Low-rank adaptation (LoRA) is a predominant parameter-efficient finetuning method for adapting large language models (LLMs) to downstream tasks. Meanwhile, Compute-in-Memory (CIM) architectures demonstrate superior energy efficiency due to their array-level parallel in-memory computing designs. In this article, we propose deploying the LoRA-finetuned LLMs on the hybrid CIM architecture (i.e., pretrained weights onto energy-efficient Resistive Random-Access Memory (RRAM) and LoRA branches onto noise-free Static Random-Access Memory (SRAM)), reducing the energy cost to about 3% compared with the Nvidia A100 GPU. However, the inherent noise of RRAM on the saved weights leads to performance degradation, simultaneously. To address this issue, we design a novel Hardware-aware Low-rank Adaptation (HaLoRA) method. The key insight is to train a LoRA branch that is robust toward such noise and then deploy it on noise-free SRAM, while the extra cost is negligible since the parameters of LoRAs are much fewer than pretrained weights (e.g., 0.15% for LLaMA-3.2 1B model). To improve the robustness towards the noise, we theoretically analyze the gap between the optimization trajectories of the LoRA branch under both ideal and noisy conditions and further design an extra loss to minimize the upper bound of this gap. Therefore, we can enjoy both energy efficiency and accuracy during inference. Experiments finetuning the Qwen and LLaMA series demonstrate the effectiveness of HaLoRA across multiple reasoning tasks, achieving up to 22.7 improvement in average score while maintaining robustness at various noise types and noise levels.

Zhi Qu, Yiran Wang, Jiannan Mao, Chenchen Ding, Hideki Tanaka, Masao Utiyama, Taro Watanabe. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Exploring multiple precisions as well as sparsities for a computingin-memory (CIM) based convolutional accelerators is challenging. To further improve energy efficiency with minimal accuracy loss, this paper develops a neural architecture search (NAS) method to identify precision for each layer of the CNN and further leverages bit-level sparsity. The results indicate that following this approach, ResNet-18 and VGG-16 not only maintain their accuracy but also implement layer-wised mixed-precision effectively. Furthermore, there is a substantial enhancement in the bit-level sparsity of weights within each layer, with an average bit-level sparsity exceeding 90% per bit, thus providing broader possibilities for hardware-level sparsity optimization. In terms of hardware design, a mixed-precision (2/4/8-bit) readout circuit as well as a bit-level sparsity-aware Analog-to-Digital Converter (ADC) are both proposed to reduce system power consumption. Based on bit-level sparsity mixed-precision CNNs benchmarks, post-layout simulation results in 28nm reveal that the proposed accelerator achieves up to 245.72 TOPS/W energy efficiency, which shows about 2.52 -- 6.57× improvement compared to the state-of-the-art SRAM-based CIM accelerators.

Compute-in-memory using emerging resistive random-access memory (RRAM) demonstrates significant potential for building energy-efficient deep neural networks. However, RRAM-based network training faces challenges from computational noise and gradient calculation overhead. In this study, we introduce NoiseZO, a forward-only training framework that leverages intrinsic RRAM noise to estimate gradients via zeroth-order (ZO) optimization. The framework maps neural networks onto dual RRAM arrays, utilizing their inherent write noise as $\mathbf{Z O}$ perturbations for training. This enables network updates through only two forward computations. A fine-grained perturbation control strategy is further developed to enhance training accuracy. Extensive experiments on vowel and image datasets, implemented with typical networks, showcase the effectiveness of our framework. Compared to conventional complementary metal-oxide-semiconductor (CMOS) implementations, our approach achieves a 21-fold reduction in energy consumption.

Resistive random-access memory (RRAM)-based compute-in-memory (CIM) systems show promise in accelerating Transformer-based vision models but face challenges from inherent device non-idealities. In this work, we systematically investigate the vulnerability of Transformer-based vision models to RRAM-induced perturbations. Our analysis reveals that earlier Transformer layers are more vulnerable than later ones, and feed-forward networks (FFNs) are more susceptible to noise than multi-head self-attention (MHSA). Based on these observations, we propose a noise-aware knowledge distillation framework that enhances model robustness by aligning both intermediate features and final outputs between weight-perturbed and noise-free models. Experimental results demonstrate that our method improves accuracy by up to 1.54% and 1.49% on ViT and DeiT models under various noise conditions compared to their vanilla counterparts.

