Sosuke Kobayashi

Sho Takase, Shun Kiyono, Sosuke Kobayashi et al.

From the perspective of the layer normalization (LN) positions, the architectures of Transformers can be categorized into two types: Post-LN and Pre-LN. Recent Transformers tend to be Pre-LN because, in Post-LN with deep Transformers (e.g., those with ten or more layers), the training is often unstable, resulting in useless models. However, Post-LN has consistently achieved better performance than Pre-LN in relatively shallow Transformers (e.g., those with six or fewer layers). This study first investigates the reason for these discrepant observations empirically and theoretically and made the following discoveries: 1, the LN in Post-LN is the main source of the vanishing gradient problem that leads to unstable training, whereas Pre-LN prevents it, and 2, Post-LN tends to preserve larger gradient norms in higher layers during the back-propagation, which may lead to effective training. Exploiting the new findings, we propose a method that can provide both high stability and effective training by a simple modification of Post-LN. We conduct experiments on a wide range of text generation tasks. The experimental results demonstrate that our method outperforms Pre-LN, and enables stable training regardless of the shallow or deep layer settings. Our code is publicly available at https://github.com/takase/b2t_connection.

Sosuke Kobayashi, Eiichi Matsumoto, Vincent Sitzmann

Emerging neural radiance fields (NeRF) are a promising scene representation for computer graphics, enabling high-quality 3D reconstruction and novel view synthesis from image observations. However, editing a scene represented by a NeRF is challenging, as the underlying connectionist representations such as MLPs or voxel grids are not object-centric or compositional. In particular, it has been difficult to selectively edit specific regions or objects. In this work, we tackle the problem of semantic scene decomposition of NeRFs to enable query-based local editing of the represented 3D scenes. We propose to distill the knowledge of off-the-shelf, self-supervised 2D image feature extractors such as CLIP-LSeg or DINO into a 3D feature field optimized in parallel to the radiance field. Given a user-specified query of various modalities such as text, an image patch, or a point-and-click selection, 3D feature fields semantically decompose 3D space without the need for re-training and enable us to semantically select and edit regions in the radiance field. Our experiments validate that the distilled feature fields (DFFs) can transfer recent progress in 2D vision and language foundation models to 3D scene representations, enabling convincing 3D segmentation and selective editing of emerging neural graphics representations.

Sosuke Kobayashi, Shun Kiyono, Jun Suzuki et al.

Ensembling is a popular method used to improve performance as a last resort. However, ensembling multiple models finetuned from a single pretrained model has been not very effective; this could be due to the lack of diversity among ensemble members. This paper proposes Multi-Ticket Ensemble, which finetunes different subnetworks of a single pretrained model and ensembles them. We empirically demonstrated that winning-ticket subnetworks produced more diverse predictions than dense networks, and their ensemble outperformed the standard ensemble on some tasks.

Kazuki Yano, Shun Kiyono, Sosuke Kobayashi et al.

We investigate the role of learning rate scheduling in the large-scale pre-training of large language models, focusing on its influence on downstream performance after supervised fine-tuning (SFT). Decay-based learning rate schedulers are widely used to minimize pre-training loss. However, despite their widespread use, how these schedulers affect performance after SFT remains underexplored. In this paper, we examine Warmup-Stable-Only (WSO), which maintains a constant learning rate after warmup without any decay. Through experiments with 1B and 8B parameter models, we show that WSO consistently outperforms decay-based schedulers in terms of performance after SFT, even though decay-based schedulers may exhibit better performance after pre-training. The result also holds across different regimes with mid-training and over-training. Loss landscape analysis further reveals that decay-based schedulers lead models into sharper minima, whereas WSO preserves flatter minima that support adaptability. These findings indicate that applying LR decay to improve pre-training metrics may compromise downstream adaptability. Our work also provides practical guidance for training and model release strategies, highlighting that pre-training models with WSO enhances their adaptability for downstream tasks.

Sho Takase, Shun Kiyono, Sosuke Kobayashi et al.

