Daisuke Miyashita

Yasuto Hoshi, Daisuke Miyashita, Youyang Ng et al.

Retrieval-augmented large language models (R-LLMs) combine pre-trained large language models (LLMs) with information retrieval systems to improve the accuracy of factual question-answering. However, current libraries for building R-LLMs provide high-level abstractions without sufficient transparency for evaluating and optimizing prompts within specific inference processes such as retrieval and generation. To address this gap, we present RaLLe, an open-source framework designed to facilitate the development, evaluation, and optimization of R-LLMs for knowledge-intensive tasks. With RaLLe, developers can easily develop and evaluate R-LLMs, improving hand-crafted prompts, assessing individual inference processes, and objectively measuring overall system performance quantitatively. By leveraging these features, developers can enhance the performance and accuracy of their R-LLMs in knowledge-intensive generation tasks. We open-source our code at https://github.com/yhoshi3/RaLLe.

Youyang Ng, Daisuke Miyashita, Yasuto Hoshi et al.

Large Language Model (LLM) based Generative AI systems have seen significant progress in recent years. Integrating a knowledge retrieval architecture allows for seamless integration of private data into publicly available Generative AI systems using pre-trained LLM without requiring additional model fine-tuning. Moreover, Retrieval-Centric Generation (RCG) approach, a promising future research direction that explicitly separates roles of LLMs and retrievers in context interpretation and knowledge memorization, potentially leads to more efficient implementation. SimplyRetrieve is an open-source tool with the goal of providing a localized, lightweight, and user-friendly interface to these sophisticated advancements to the machine learning community. SimplyRetrieve features a GUI and API based RCG platform, assisted by a Private Knowledge Base Constructor and a Retrieval Tuning Module. By leveraging these capabilities, users can explore the potential of RCG for improving generative AI performance while maintaining privacy standards. The tool is available at https://github.com/RCGAI/SimplyRetrieve with an MIT license.

Kengo Nakata, Youyang Ng, Daisuke Miyashita et al.

In existing image classification systems that use deep neural networks, the knowledge needed for image classification is implicitly stored in model parameters. If users want to update this knowledge, then they need to fine-tune the model parameters. Moreover, users cannot verify the validity of inference results or evaluate the contribution of knowledge to the results. In this paper, we investigate a system that stores knowledge for image classification, such as image feature maps, labels, and original images, not in model parameters but in external high-capacity storage. Our system refers to the storage like a database when classifying input images. To increase knowledge, our system updates the database instead of fine-tuning model parameters, which avoids catastrophic forgetting in incremental learning scenarios. We revisit a kNN (k-Nearest Neighbor) classifier and employ it in our system. By analyzing the neighborhood samples referred by the kNN algorithm, we can interpret how knowledge learned in the past is used for inference results. Our system achieves 79.8% top-1 accuracy on the ImageNet dataset without fine-tuning model parameters after pretraining, and 90.8% accuracy on the Split CIFAR-100 dataset in the task incremental learning setting.

Yasuto Hoshi, Daisuke Miyashita, Jun Deguchi

Long-context generation is increasingly limited by decode-time key-value (KV) cache traffic, particularly when KV is offloaded beyond GPU memory. Query-aware retrieval (e.g., Top-K selection) reduces this traffic by loading only a subset of KV pairs, but renormalizing the softmax over the subset introduces bias when attention mass is spread over unretrieved tokens. We propose a retrieval-completion attention module that keeps backbone weights and the KV-cache format unchanged. For each query, we compute exact attention over sink/tail anchors and the query-dependent retrieved Top-K tokens, and estimate the remaining mid-region numerator and denominator using a fixed-size feature-map summary computed at prefill time. We add the exact and estimated contributions in the unnormalized domain and apply a single normalization, recovering the missing softmax mass without additional attention-side KV reads. Across long-context benchmarks, the proposed method improves over selection-only Top-K at matched token-equivalent read budgets, with the largest gains in high-entropy heads.

Kengo Nakata, Daisuke Miyashita, Youyang Ng et al.

In this paper, we rethink sparse lexical representations for image retrieval. By utilizing multi-modal large language models (M-LLMs) that support visual prompting, we can extract image features and convert them into textual data, enabling us to utilize efficient sparse retrieval algorithms employed in natural language processing for image retrieval tasks. To assist the LLM in extracting image features, we apply data augmentation techniques for key expansion and analyze the impact with a metric for relevance between images and textual data. We empirically show the superior precision and recall performance of our image retrieval method compared to conventional vision-language model-based methods on the MS-COCO, PASCAL VOC, and NUS-WIDE datasets in a keyword-based image retrieval scenario, where keywords serve as search queries. We also demonstrate that the retrieval performance can be improved by iteratively incorporating keywords into search queries.

Yasuto Hoshi, Daisuke Miyashita, Yasuhiro Morioka et al.

