The rapid development and application of foundation models have revolutionized the field of artificial intelligence. Large diffusion models have gained significant attention for their ability to generate photorealistic images and support various tasks. On-device deployment of these models provides benefits such as lower server costs, offline functionality, and improved user privacy. However, common large diffusion models have over 1 billion parameters and pose challenges due to restricted computational and memory resources on devices. We present a series of implementation optimizations for large diffusion models that achieve the fastest reported inference latency to-date (under 12 seconds for Stable Diffusion 1.4 without int8 quantization on Samsung S23 Ultra for a 512x512 image with 20 iterations) on GPU-equipped mobile devices. These enhancements broaden the applicability of generative AI and improve the overall user experience across a wide range of devices.

In this work, we propose a new class of learnable optimizers, called \textit{Mnemosyne}. It is based on the novel spatio-temporal low-rank implicit attention Transformers that can learn to train entire neural network architectures, including other Transformers, without any task-specific optimizer tuning. We show that Mnemosyne: (a) outperforms popular LSTM optimizers (also with new feature engineering to mitigate catastrophic forgetting of LSTMs), (b) can successfully train Transformers while using simple meta-training strategies that require minimal computational resources, (c) matches accuracy-wise SOTA hand-designed optimizers with carefully tuned hyper-parameters (often producing top performing models). Furthermore, Mnemosyne provides space complexity comparable to that of its hand-designed first-order counterparts, which allows it to scale to training larger sets of parameters. We conduct an extensive empirical evaluation of Mnemosyne on: (a) fine-tuning a wide range of Vision Transformers (ViTs) from medium-size architectures to massive ViT-Hs (36 layers, 16 heads), (b) pre-training BERT models and (c) soft prompt-tuning large 11B+ T5XXL models. We complement our results with a comprehensive theoretical analysis of the compact associative memory used by Mnemosyne which we believe was never done before.

Deep neural networks have been shown to be vulnerable to small perturbations of their inputs, known as adversarial attacks. In this paper, we investigate the vulnerability of Neural Machine Translation (NMT) models to adversarial attacks and propose a new attack algorithm called TransFool. To fool NMT models, TransFool builds on a multi-term optimization problem and a gradient projection step. By integrating the embedding representation of a language model, we generate fluent adversarial examples in the source language that maintain a high level of semantic similarity with the clean samples. Experimental results demonstrate that, for different translation tasks and NMT architectures, our white-box attack can severely degrade the translation quality while the semantic similarity between the original and the adversarial sentences stays high. Moreover, we show that TransFool is transferable to unknown target models. Finally, based on automatic and human evaluations, TransFool leads to improvement in terms of success rate, semantic similarity, and fluency compared to the existing attacks both in white-box and black-box settings. Thus, TransFool permits us to better characterize the vulnerability of NMT models and outlines the necessity to design strong defense mechanisms and more robust NMT systems for real-life applications.

Few-shot learning (FSL) is one of the significant and hard problems in the field of image classification. However, in contrast to the rapid development of the visible light dataset, the progress in SAR target image classification is much slower. The lack of unified benchmark is a key reason for this phenomenon, which may be severely overlooked by the current literature. The researchers of SAR target image classification always report their new results on their own datasets and experimental setup. It leads to inefficiency in result comparison and impedes the further progress of this area. Motivated by this observation, we propose a novel few-shot SAR image classification benchmark (FewSAR) to address this issue. FewSAR consists of an open-source Python code library of 15 classic methods in three categories for few-shot SAR image classification. It provides an accessible and customizable testbed for different few-shot SAR image classification task. To further understanding the performance of different few-shot methods, we establish evaluation protocols and conduct extensive experiments within the benchmark. By analyzing the quantitative results and runtime under the same setting, we observe that the accuracy of metric learning methods can achieve the best results. Meta-learning methods and fine-tuning methods perform poorly on few-shot SAR images, which is primarily due to the bias of existing datasets. We believe that FewSAR will open up a new avenue for future research and development, on real-world challenges at the intersection of SAR image classification and few-shot deep learning. We will provide our code for the proposed FewSAR at https://github.com/solarlee/FewSAR.

This paper investigates the effect of tokenizers on the downstream performance of pretrained language models (PLMs) in scriptio continua languages where no explicit spaces exist between words, using Japanese as a case study. The tokenizer for such languages often consists of a morphological analyzer and a subword tokenizer, requiring us to conduct a comprehensive study of all possible pairs. However, previous studies lack this comprehensiveness. We therefore train extensive sets of tokenizers, build a PLM using each, and measure the downstream performance on a wide range of tasks. Our results demonstrate that each downstream task has a different optimal morphological analyzer, and that it is better to use Byte-Pair-Encoding or Unigram rather than WordPiece as a subword tokenizer, regardless of the type of task.