Loss spikes often occur during pre-training of large language models. The spikes degrade the performance of large language models and sometimes ruin the pre-training. Since the pre-training needs a vast computational budget, we should avoid such spikes. Based on the assumption that the loss spike is caused by the sudden growth of the gradient norm, we explore factors to keep the gradient norm small through an analysis of the spectral norms of the Jacobian matrices for the sub-layers. Our findings suggest that stabilizing the pre-training process requires two conditions: small sub-layers and large shortcut. We conduct various experiments to empirically verify our theoretical analyses. Experimental results demonstrate that methods satisfying the conditions effectively prevent loss spikes during pre-training.

Kazuki Yano, Sho Takase, Sosuke Kobayashi et al.

As Large Language Models (LLMs) gain widespread practical application, providing the model family of different parameter sizes has become standard practice to address diverse computational requirements. Conventionally, each model in a family is trained independently, resulting in computational costs that scale additively with the number of models. We propose an efficient method for constructing the model family through progressive training, where smaller models are incrementally expanded to larger sizes to create a complete model family. Through extensive experiments with a model family ranging from 1B to 8B parameters, we demonstrate that our method reduces computational costs by approximately 25% while maintaining comparable performance to independently trained models. Furthermore, by strategically adjusting maximum learning rates based on model size, our method outperforms the independent training across various metrics. Beyond performance gains, our approach offers an additional advantage: models in our family tend to yield more consistent behavior across different model sizes.

Rei Higuchi, Ryotaro Kawata, Naoki Nishikawa et al.

The ability to acquire latent semantics is one of the key properties that determines the performance of language models. One convenient approach to invoke this ability is to prepend metadata (e.g. URLs, domains, and styles) at the beginning of texts in the pre-training data, making it easier for the model to access latent semantics before observing the entire text. Previous studies have reported that this technique actually improves the performance of trained models in downstream tasks; however, this improvement has been observed only in specific downstream tasks, without consistent enhancement in average next-token prediction loss. To understand this phenomenon, we closely investigate how prepending metadata during pre-training affects model performance by examining its behavior using artificial data. Interestingly, we found that this approach produces both positive and negative effects on the downstream tasks. We demonstrate that the effectiveness of the approach depends on whether latent semantics can be inferred from the downstream task's prompt. Specifically, through investigations using data generated by probabilistic context-free grammars, we show that training with metadata helps improve model's performance when the given context is long enough to infer the latent semantics. In contrast, the technique negatively impacts performance when the context lacks the necessary information to make an accurate posterior inference.

Hiroki Ouchi, Jun Suzuki, Sosuke Kobayashi et al.

Interpretable rationales for model predictions are crucial in practical applications. We develop neural models that possess an interpretable inference process for dependency parsing. Our models adopt instance-based inference, where dependency edges are extracted and labeled by comparing them to edges in a training set. The training edges are explicitly used for the predictions; thus, it is easy to grasp the contribution of each edge to the predictions. Our experiments show that our instance-based models achieve competitive accuracy with standard neural models and have the reasonable plausibility of instance-based explanations.

Shun Kiyono, Sosuke Kobayashi, Jun Suzuki et al.

Position representation is crucial for building position-aware representations in Transformers. Existing position representations suffer from a lack of generalization to test data with unseen lengths or high computational cost. We investigate shifted absolute position embedding (SHAPE) to address both issues. The basic idea of SHAPE is to achieve shift invariance, which is a key property of recent successful position representations, by randomly shifting absolute positions during training. We demonstrate that SHAPE is empirically comparable to its counterpart while being simpler and faster.

Sosuke Kobayashi, Sho Yokoi, Jun Suzuki et al.

Understanding the influence of a training instance on a neural network model leads to improving interpretability. However, it is difficult and inefficient to evaluate the influence, which shows how a model's prediction would be changed if a training instance were not used. In this paper, we propose an efficient method for estimating the influence. Our method is inspired by dropout, which zero-masks a sub-network and prevents the sub-network from learning each training instance. By switching between dropout masks, we can use sub-networks that learned or did not learn each training instance and estimate its influence. Through experiments with BERT and VGGNet on classification datasets, we demonstrate that the proposed method can capture training influences, enhance the interpretability of error predictions, and cleanse the training dataset for improving generalization.

