapted to large-scale text-to-image diffusion models, which are challenging for conventional methods given the fact that the training sets are often large and difficult to access. We demonstrate the effectiveness of our approach on several benchmark datasets in the DDIM setting, achieving comparable generation quality while being orders of magnitude faster than the diffusion teacher. The text-to-image results show that the proposed approach is able to handle highly complex distributions, shedding light on more efficient generative modeling.
Unsupervised person re-identification has achieved great success through the self-improvement of individual neural networks. However, limited by the lack of diversity of discriminant information, a single network has difficulty learning sufficient discrimination ability by itself under unsupervised conditions. To address this limit, we develop a population-based evolutionary gaming (PEG) framework in which a population of diverse neural networks is trained concurrently through selection, reproduction, mutation, and population mutual learning iteratively. Specifically, the selection of networks to preserve is modeled as a cooperative game and solved by the best-response dynamics, then the reproduction and mutation are implemented by cloning and fluctuating hyper-parameters of networks to learn more diversity, and population mutual learning improves the discrimination of networks by knowledge distillation from each other within the population. In addition, we propose a cross-reference scatter (CRS) to approximately evaluate re-ID models without labeled samples and adopt it as the criterion of network selection in PEG. CRS measures a model's performance by indirectly estimating the accuracy of its predicted pseudo-labels according to the cohesion and separation of the feature space. Extensive experiments demonstrate that (1) CRS approximately measures the performance of models without labeled samples; (2) and PEG produces new state-of-the-art accuracy for person re-identification, indicating the great potential of population-based network cooperative training for unsupervised learning.
Knowledge distillation (KD) has received much attention due to its success in compressing networks to allow for their deployment in resource-constrained systems. While the problem of adversarial robustness has been studied before in the KD setting, previous works overlook what we term the relative calibration of the student network with respect to its teacher in terms of soft confidences. In particular, we focus on two crucial questions with regard to a teacher-student pair: (i) do the teacher and student disagree at points close to correctly classified dataset examples, and (ii) is the distilled student as confident as the teacher around dataset examples? These are critical questions when considering the deployment of a smaller student network trained from a robust teacher within a safety-critical setting. To address these questions, we introduce a faithful imitation framework to discuss the relative calibration of confidences and provide empirical and certified methods to evaluate the relative calibration of a student w.r.t. its teacher. Further, to verifiably align the relative calibration incentives of the student to those of its teacher, we introduce faithful distillation. Our experiments on the MNIST, Fashion-MNIST and CIFAR-10 datasets demonstrate the need for such an analysis and the advantages of the increased verifiability of faithful distillation over alternative adversarial distillation methods.
Self-supervised learning (SSL) has emerged as a popular approach for learning audio representations. One goal of audio self-supervised pre-training is to transfer knowledge to downstream audio tasks, generally including clip-level and frame-level tasks. While frame-level tasks are important for fine-grained acoustic scene/event understanding, prior studies primarily evaluate on clip-level downstream tasks. In order to tackle both clip-level and frame-level tasks, this paper proposes Audio Teacher-Student Transformer (ATST), with a clip-level version (named ATST-Clip) and a frame-level version (named ATST-Frame), responsible for learning clip-level and frame-level representations, respectively. Both methods use a Transformer encoder and a teacher-student training scheme. We have carefully designed the view creation strategy for ATST-Clip and ATST-Frame. Specifically, ATST-Clip uses segment-wise data augmentations, and ATST-Frame integrates frame-wise data augmentations and masking. Experimental results show that our ATST-Frame model obtains state-of-the-art (SOTA) performances on most of the clip-level and frame-level downstream tasks. Especially, it outperforms other models by a large margin on the frame-level sound event detection task. In addition, the performance can be further improved by combining the two models through knowledge distillation. Our code is available online.
Pretraining reinforcement learning (RL) models on offline datasets is a promising way to improve their training efficiency in online tasks, but challenging due to the inherent mismatch in dynamics and behaviors across various tasks. We present a model-based RL method that learns to transfer potentially useful dynamics and action demonstrations from offline data to a novel task. The main idea is to use the world models not only as simulators for behavior learning but also as tools to measure the task relevance for both dynamics representation transfer and policy transfer. We build a time-varying, domain-selective distillation loss to generate a set of offline-to-online similarity weights. These weights serve two purposes: (i) adaptively transferring the task-agnostic knowledge of physical dynamics to facilitate world model training, and (ii) learning to replay relevant source actions to guide the target policy. We demonstrate the advantages of our approach compared with the state-of-the-art methods in Meta-World and DeepMind Control Suite.