3D Gaussian inpainting, a critical technique for numerous applications in virtual reality and multimedia, has made significant progress with pretrained diffusion models. However, ensuring multi-view consistency, an essential requirement for high-quality inpainting, remains a key challenge. In this work, we present PAInpainter, a novel approach designed to advance 3D Gaussian inpainting by leveraging perspective-aware content propagation and consistency verification across multi-view inpainted images. Our method iteratively refines inpainting and optimizes the 3D Gaussian representation with multiple views adaptively sampled from a perspective graph. By propagating inpainted images as prior information and verifying consistency across neighboring views, PAInpainter substantially enhances global consistency and texture fidelity in restored 3D scenes. Extensive experiments demonstrate the superiority of PAInpainter over existing methods. Our approach achieves superior 3D inpainting quality, with PSNR scores of 26.03 dB and 29.51 dB on the SPIn-NeRF and NeRFiller datasets, respectively, highlighting its effectiveness and generalization capability.

Understanding representation transfer in multilingual neural machine translation (MNMT) can reveal the reason for the zero-shot translation deficiency. In this work, we systematically analyze the representational issue of MNMT models. We first introduce the identity pair, translating a sentence to itself, to address the lack of the base measure in multilingual investigations, as the identity pair can reflect the representation of a language within the model. Then, we demonstrate that the encoder transfers the source language to the representational subspace of the target language instead of the language-agnostic state. Thus, the zero-shot translation deficiency arises because the representation of a translation is entangled with other languages and not transferred to the target language effectively. Based on our findings, we propose two methods: 1) low-rank language-specific embedding at the encoder, and 2) language-specific contrastive learning of the representation at the decoder. The experimental results on Europarl-15, TED-19, and OPUS-100 datasets show that our methods substantially enhance the performance of zero-shot translations without sacrifices in supervised directions by improving language transfer capacity, thereby providing practical evidence to support our conclusions. Codes are available at https://github.com/zhiqu22/ZeroTrans.

The real-time detection of 3D objects represents a grand challenge on edge devices. Existing 3D point clouds models are over-parameterized with heavy computation load. This paper proposes a highly compact model for 3D point clouds detection using tensor-compression. Compared to conventional methods, we propose a fused multi-head transformer tensor-compression (FMTT) to achieve both compact size yet with high accuracy. The FMTT leverages different ranks to extract both high and low-level features and then fuses them together to improve the accuracy. Experiments on the KITTI dataset show that the proposed FMTT can achieve 6.04× smaller than the uncompressed model from 55.09MB to 9.12MB such that the compressed model can be implemented on edge devices. It also achieves 2.62% improved accuracy in easy mode and 0.28% improved accuracy in hard mode.

The increasing model size and computation load of convolutional neural networks (CNN) pose a grand challenge to deploy CNN models on edge computing devices. To further improve performance without significant accuracy loss, this paper developed a neural architecture search (NAS) method to achieve a layer-wise mixed-precision-and-sparsity (LAMPS) CNN. However, this optimization cannot be fully utilized and directly mapped to existing AI accelerators due to the irregu- lar computation of sparse and multi-precision data. To tackle this challenge, this work proposed a LAMPS vector systolic accelerator and demonstrated state-of-the-art results. Experi- mental results show that the LAMPS accelerator on Xilinx ZCU102 achieves an average performance of 756.83 GOPS and 470.25 GOPS when accelerating the NAS-optimized VGG16 and Resnet18, respectively, leading to 1.3-6.0x speed-up over the state- of-the-art accelerators on FPGA.

Various industrial and domestic applications call for optimized lightweight video LSTM network models on edge. The recent tensor-train method can transform space-time features into tensors, which can be further decomposed into low-rank network models for lightweight video analysis on edge. The rank selection of tensor is however manually performed with no optimization. This paper formulates a rank search algorithm to automatically decide tensor ranks with consideration of the trade-off between network accuracy and complexity. A fast rank search method, called RankSearch, is developed to find optimized low-rank video LSTM network models on edge. Results from experiments show that RankSearch achieves a$4.84 >$reduction in model complexity, and$1.96\times$speed-up in run time while delivering a 3.86% accuracy improvement compared with the manual-ranked models.

Low-resource speech recognition has been long-suffering from insufficient training data. In this paper, we propose an approach that leverages neighboring languages to improve low-resource scenario performance, founded on the hypothesis that similar linguistic units in neighboring languages exhibit comparable term frequency distributions, which enables us to construct a Huffman tree for performing multilingual hierarchical Softmax decoding. This hierarchical structure enables cross-lingual knowledge sharing among similar tokens, thereby enhancing low-resource training outcomes. Empirical analyses demonstrate that our method is effective in improving the accuracy and efficiency of low-resource speech recognition.