Hiroki Ouchi, Jun Suzuki, Sosuke Kobayashi et al.

Interpretable rationales for model predictions play a critical role in practical applications. In this study, we develop models possessing interpretable inference process for structured prediction. Specifically, we present a method of instance-based learning that learns similarities between spans. At inference time, each span is assigned a class label based on its similar spans in the training set, where it is easy to understand how much each training instance contributes to the predictions. Through empirical analysis on named entity recognition, we demonstrate that our method enables to build models that have high interpretability without sacrificing performance.

Sho Takase, Sosuke Kobayashi

In neural network-based models for natural language processing (NLP), the largest part of the parameters often consists of word embeddings. Conventional models prepare a large embedding matrix whose size depends on the vocabulary size. Therefore, storing these models in memory and disk storage is costly. In this study, to reduce the total number of parameters, the embeddings for all words are represented by transforming a shared embedding. The proposed method, ALONE (all word embeddings from one), constructs the embedding of a word by modifying the shared embedding with a filter vector, which is word-specific but non-trainable. Then, we input the constructed embedding into a feed-forward neural network to increase its expressiveness. Naively, the filter vectors occupy the same memory size as the conventional embedding matrix, which depends on the vocabulary size. To solve this issue, we also introduce a memory-efficient filter construction approach. We indicate our ALONE can be used as word representation sufficiently through an experiment on the reconstruction of pre-trained word embeddings. In addition, we also conduct experiments on NLP application tasks: machine translation and summarization. We combined ALONE with the current state-of-the-art encoder-decoder model, the Transformer, and achieved comparable scores on WMT 2014 English-to-German translation and DUC 2004 very short summarization with less parameters.

Takuya Shimada, Shoichiro Yamaguchi, Kohei Hayashi et al.

Data augmentation by mixing samples, such as Mixup, has widely been used typically for classification tasks. However, this strategy is not always effective due to the gap between augmented samples for training and original samples for testing. This gap may prevent a classifier from learning the optimal decision boundary and increase the generalization error. To overcome this problem, we propose an alternative framework called Data Interpolating Prediction (DIP). Unlike common data augmentations, we encapsulate the sample-mixing process in the hypothesis class of a classifier so that train and test samples are treated equally. We derive the generalization bound and show that DIP helps to reduce the original Rademacher complexity. Also, we empirically demonstrate that DIP can outperform existing Mixup.

Masaki Saito, Shunta Saito, Masanori Koyama et al.

Training of Generative Adversarial Network (GAN) on a video dataset is a challenge because of the sheer size of the dataset and the complexity of each observation. In general, the computational cost of training GAN scales exponentially with the resolution. In this study, we present a novel memory efficient method of unsupervised learning of high-resolution video dataset whose computational cost scales only linearly with the resolution. We achieve this by designing the generator model as a stack of small sub-generators and training the model in a specific way. We train each sub-generator with its own specific discriminator. At the time of the training, we introduce between each pair of consecutive sub-generators an auxiliary subsampling layer that reduces the frame-rate by a certain ratio. This procedure can allow each sub-generator to learn the distribution of the video at different levels of resolution. We also need only a few GPUs to train a highly complex generator that far outperforms the predecessor in terms of inception scores.

Riku Arakawa, Sosuke Kobayashi, Yuya Unno et al.

Exploration has been one of the greatest challenges in reinforcement learning (RL), which is a large obstacle in the application of RL to robotics. Even with state-of-the-art RL algorithms, building a well-learned agent often requires too many trials, mainly due to the difficulty of matching its actions with rewards in the distant future. A remedy for this is to train an agent with real-time feedback from a human observer who immediately gives rewards for some actions. This study tackles a series of challenges for introducing such a human-in-the-loop RL scheme. The first contribution of this work is our experiments with a precisely modeled human observer: binary, delay, stochasticity, unsustainability, and natural reaction. We also propose an RL method called DQN-TAMER, which efficiently uses both human feedback and distant rewards. We find that DQN-TAMER agents outperform their baselines in Maze and Taxi simulated environments. Furthermore, we demonstrate a real-world human-in-the-loop RL application where a camera automatically recognizes a user's facial expressions as feedback to the agent while the agent explores a maze.