Neural document retrievers, including dense passage retrieval (DPR), have outperformed classical lexical-matching retrievers, such as BM25, when fine-tuned and tested on specific question-answering datasets. However, it has been shown that the existing dense retrievers do not generalize well not only out of domain but even in domain such as Wikipedia, especially when a named entity in a question is a dominant clue for retrieval. In this paper, we propose an approach toward in-domain generalization using the embeddings generated by the frozen language model trained with the entities in the domain. By not fine-tuning, we explore the possibility that the rich knowledge contained in a pretrained language model can be used for retrieval tasks. The proposed method outperforms conventional DPRs on entity-centric questions in Wikipedia domain and achieves almost comparable performance to BM25 and state-of-the-art SPAR model. We also show that the contextualized keys lead to strong improvements compared to BM25 when the entity names consist of common words. Our results demonstrate the feasibility of the zero-shot retrieval method for entity-centric questions of Wikipedia domain, where DPR has struggled to perform.

Kento Tatsuno, Daisuke Miyashita, Taiga Ikeda et al.

Graph-based approximate nearest neighbor search (ANNS) algorithms work effectively against large-scale vector retrieval. Among such methods, DiskANN achieves good recall-speed tradeoffs using both DRAM and storage. DiskANN adopts product quantization (PQ) to reduce memory usage, which is still proportional to the scale of datasets. In this paper, we propose All-in-Storage ANNS with Product Quantization (AiSAQ), which offloads compressed vectors to the SSD index. Our method achieves $\sim$10 MB memory usage in query search with billion-scale datasets without critical latency degradation. AiSAQ also reduces the index load time for query search preparation, which enables fast switch between muitiple billion-scale indices.This method can be applied to retrievers of retrieval-augmented generation (RAG) and be scaled out with multiple-server systems for emerging datasets. Our DiskANN-based implementation is available on GitHub.

Hiroaki Yamanaka, Daisuke Miyashita, Takashi Toi et al.

Driven by our mission of "uplifting the world with memory," this paper explores the design concept of "memory" that is essential for achieving artificial superintelligence (ASI). Rather than proposing novel methods, we focus on several alternative approaches whose potential benefits are widely imaginable, yet have remained largely unexplored. The currently dominant paradigm, which can be termed "extract then store," involves extracting information judged to be useful from experiences and saving only the extracted content. However, this approach inherently risks the loss of information, as some valuable knowledge particularly for different tasks may be discarded in the extraction process. In contrast, we emphasize the "store then on-demand extract" approach, which seeks to retain raw experiences and flexibly apply them to various tasks as needed, thus avoiding such information loss. In addition, we highlight two further approaches: discovering deeper insights from large collections of probabilistic experiences, and improving experience collection efficiency by sharing stored experiences. While these approaches seem intuitively effective, our simple experiments demonstrate that this is indeed the case. Finally, we discuss major challenges that have limited investigation into these promising directions and propose research topics to address them.

Taiga Ikeda, Daisuke Miyashita, Jun Deguchi

Large-scale approximate nearest neighbor search (ANN) has been gaining attention along with the latest machine learning researches employing ANNs. If the data is too large to fit in memory, it is necessary to search for the most similar vectors to a given query vector from the data stored in storage devices, not from that in memory. The storage device such as NAND flash memory has larger capacity than the memory device such as DRAM, but they also have larger latency to read data. Therefore, ANN methods for storage require completely different approaches from conventional in-memory ANN methods. Since the approximation that the time required for search is determined only by the amount of data fetched from storage holds under reasonable assumptions, our goal is to minimize it while maximizing recall. For partitioning-based ANNs, vectors are partitioned into clusters in the index building phase. In the search phase, some of the clusters are chosen, the vectors in the chosen clusters are fetched from storage, and the nearest vector is retrieved from the fetched vectors. Thus, the key point is to accurately select the clusters containing the ground truth nearest neighbor vectors. We accomplish this by proposing a method to predict the correct clusters by means of a neural network that is gradually refined by alternating supervised learning and duplicated cluster assignment. Compared to state-of-the-art SPANN and an exhaustive method using k-means clustering and linear search, the proposed method achieves 90% recall on SIFT1M with 80% and 58% less data fetched from storage, respectively.

Daisuke Miyashita, Edward H. Lee, Boris Murmann

Recent advances in convolutional neural networks have considered model complexity and hardware efficiency to enable deployment onto embedded systems and mobile devices. For example, it is now well-known that the arithmetic operations of deep networks can be encoded down to 8-bit fixed-point without significant deterioration in performance. However, further reduction in precision down to as low as 3-bit fixed-point results in significant losses in performance. In this paper we propose a new data representation that enables state-of-the-art networks to be encoded to 3 bits with negligible loss in classification performance. To perform this, we take advantage of the fact that the weights and activations in a trained network naturally have non-uniform distributions. Using non-uniform, base-2 logarithmic representation to encode weights, communicate activations, and perform dot-products enables networks to 1) achieve higher classification accuracies than fixed-point at the same resolution and 2) eliminate bulky digital multipliers. Finally, we propose an end-to-end training procedure that uses log representation at 5-bits, which achieves higher final test accuracy than linear at 5-bits.