Low-resource machine translation (LRMT) poses a substantial challenge due to the scarcity of parallel training data. This paper introduces a new method to improve the transfer of the embedding layer from the Parent model to the Child model in LRMT, utilizing trained token embeddings in the Parent model’s high-resource vocabulary. Our approach involves projecting all tokens into a shared semantic space and measuring the semantic similarity between tokens in the low-resource and high-resource languages. These measures are then utilized to initialize token representations in the Child model’s low-resource vocabulary. We evaluated our approach on three benchmark datasets of low-resource language pairs: Myanmar-English, Indonesian-English, and Turkish-English. The experimental results demonstrate that our method outperforms previous methods regarding translation quality. Additionally, our approach is computationally efficient, leading to reduced training time compared to prior works.

This article investigates approaches to effectively harness source-side linguistic features for low-resource multilingual neural machine translation (MNMT). Previous works focus on using various features of a word such as lemma, part-of-speech tag, dependency label, and so on, to improve translation quality in a low-resource scenario. However, these studies deal with bilingual translation and do not focus on using features in multilingual training setups. Our work focuses on this particular point and experiments with low-resource multilingual models incorporating source-side linguistic features. Although techniques for integrating features into an NMT model such as concatenation and feature relevance perform quite well in bilingual settings, they do not work well in multilingual settings. To remedy this, we propose the use of dummy features and language indicator features in MNMT models. Experiments are conducted on English to Asian language translation on a multilingual, multi-parallel corpus spanning English and eight Asian languages where for each language pair, the training data size does not exceed 20,000 parallel sentences. After establishing strong bilingual baselines using feature relevance mechanisms and multilingual baselines without any features, we show that our proposed dummy features and language indicator features, in combination with feature relevance mechanisms, yield significant improvements in BLEU points for all language pairs. We then analyze our models from the perspectives of model sizes, the impact of individual linguistic features, validation perplexity computed during training, visualization of the activations of the relevance mechanisms, and exhaustive tuning of hyperparameters. We also report preliminary results for multilingual multi-way models using linguistic features.

Neural machine translation uses a decoder to generate target words auto-regressively by predicting the next target word conditioned on a given source sentence and its previously predicted target words, i.e, its translation history, which suffers from two limitations: 1) the prediction of next word depends heavily on the quality of its history information. Moreover, the discrepancy between training and inference exacerbates this limitation; 2) this left-to-right decoding way cannot make full use of the target-side future information, which leads to the issue of unbalanced outputs. On the one hand, we alleviate the first limitation with a history-refining module, which learns to examine the quality of each history word by assigning it a confidence score. The confidence score is further used as a gate to control the amount of its word embedding flowing to the decoder. On the other hand, we attack the second limitation with a future-foreseeing module, which learns the distribution of future translation at each decoding time step. More importantly, we further propose refining history for future-aware NMT since the two modules can be closely incorporated as they focus on different kinds of context. Experimental results on various translation tasks with different scaled datasets, including WMT English$\leftrightarrow${German, French, Romanian}, show that our proposed approach achieves significant improvements over strong Transformer-based NMT baselines.

Transformer-like network has shown remarkable high performance in both natural language processing and computer vision. However, the huge computational demands in non-linear floating-point arithmetic and the irregular memory access requirement in self-attention mechanism make it still a challenge to deploy Transformer on edge. To address the above issues, we propose integer-only quantization scheme for the simplification of non-linear operations (such as LayerNorm, Softmax and Gelu), meanwhile algorithm-hardware co-design strategy is applied to guarantee both the high accuracy and high efficiency. Besides, we construct general-purpose group vector systolic array to efficiently accelerate the matrix multiplication operations including both regular matrix-multiplication/convolution and the irregular multi-head self-attention mechanism. Unified data-package strategy and flexible on-/off-chip data storage management strategy are also proposed to further improve the performance. The design has been deployed on Xilinx ZCU102 FPGA platform, achieving an overall inference latency of 4.077ms and 11.15ms per image for ViT-tiny and ViT-s, respectively. The average throughput can reach as high as 762.7 GOPs, which shows significant improvement over the previous state-of-the-art FPGA Transformer accelerator.

In this paper we present FeatureBART, a linguistically motivated sequence-to-sequence monolingual pre-training strategy in which syntactic features such as lemma, part-of-speech and dependency labels are incorporated into the span prediction based pre-training framework (BART). These automatically extracted features are incorporated via approaches such as concatenation and relevance mechanisms, among which the latter is known to be better than the former. When used for low-resource NMT as a downstream task, we show that these feature based models give large improvements in bilingual settings and modest ones in multilingual settings over their counterparts that do not use features.