Passage retrieval is a fundamental task in many information systems, such as web search and question answering, where both efficiency and effectiveness are critical concerns. In recent years, neural retrievers based on pre-trained language models (PLM), such as dual-encoders, have achieved huge success. Yet, studies have found that the performance of dual-encoders are often limited due to the neglecting of the interaction information between queries and candidate passages. Therefore, various interaction paradigms have been proposed to improve the performance of vanilla dual-encoders. Particularly, recent state-of-the-art methods often introduce late-interaction during the model inference process. However, such late-interaction based methods usually bring extensive computation and storage cost on large corpus. Despite their effectiveness, the concern of efficiency and space footprint is still an important factor that limits the application of interaction-based neural retrieval models. To tackle this issue, we incorporate implicit interaction into dual-encoders, and propose I^3 retriever. In particular, our implicit interaction paradigm leverages generated pseudo-queries to simulate query-passage interaction, which jointly optimizes with query and passage encoders in an end-to-end manner. It can be fully pre-computed and cached, and its inference process only involves simple dot product operation of the query vector and passage vector, which makes it as efficient as the vanilla dual encoders. We conduct comprehensive experiments on MSMARCO and TREC2019 Deep Learning Datasets, demonstrating the I^3 retriever's superiority in terms of both effectiveness and efficiency. Moreover, the proposed implicit interaction is compatible with special pre-training and knowledge distillation for passage retrieval, which brings a new state-of-the-art performance.
Pre-trained Transformer models like T5 and BART have advanced the state of the art on a wide range of text generation tasks. Compressing these models into smaller ones has become critically important for practical use. Common neural network compression techniques such as knowledge distillation or quantization are limited to static compression where the compression ratio is fixed. In this paper, we introduce Modular Transformers, a modularized encoder-decoder framework for flexible sequence-to-sequence model compression. Modular Transformers train modularized layers that have the same function of two or more consecutive layers in the original model via module replacing and knowledge distillation. After training, the modularized layers can be flexibly assembled into sequence-to-sequence models that meet different performance-efficiency trade-offs. Experimental results show that after a single training phase, by simply varying the assembling strategy, Modular Transformers can achieve flexible compression ratios from 1.1x to 6x with little to moderate relative performance drop.
Data-free knowledge distillation (KD) helps transfer knowledge from a pre-trained model (known as the teacher model) to a smaller model (known as the student model) without access to the original training data used for training the teacher model. However, the security of the synthetic or out-of-distribution (OOD) data required in data-free KD is largely unknown and under-explored. In this work, we make the first effort to uncover the security risk of data-free KD w.r.t. untrusted pre-trained models. We then propose Anti-Backdoor Data-Free KD (ABD), the first plug-in defensive method for data-free KD methods to mitigate the chance of potential backdoors being transferred. We empirically evaluate the effectiveness of our proposed ABD in diminishing transferred backdoor knowledge while maintaining compatible downstream performances as the vanilla KD. We envision this work as a milestone for alarming and mitigating the potential backdoors in data-free KD. Codes are released at https://github.com/illidanlab/ABD.
Recent research has focused on enhancing the capability of smaller models through imitation learning, drawing on the outputs generated by large foundation models (LFMs). A number of issues impact the quality of these models, ranging from limited imitation signals from shallow LFM outputs; small scale homogeneous training data; and most notably a lack of rigorous evaluation resulting in overestimating the small model's capability as they tend to learn to imitate the style, but not the reasoning process of LFMs. To address these challenges, we develop Orca (We are working with our legal team to publicly release a diff of the model weights in accordance with LLaMA's release policy to be published at https://aka.ms/orca-lm), a 13-billion parameter model that learns to imitate the reasoning process of LFMs. Orca learns from rich signals from GPT-4 including explanation traces; step-by-step thought processes; and other complex instructions, guided by teacher assistance from ChatGPT. To promote this progressive learning, we tap into large-scale and diverse imitation data with judicious sampling and selection. Orca surpasses conventional state-of-the-art instruction-tuned models such as Vicuna-13B by more than 100% in complex zero-shot reasoning benchmarks like Big-Bench Hard (BBH) and 42% on AGIEval. Moreover, Orca reaches parity with ChatGPT on the BBH benchmark and shows competitive performance (4 pts gap with optimized system message) in professional and academic examinations like the SAT, LSAT, GRE, and GMAT, both in zero-shot settings without CoT; while trailing behind GPT-4. Our research indicates that learning from step-by-step explanations, whether these are generated by humans or more advanced AI models, is a promising direction to improve model capabilities and skills.
In this paper, we present the ``joint pre-training and local re-training'' framework for learning and applying multi-source knowledge graph (KG) embeddings. We are motivated by the fact that different KGs contain complementary information to improve KG embeddings and downstream tasks. We pre-train a large teacher KG embedding model over linked multi-source KGs and distill knowledge to train a student model for a task-specific KG. To enable knowledge transfer across different KGs, we use entity alignment to build a linked subgraph for connecting the pre-trained KGs and the target KG. The linked subgraph is re-trained for three-level knowledge distillation from the teacher to the student, i.e., feature knowledge distillation, network knowledge distillation, and prediction knowledge distillation, to generate more expressive embeddings. The teacher model can be reused for different target KGs and tasks without having to train from scratch. We conduct extensive experiments to demonstrate the effectiveness and efficiency of our framework.