Sho Yokoi, Sosuke Kobayashi, Kenji Fukumizu et al.

In this paper, we propose a new kernel-based co-occurrence measure that can be applied to sparse linguistic expressions (e.g., sentences) with a very short learning time, as an alternative to pointwise mutual information (PMI). As well as deriving PMI from mutual information, we derive this new measure from the Hilbert--Schmidt independence criterion (HSIC); thus, we call the new measure the pointwise HSIC (PHSIC). PHSIC can be interpreted as a smoothed variant of PMI that allows various similarity metrics (e.g., sentence embeddings) to be plugged in as kernels. Moreover, PHSIC can be estimated by simple and fast (linear in the size of the data) matrix calculations regardless of whether we use linear or nonlinear kernels. Empirically, in a dialogue response selection task, PHSIC is learned thousands of times faster than an RNN-based PMI while outperforming PMI in accuracy. In addition, we also demonstrate that PHSIC is beneficial as a criterion of a data selection task for machine translation owing to its ability to give high (low) scores to a consistent (inconsistent) pair with other pairs.

Sosuke Kobayashi

We propose a novel data augmentation for labeled sentences called contextual augmentation. We assume an invariance that sentences are natural even if the words in the sentences are replaced with other words with paradigmatic relations. We stochastically replace words with other words that are predicted by a bi-directional language model at the word positions. Words predicted according to a context are numerous but appropriate for the augmentation of the original words. Furthermore, we retrofit a language model with a label-conditional architecture, which allows the model to augment sentences without breaking the label-compatibility. Through the experiments for six various different text classification tasks, we demonstrate that the proposed method improves classifiers based on the convolutional or recurrent neural networks.

Reina Akama, Kento Watanabe, Sho Yokoi et al.

This paper presents the first study aimed at capturing stylistic similarity between words in an unsupervised manner. We propose extending the continuous bag of words (CBOW) model (Mikolov et al., 2013) to learn style-sensitive word vectors using a wider context window under the assumption that the style of all the words in an utterance is consistent. In addition, we introduce a novel task to predict lexical stylistic similarity and to create a benchmark dataset for this task. Our experiment with this dataset supports our assumption and demonstrates that the proposed extensions contribute to the acquisition of style-sensitive word embeddings.

Jun Hatori, Yuta Kikuchi, Sosuke Kobayashi et al.

Comprehension of spoken natural language is an essential component for robots to communicate with human effectively. However, handling unconstrained spoken instructions is challenging due to (1) complex structures including a wide variety of expressions used in spoken language and (2) inherent ambiguity in interpretation of human instructions. In this paper, we propose the first comprehensive system that can handle unconstrained spoken language and is able to effectively resolve ambiguity in spoken instructions. Specifically, we integrate deep-learning-based object detection together with natural language processing technologies to handle unconstrained spoken instructions, and propose a method for robots to resolve instruction ambiguity through dialogue. Through our experiments on both a simulated environment as well as a physical industrial robot arm, we demonstrate the ability of our system to understand natural instructions from human operators effectively, and how higher success rates of the object picking task can be achieved through an interactive clarification process.

Sosuke Kobayashi, Naoaki Okazaki, Kentaro Inui

This study addresses the problem of identifying the meaning of unknown words or entities in a discourse with respect to the word embedding approaches used in neural language models. We proposed a method for on-the-fly construction and exploitation of word embeddings in both the input and output layers of a neural model by tracking contexts. This extends the dynamic entity representation used in Kobayashi et al. (2016) and incorporates a copy mechanism proposed independently by Gu et al. (2016) and Gulcehre et al. (2016). In addition, we construct a new task and dataset called Anonymized Language Modeling for evaluating the ability to capture word meanings while reading. Experiments conducted using our novel dataset show that the proposed variant of RNN language model outperformed the baseline model. Furthermore, the experiments also demonstrate that dynamic updates of an output layer help a model predict reappearing entities, whereas those of an input layer are effective to predict words following reappearing entities.