Knowledge distillation (KD) has been widely adopted to compress large language models (LLMs). Existing KD methods investigate various divergence measures including the Kullback-Leibler (KL), reverse Kullback-Leibler (RKL), and Jensen-Shannon (JS) divergences. However, due to limitations inherent in their assumptions and definitions, these measures fail to deliver effective supervision when few distribution overlap exists between the teacher and the student. In this paper, we show that the aforementioned KL, RKL, and JS divergences respectively suffer from issues of mode-averaging, mode-collapsing, and mode-underestimation, which deteriorates logits-based KD for diverse NLP tasks. We propose the Sinkhorn Knowledge Distillation (SinKD) that exploits the Sinkhorn distance to ensure a nuanced and precise assessment of the disparity between teacher and student distributions. Besides, profit by properties of the Sinkhorn metric, we can get rid of sample-wise KD that restricts the perception of divergence in each teacher-student sample pair. Instead, we propose a batch-wise reformulation to capture geometric intricacies of distributions across samples in the high-dimensional space. Comprehensive evaluation on GLUE and SuperGLUE, in terms of comparability, validity, and generalizability, highlights our superiority over state-of-the-art methods on all kinds of LLMs with encoder-only, encoder-decoder, and decoder-only architectures.
Retrieval augmentation has become an effective solution to empower large language models (LLMs) with external and verified knowledge sources from the database, which overcomes the limitations and hallucinations of LLMs in handling up-to-date and domain-specific information. However, existing embedding models for text retrieval usually have three non-negligible limitations. First, the number and diversity of samples in a batch are too restricted to supervise the modeling of textual nuances at scale. Second, the high proportional noise are detrimental to the semantic correctness and consistency of embeddings. Third, the equal treatment to easy and difficult samples would cause sub-optimum convergence of embeddings with poorer generalization. In this paper, we propose the PEG, a progressively learned embeddings for robust text retrieval. Specifically, we increase the training in-batch negative samples to 80,000, and for each query, we extracted five hard negatives. Concurrently, we incorporated a progressive learning mechanism, enabling the model to dynamically modulate its attention to the samples throughout the entire training process. Additionally, PEG is trained on more than 100 million data, encompassing a wide range of domains (e.g., finance, medicine, and tourism) and covering various tasks (e.g., question-answering, machine reading comprehension, and similarity matching). Extensive experiments conducted on C-MTEB and DuReader demonstrate that PEG surpasses state-of-the-art embeddings in retrieving true positives, highlighting its significant potential for applications in LLMs. Our model is publicly available at https://huggingface.co/TownsWu/PEG.
In order to appropriately filter multi-modality data sets on a web-scale, it becomes crucial to employ suitable filtering methods to boost performance and reduce training costs. For instance, LAION papers employs the CLIP score filter to select data with CLIP scores surpassing a certain threshold. On the other hand, T-MARS achieves high-quality data filtering by detecting and masking text within images and then filtering by CLIP score. Through analyzing the dataset, we observe a significant proportion of redundant information, such as numbers, present in the textual content. Our experiments on a subset of the data unveil the profound impact of these redundant elements on the CLIP scores. A logical approach would involve reevaluating the CLIP scores after eliminating these influences. Experimentally, our text-based CLIP filter outperforms the top-ranked method on the ``small scale" of DataComp (a data filtering benchmark) on ImageNet distribution shifts, achieving a 3.6% performance improvement. The results also demonstrate that our proposed text-masked filter outperforms the original CLIP score filter when selecting the top 40% of the data. The impact of numbers on CLIP and their handling provide valuable insights for improving the effectiveness of CLIP training, including language rewrite techniques.
During the preceding biennium, vision-language pre-training has achieved noteworthy success on several downstream tasks. Nevertheless, acquiring high-quality image-text pairs, where the pairs are entirely exclusive of each other, remains a challenging task, and noise exists in the commonly used datasets. To address this issue, we propose SoftCLIP, a novel approach that relaxes the strict one-to-one constraint and achieves a soft cross-modal alignment by introducing a softened target, which is generated from the fine-grained intra-modal self-similarity. The intra-modal guidance is indicative to enable two pairs have some local similarities and model many-to-many relationships between the two modalities. Besides, since the positive still dominates in the softened target distribution, we disentangle the negatives in the distribution to further boost the relation alignment with the negatives in the cross-modal learning. Extensive experiments demonstrate the effectiveness of SoftCLIP. In particular, on ImageNet zero-shot classification task, using CC3M/CC12M as pre-training dataset, SoftCLIP brings a top-1 accuracy improvement of 6.8%/7.2% over the CLIP baseline.
