Ji‐Hwan Kim

WonJun Moon, Ji-Hwan Kim, Jae-Pil Heo

Recently, it is shown that deploying a proper self-supervision is a prospective way to enhance the performance of supervised learning. Yet, the benefits of self-supervision are not fully exploited as previous pretext tasks are specialized for unsupervised representation learning. To this end, we begin by presenting three desirable properties for such auxiliary tasks to assist the supervised objective. First, the tasks need to guide the model to learn rich features. Second, the transformations involved in the self-supervision should not significantly alter the training distribution. Third, the tasks are preferred to be light and generic for high applicability to prior arts. Subsequently, to show how existing pretext tasks can fulfill these and be tailored for supervised learning, we propose a simple auxiliary self-supervision task, predicting localizable rotation (LoRot). Our exhaustive experiments validate the merits of LoRot as a pretext task tailored for supervised learning in terms of robustness and generalization capability. Our code is available at https://github.com/wjun0830/Localizable-Rotation.

Jihwan Kim, Miso Lee, Jae-Pil Heo

Temporal Action Detection (TAD) is challenging but fundamental for real-world video applications. Recently, DETR-based models have been devised for TAD but have not performed well yet. In this paper, we point out the problem in the self-attention of DETR for TAD; the attention modules focus on a few key elements, called temporal collapse problem. It degrades the capability of the encoder and decoder since their self-attention modules play no role. To solve the problem, we propose a novel framework, Self-DETR, which utilizes cross-attention maps of the decoder to reactivate self-attention modules. We recover the relationship between encoder features by simple matrix multiplication of the cross-attention map and its transpose. Likewise, we also get the information within decoder queries. By guiding collapsed self-attention maps with the guidance map calculated, we settle down the temporal collapse of self-attention modules in the encoder and decoder. Our extensive experiments demonstrate that Self-DETR resolves the temporal collapse problem by keeping high diversity of attention over all layers.

Won Ik Cho, Yoon Kyung Lee, Seoyeon Bae et al.

Building a natural language dataset requires caution since word semantics is vulnerable to subtle text change or the definition of the annotated concept. Such a tendency can be seen in generative tasks like question-answering and dialogue generation and also in tasks that create a categorization-based corpus, like topic classification or sentiment analysis. Open-domain conversations involve two or more crowdworkers freely conversing about any topic, and collecting such data is particularly difficult for two reasons: 1) the dataset should be ``crafted" rather than ``obtained" due to privacy concerns, and 2) paid creation of such dialogues may differ from how crowdworkers behave in real-world settings. In this study, we tackle these issues when creating a large-scale open-domain persona dialogue corpus, where persona implies that the conversation is performed by several actors with a fixed persona and user-side workers from an unspecified crowd.

Yerim Jeon, Subeen Lee, Jihwan Kim et al.

Few-shot object counting has garnered significant attention for its practicality as it aims to count target objects in a query image based on given exemplars without additional training. However, the prevailing extract-and-match approach has a shortcoming: query and exemplar features lack interaction during feature extraction since they are extracted independently and later correlated based on similarity. This can lead to insufficient target awareness and confusion in identifying the actual target when multiple class objects coexist. To address this, we propose a novel framework, Mutually-Aware FEAture learning (MAFEA), which encodes query and exemplar features with mutual awareness from the outset. By encouraging interaction throughout the pipeline, we obtain target-aware features robust to a multi-category scenario. Furthermore, we introduce background token to effectively associate the query's target region with exemplars and decouple its background region. Our extensive experiments demonstrate that our model achieves state-of-the-art performance on FSCD-LVIS and FSC-147 benchmarks with remarkably reduced target confusion.

Alex Mathai, Chenxi Huang, Suwei Ma et al.