The current popular knowledge distillation (KD) methods effectively compress the transformer-based end-to-end speech recognition model. However, existing methods fail to utilize complete information of the teacher model, and they distill only a limited number of blocks of the teacher model. In this study, we first integrate a task-driven loss function into the decoder’s intermediate blocks to generate task-related feature representations. Then, we propose an attention-based multi-level feature distillation to automatically learn the feature representation summarized by all blocks of the teacher model. Under the 1.1M parameters model, the experimental results on the Wall Street Journal dataset reveal that our approach achieves a 12.1% WER reduction compared with the baseline system.

As a highly analytic language, Khmer has considerable ambiguities in tokenization and part-of-speech (POS) tagging processing. This topic is investigated in this study. Specifically, a 20,000-sentence Khmer corpus with manual tokenization and POS-tagging annotation is released after a series of work over the last 4 years. This is the largest morphologically annotated Khmer dataset as of 2020, when this article was prepared. Based on the annotated data, experiments were conducted to establish a comprehensive benchmark on the automatic processing of tokenization and POS-tagging for Khmer. Specifically, a support vector machine, a conditional random field (CRF) , a long short-term memory (LSTM) -based recurrent neural network, and an integrated LSTM-CRF model have been investigated and discussed. As a primary conclusion, processing at morpheme-level is satisfactory for the provided data. However, it is intrinsically difficult to identify further grammatical constituents of compounds or phrases because of the complex analytic features of the language. Syntactic annotation and automatic parsing for Khmer will be scheduled in the near future.

We present that, the rank-frequency relation in textual data follows f ∝ r -α (r + γ) -β , where f is the token frequency and r is the rank by frequency, with (α, β, γ) as parameters.The formulation is derived based on the empirical observation that d 2 (x+y)/dx 2 is a typical impulse function, where (x, y) = (log r, log f ).The formulation is the power law when β = 0 and the Zipf-Mandelbrot law when α = 0. We illustrate that α is related to the analytic features of syntax and β + γ to those of morphology in natural languages from an investigation of multilingual corpora.

In this study, linguistic knowledge at different levels are incorporated into the neural machine translation (NMT) framework to improve translation quality for language pairs with extremely limited data.Integrating manually designed or automatically extracted features into the NMT framework is known to be beneficial.However, this study emphasizes that the relevance of the features is crucial to the performance.Specifically, we propose two methods, 1) self relevance and 2) word-based relevance, to improve the representation of features for NMT.Experiments are conducted on translation tasks from English to eight Asian languages, with no more than twenty thousand sentences for training.The proposed methods improve translation quality for all tasks by up to 3.09 BLEU points.Discussions with visualization provide the explainability of the proposed methods where we show that the relevance methods provide weights to features thereby enhancing their impact on low-resource machine translation.

In this study, we focus on detecting articulatory attribute errors for dysarthric patients with cerebral palsy (CP) or amyotrophic lateral sclerosis (ALS). There are two major challenges for this task. The pronunciation of dysarthric patients is unclear and inaccurate, which results in poor performances of traditional automatic speech recognition (ASR) systems and traditional automatic speech attribute transcription (ASAT). In addition, the data is limited because of the difficulty of recording. This study proposes an end-to-end automatic speech attribute transcription (E2E-ASAT) method for detecting articulatory attribute errors more precisely. To use the limited data more effectively, the parameters of the acoustic model are refactored into two layers and only one layer is retrained. Our proposed method showed good performances in both ASR and articulatory attribute detection. Our system has a potential as a rehabilitation tool.

Transliteration is generally a phonetically based transcription across different writing systems. It is a crucial task for various downstream natural language processing applications. For the Myanmar (Burmese) language, robust automatic transliteration for borrowed English words is a challenging task because of the complex Myanmar writing system and the lack of data. In this study, we constructed a Myanmar-English named entity dictionary containing more than eighty thousand transliteration instances. The data have been released under a CC BY-NC-SA license. We evaluated the automatic transliteration performance using statistical and neural network-based approaches based on the prepared data. The neural network model outperformed the statistical model significantly in terms of the BLEU score on the character level. Different units used in the Myanmar script for processing were also compared and discussed.