Large-scale vision-language pre-training has achieved promising results on downstream tasks. Existing methods highly rely on the assumption that the image-text pairs crawled from the Internet are in perfect one-to-one correspondence. However, in real scenarios, this assumption can be difficult to hold: the text description, obtained by crawling the affiliated metadata of the image, often suffer from semantic mismatch and mutual compatibility. To address these issues, here we introduce PyramidCLIP, which constructs an input pyramid with different semantic levels, and aligns visual elements and linguistic elements in the form of hierarchy via intra-level semantics alignment and cross-level relation alignment. Furthermore, we adjust the objective function by softening the loss of negative samples (unpaired samples) so as to weaken the strict constraint during the pre-training stage, thus mitigating the risk of the model being over-confident. Experiments on three downstream tasks, including zero-shot image classification, zero-shot image-text retrieval and image object detection, verify the effectiveness of the proposed PyramidCLIP. In particular, with the same amount of pre-training data of 15 millions image-text pairs, PyramidCLIP exceeds CLIP by 19.2%/18.5%/19.6% respectively, with the image encoder being ResNet-50/ViT-B32/ViT-B16 on ImageNet zero-shot classification top-1 accuracy. When scaling to larger datasets, the results of PyramidCLIP only trained for 8 epochs using 128M image-text pairs are very close to that of CLIP trained for 32 epochs using 400M training data.
The bound of the information transmission rate of direct current biased optical orthogonal frequency division multiplexing (DCO-OFDM) for visible light communication (VLC) with finite-alphabet inputs is yet unknown, where the corresponding spectral efficiency (SE) and energy efficiency (EE) stems out as the open research problems. In this paper, we derive the exact achievable rate of {the} DCO-OFDM system with finite-alphabet inputs for the first time. Furthermore, we investigate SE maximization problems of {the} DCO-OFDM system subject to both electrical and optical power constraints. By exploiting the relationship between the mutual information and the minimum mean-squared error, we propose a multi-level mercury-water-filling power allocation scheme to achieve the maximum SE. Moreover, the EE maximization problems of {the} DCO-OFDM system are studied, and the Dinkelbach-type power allocation scheme is developed for the maximum EE. Numerical results verify the effectiveness of the proposed theories and power allocation schemes.
Network sparsity receives popularity mostly due to its capability to reduce the network complexity. Extensive studies excavate gradient-driven sparsity. Typically, these methods are constructed upon premise of weight independence, which however, is contrary to the fact that weights are mutually influenced. Thus, their performance remains to be improved. In this paper, we propose to further optimize gradient-driven sparsity (OptG) by solving this independence paradox. Our motive comes from the recent advances on supermask training which shows that sparse subnetworks can be located in a randomly initialized network by simply updating mask values without modifying any weight. We prove that supermask training is to accumulate the weight gradients and can partly solve the independence paradox. Consequently, OptG integrates supermask training into gradient-driven sparsity, and a specialized mask optimizer is designed to solve the independence paradox. Experiments show that OptG can well surpass many existing state-of-the-art competitors. Our code is available at \url{https://github.com/zyxxmu/OptG}.
Given a database schema, Text-to-SQL aims to translate a natural language question into the corresponding SQL query. Under the setup of cross-domain, traditional semantic parsing models struggle to adapt to unseen database schemas. To improve the model generalization capability for rare and unseen schemas, we propose a new architecture, ShadowGNN, which processes schemas at abstract and semantic levels. By ignoring names of semantic items in databases, abstract schemas are exploited in a well-designed graph projection neural network to obtain delexicalized representation of question and schema. Based on the domain-independent representations, a relation-aware transformer is utilized to further extract logical linking between question and schema. Finally, a SQL decoder with context-free grammar is applied. On the challenging Text-to-SQL benchmark Spider, empirical results show that ShadowGNN outperforms state-of-the-art models. When the annotated data is extremely limited (only 10\% training set), ShadowGNN gets over absolute 5\% performance gain, which shows its powerful generalization ability. Our implementation will be open-sourced at \url{https://github.com/WowCZ/shadowgnn}.
Binary neural networks (BNNs) have received increasing attention due to their superior reductions of computation and memory. Most existing works focus on either lessening the quantization error by minimizing the gap between the full-precision weights and their binarization or designing a gradient approximation to mitigate the gradient mismatch, while leaving the "dead weights" untouched. This leads to slow convergence when training BNNs. In this paper, for the first time, we explore the influence of "dead weights" which refer to a group of weights that are barely updated during the training of BNNs, and then introduce rectified clamp unit (ReCU) to revive the "dead weights" for updating. We prove that reviving the "dead weights" by ReCU can result in a smaller quantization error. Besides, we also take into account the information entropy of the weights, and then mathematically analyze why the weight standardization can benefit BNNs. We demonstrate the inherent contradiction between minimizing the quantization error and maximizing the information entropy, and then propose an adaptive exponential scheduler to identify the range of the "dead weights". By considering the "dead weights", our method offers not only faster BNN training, but also state-of-the-art performance on CIFAR-10 and ImageNet, compared with recent methods. Code can be available at [this https URL](https://github.com/z-hXu/ReCU).