Code large language models (LLMs) have shown impressive capabilities on a multitude of software engineering tasks. In particular, they have demonstrated remarkable utility in the task of code repair. However, common benchmarks used to evaluate the performance of code LLMs are often limited to small-scale settings. In this work, we build upon kGym, which shares a benchmark for system-level Linux kernel bugs and a platform to run experiments on the Linux kernel. This paper introduces CrashFixer, the first LLM-based software repair agent that is applicable to Linux kernel bugs. Inspired by the typical workflow of a kernel developer, we identify the key capabilities an expert developer leverages to resolve a kernel crash. Using this as our guide, we revisit the kGym platform and identify key system improvements needed to practically run LLM-based agents at the scale of the Linux kernel (50K files and 20M lines of code). We implement these changes by extending kGym to create an improved platform - called kGymSuite, which will be open-sourced. Finally, the paper presents an evaluation of various repair strategies for such complex kernel bugs and showcases the value of explicitly generating a hypothesis before attempting to fix bugs in complex systems such as the Linux kernel. We also evaluated CrashFixer's capabilities on still open bugs, and found at least two patch suggestions considered plausible to resolve the reported bug.

Jihwan Kim, Junoh Kang, Jinyoung Choi et al.

We propose a novel inference technique based on a pretrained diffusion model for text-conditional video generation. Our approach, called FIFO-Diffusion, is conceptually capable of generating infinitely long videos without additional training. This is achieved by iteratively performing diagonal denoising, which simultaneously processes a series of consecutive frames with increasing noise levels in a queue; our method dequeues a fully denoised frame at the head while enqueuing a new random noise frame at the tail. However, diagonal denoising is a double-edged sword as the frames near the tail can take advantage of cleaner frames by forward reference but such a strategy induces the discrepancy between training and inference. Hence, we introduce latent partitioning to reduce the training-inference gap and lookahead denoising to leverage the benefit of forward referencing. Practically, FIFO-Diffusion consumes a constant amount of memory regardless of the target video length given a baseline model, while well-suited for parallel inference on multiple GPUs. We have demonstrated the promising results and effectiveness of the proposed methods on existing text-to-video generation baselines. Generated video examples and source codes are available at our project page.

Miso Lee, Jihwan Kim, Jae-Pil Heo

Multi-scene absolute pose regression addresses the demand for fast and memory-efficient camera pose estimation across various real-world environments. Nowadays, transformer-based model has been devised to regress the camera pose directly in multi-scenes. Despite its potential, transformer encoders are underutilized due to the collapsed self-attention map, having low representation capacity. This work highlights the problem and investigates it from a new perspective: distortion of query-key embedding space. Based on the statistical analysis, we reveal that queries and keys are mapped in completely different spaces while only a few keys are blended into the query region. This leads to the collapse of the self-attention map as all queries are considered similar to those few keys. Therefore, we propose simple but effective solutions to activate self-attention. Concretely, we present an auxiliary loss that aligns queries and keys, preventing the distortion of query-key space and encouraging the model to find global relations by self-attention. In addition, the fixed sinusoidal positional encoding is adopted instead of undertrained learnable one to reflect appropriate positional clues into the inputs of self-attention. As a result, our approach resolves the aforementioned problem effectively, thus outperforming existing methods in both outdoor and indoor scenes.

Sung Hwan Mun, Min Hyun Han, Dongjune Lee et al.

In this paper, we propose self-supervised speaker representation learning strategies, which comprise of a bootstrap equilibrium speaker representation learning in the front-end and an uncertainty-aware probabilistic speaker embedding training in the back-end. In the front-end stage, we learn the speaker representations via the bootstrap training scheme with the uniformity regularization term. In the back-end stage, the probabilistic speaker embeddings are estimated by maximizing the mutual likelihood score between the speech samples belonging to the same speaker, which provide not only speaker representations but also data uncertainty. Experimental results show that the proposed bootstrap equilibrium training strategy can effectively help learn the speaker representations and outperforms the conventional methods based on contrastive learning. Also, we demonstrate that the integrated two-stage framework further improves the speaker verification performance on the VoxCeleb1 test set in terms of EER and MinDCF.

Soonshin Seo, Ji-Hwan Kim

One of the most important parts of an end-to-end speaker verification system is the speaker embedding generation. In our previous paper, we reported that shortcut connections-based multi-layer aggregation improves the representational power of the speaker embedding. However, the number of model parameters is relatively large and the unspecified variations increase in the multi-layer aggregation. Therefore, we propose a self-attentive multi-layer aggregation with feature recalibration and normalization for end-to-end speaker verification system. To reduce the number of model parameters, the ResNet, which scaled channel width and layer depth, is used as a baseline. To control the variability in the training, a self-attention mechanism is applied to perform the multi-layer aggregation with dropout regularizations and batch normalizations. Then, a feature recalibration layer is applied to the aggregated feature using fully-connected layers and nonlinear activation functions. Deep length normalization is also used on a recalibrated feature in the end-to-end training process. Experimental results using the VoxCeleb1 evaluation dataset showed that the performance of the proposed methods was comparable to that of state-of-the-art models (equal error rate of 4.95% and 2.86%, using the VoxCeleb1 and VoxCeleb2 training datasets, respectively).