In this work we propose a novel concept of a hierarchical confusion matrix, opening the door for popular confusion matrix based (flat) evaluation measures from binary classification problems, while considering the peculiarities of hierarchical classification problems. We develop the concept to a generalized form and prove its applicability to all types of hierarchical classification problems including directed acyclic graphs, multi path labelling, and non mandatory leaf node prediction. Finally, we use measures based on the novel confusion matrix to evaluate models within a benchmark for three real world hierarchical classification applications and compare the results to established evaluation measures. The results outline the reasonability of this approach and its usefulness to evaluate hierarchical classification problems. The implementation of hierarchical confusion matrix is available on GitHub.

This paper presents an analysis of annotation using an automatic pre-annotation for a mid-level annotation complexity task -- dependency syntax annotation. It compares the annotation efforts made by annotators using a pre-annotated version (with a high-accuracy parser) and those made by fully manual annotation. The aim of the experiment is to judge the final annotation quality when pre-annotation is used. In addition, it evaluates the effect of automatic linguistically-based (rule-formulated) checks and another annotation on the same data available to the annotators, and their influence on annotation quality and efficiency. The experiment confirmed that the pre-annotation is an efficient tool for faster manual syntactic annotation which increases the consistency of the resulting annotation without reducing its quality.

Despite the successes of probabilistic models based on passing noise through neural networks, recent work has identified that such methods often fail to capture tail behavior accurately--unless the tails of the base distribution are appropriately calibrated. To overcome this deficiency, we propose a systematic approach for analyzing the tails of random variables, and we illustrate how this approach can be used during the static analysis (before drawing samples) pass of a probabilistic programming language compiler. To characterize how the tails change under various operations, we develop an algebra which acts on a three-parameter family of tail asymptotics and which is based on the generalized Gamma distribution. Our algebraic operations are closed under addition and multiplication; they are capable of distinguishing sub-Gaussians with differing scales; and they handle ratios sufficiently well to reproduce the tails of most important statistical distributions directly from their definitions. Our empirical results confirm that inference algorithms that leverage our heavy-tailed algebra attain superior performance across a number of density modeling and variational inference tasks.

As the capabilities of large language models (LLMs) continue to advance, evaluating their performance becomes increasingly crucial and challenging. This paper aims to bridge this gap by introducing CMMLU, a comprehensive Chinese benchmark that covers various subjects, including natural science, social sciences, engineering, and humanities. We conduct a thorough evaluation of 18 advanced multilingual- and Chinese-oriented LLMs, assessing their performance across different subjects and settings. The results reveal that most existing LLMs struggle to achieve an average accuracy of 50%, even when provided with in-context examples and chain-of-thought prompts, whereas the random baseline stands at 25%. This highlights significant room for improvement in LLMs. Additionally, we conduct extensive experiments to identify factors impacting the models' performance and propose directions for enhancing LLMs. CMMLU fills the gap in evaluating the knowledge and reasoning capabilities of large language models within the Chinese context.

Dual-task dialog language understanding aims to tackle two correlative dialog language understanding tasks simultaneously via leveraging their inherent correlations. In this paper, we put forward a new framework, whose core is relational temporal graph reasoning.We propose a speaker-aware temporal graph (SATG) and a dual-task relational temporal graph (DRTG) to facilitate relational temporal modeling in dialog understanding and dual-task reasoning. Besides, different from previous works that only achieve implicit semantics-level interactions, we propose to model the explicit dependencies via integrating prediction-level interactions. To implement our framework, we first propose a novel model Dual-tAsk temporal Relational rEcurrent Reasoning network (DARER), which first generates the context-, speaker- and temporal-sensitive utterance representations through relational temporal modeling of SATG, then conducts recurrent dual-task relational temporal graph reasoning on DRTG, in which process the estimated label distributions act as key clues in prediction-level interactions. And the relational temporal modeling in DARER is achieved by relational convolutional networks (RGCNs). Then we further propose Relational Temporal Transformer (ReTeFormer), which achieves fine-grained relational temporal modeling via Relation- and Structure-aware Disentangled Multi-head Attention. Accordingly, we propose DARER with ReTeFormer (DARER2), which adopts two variants of ReTeFormer to achieve the relational temporal modeling of SATG and DTRG, respectively. The extensive experiments on different scenarios verify that our models outperform state-of-the-art models by a large margin. Remarkably, on the dialog sentiment classification task in the Mastodon dataset, DARER and DARER2 gain relative improvements of about 28% and 34% over the previous best model in terms of F1.