This article presents a comprehensive study on two primary tasks in Burmese (Myanmar) morphological analysis: tokenization and part-of-speech (POS) tagging. Twenty thousand Burmese sentences of newswire are annotated with two-layer tokenization and POS-tagging information, as one component of the Asian Language Treebank Project. The annotated corpus has been released under a CC BY-NC-SA license, and it is the largest open-access database of annotated Burmese when this manuscript was prepared in 2017. Detailed descriptions of the preparation, refinement, and features of the annotated corpus are provided in the first half of the article. Facilitated by the annotated corpus, experiment-based investigations are presented in the second half of the article, wherein the standard sequence-labeling approach of conditional random fields and a long short-term memory (LSTM)-based recurrent neural network (RNN) are applied and discussed. We obtained several general conclusions, covering the effect of joint tokenization and POS-tagging and importance of ensemble from the viewpoint of stabilizing the performance of LSTM-based RNN. This study provides a solid basis for further studies on Burmese processing.

A 20,000-sentence Burmese (Myanmar) treebank on news articles has been released under a CC BY-NC-SA license. Complete phrase structure annotation was developed for each sentence from the morphologically annotated data prepared in previous work of Ding et al. [1]. As the final result of the Burmese component in the Asian Language Treebank Project , this is the first large-scale, open-access treebank for the Burmese language. The annotation details and features of this treebank are presented.

A feasible and flexible annotation system is designed for joint tokenization and part-of-speech (POS) tagging to annotate those languages without natural definitions of words . This design was motivated by the fact that word separators are not used in many highly analytic East and Southeast Asian languages. Although several of the languages are well-studied, e.g., Chinese and Japanese, many are understudied with low resources, e.g., Burmese (Myanmar) and Khmer. In the first part of the article, the proposed annotation system, named nova, is introduced. nova contains only four basic tags (n, v, a, and o); these tags can be further modified and combined to adapt complex linguistic phenomena in tokenization and POS tagging. In the second part of the article, the feasibility and flexibility of nova is illustrated from the annotation practice on Burmese and Khmer. The relation between nova and two universal POS tagsets is discussed in the final part of the article.

A rule-based pre-ordering approach is proposed for statistical Japanese-to-English machine translation using the dependency structure of source-side sentences. A Japanese sentence is pre-ordered to an English-like order at the morpheme level for a statistical machine translation system during the training and decoding phase to resolve the reordering problem. In this article, extra-chunk pre-ordering of morphemes is proposed, which allows Japanese functional morphemes to move across chunk boundaries. This contrasts with the intra-chunk reordering used in previous approaches, which restricts the reordering of morphemes within a chunk. Linguistically oriented discussions show that correct pre-ordering cannot be realized without extra-chunk movement of morphemes. The proposed approach is compared with five rule-based pre-ordering approaches designed for Japanese-to-English translation and with a language independent statistical pre-ordering approach on a standard patent dataset and on a news dataset obtained by crawling Internet news sites. Two state-of-the-art statistical machine translation systems, one phrase-based and the other hierarchical phrase-based, are used in experiments. Experimental results show that the proposed approach outperforms the compared approaches on automatic reordering measures (Kendall’s τ, Spearman’s ρ, fuzzy reordering score, and test set RIBES) and on the automatic translation precision measure of test set BLEU score.

Experiments on various word segmentation approaches for the Burmese language are conducted and discussed in this note. Specifically, dictionary-based, statistical, and machine learning approaches are tested. Experimental results demonstrate that statistical and machine learning approaches perform significantly better than dictionary-based approaches. We believe that this note, based on an annotated corpus of relatively considerable size (containing approximately a half million words), is the first systematic comparison of word segmentation approaches for Burmese. This work aims to discover the properties and proper approaches to Burmese textual processing and to promote further researches on this understudied language.

fast align is a simple, fast, and efficient approach for word alignment based on the IBM model 2. fast align performs well for language pairs with relatively similar word orders; however, it does not perform well for language pairs with drastically different word orders.We propose a segmenting-reversing reordering process to solve this problem by alternately applying fast align and reordering source sentences during training.Experimental results with Japanese-English translation demonstrate that the proposed approach improves the performance of fast align significantly without the loss of efficiency.Experiments using other languages are also reported.

Word reordering is a crucial technique in statistical machine translation in which syntactic information plays an important role.Synchronous context-free grammar has typically been used for this purpose with various modifications for adding flexibilities to its synchronized tree generation.We permit further flexibilities in the synchronous context-free grammar in order to translate between languages with drastically different word order.Our method pre-processes a parallel corpus by abstracting source-side dependency trees, and performs long-distance reordering on top of an off-the-shelf phrase-based system.Experimental results show that our method significantly outperforms previous phrase-based and syntax-based models for translation between English and Japanese.