Soonshin Seo, Ji-Hwan Kim

In general, a self-attention mechanism has been applied for speaker embedding encoding. Previous studies focused on training the self-attention in a high-level layer, such as the last pooling layer. However, the effect of low-level features was reduced in the speaker embedding encoding. Therefore, we propose masked cross self-attentive encoding (MCSAE) using ResNet. It focuses on the features of both high-level and lowlevel layers. Based on multi-layer aggregation, the output features of each residual layer are used for the MCSAE. In the MCSAE, cross self-attention module is trained the interdependence of each input features. A random masking regularization module also applied to preventing overfitting problem. As such, the MCSAE enhances the weight of frames representing the speaker information. Then, the output features are concatenated and encoded to the speaker embedding. Therefore, a more informative speaker embedding is encoded by using the MCSAE. The experimental results showed an equal error rate of 2.63% and a minimum detection cost function of 0.1453 using the VoxCeleb1 evaluation dataset. These were improved performances compared with the previous self-attentive encoding and state-of-the-art encoding methods.

Jisung Wang, Jihwan Kim, Sangki Kim et al.

As the character-based end-to-end automatic speech recognition (ASR) models evolve, the choice of acoustic modeling units becomes important. Since Korean is a fairly phonetic language and has a unique writing system with its own Korean alphabet, it's worth investigating modeling units for an end-to-end Korean ASR task. In this work, we introduce lexicon-free modeling units in Korean, and explore them using a hybrid CTC/Attention-based encoder-decoder model. Five lexicon-free units are investigated: Syllable-based Korean character (with English character for a code-switching task), Korean Jamo character (with English character), sub-word on syllable-based character (with sub-word in English), sub-word on Jamo character (with sub-words in English), and finally byte unit, which is a universal one across language. Experiments on Zeroth-Korean (51.6 hrs) and Medical Record (2530 hrs) are done for Korean and Korean-English code-switching ASR tasks, respectively. Sequence-to-sequence learning with sub-words based on Korean syllables (and sub-words in English) performs the best for both tasks without lexicon and an extra language model integration.

Minkyu Lim, Ji-Hwan Kim

While the Kaldi framework provides state-of-the-art components for speech recognition like feature extraction, deep neural network (DNN)-based acoustic models, and a weighted finite state transducer (WFST)-based decoder, it is difficult to implement a new flexible DNN model. By contrast, a general-purpose deep learning framework, such as TensorFlow, can easily build various types of neural network architectures using a tensor-based computation method, but it is difficult to apply them to WFST-based speech recognition. In this study, a TensorFlow-based acoustic model is integrated with a WFST-based Kaldi decoder to combine the two frameworks. The features and alignments used in Kaldi are converted so they can be trained by the TensorFlow model, and the DNN-based acoustic model is then trained. In the integrated Kaldi decoder, the posterior probabilities are calculated by querying the trained TensorFlow model, and a beam search is performed to generate the lattice. The advantages of the proposed one-pass decoder include the application of various types of neural networks to WFST-based speech recognition and WFST-based online decoding using a TensorFlow-based acoustic model. The TensorFlow based acoustic models trained using the RM, WSJ, and LibriSpeech datasets show the same level of performance as the model trained using the Kaldi framework.

Hosung Park, Donghyun Lee, Minkyu Lim et al.

In this paper, a time delay neural network (TDNN) based acoustic model is proposed to implement a fast-converged acoustic modeling for Korean speech recognition. The TDNN has an advantage in fast-convergence where the amount of training data is limited, due to subsampling which excludes duplicated weights. The TDNN showed an absolute improvement of 2.12% in terms of character error rate compared to feed forward neural network (FFNN) based modelling for Korean speech corpora. The proposed model converged 1.67 times faster than a FFNN-based model did.