Kangil Kim

Hee-Jun Jung, Doyeon Kim, Seung-Hoon Na et al.

Knowledge distillation is an approach to transfer information on representations from a teacher to a student by reducing their difference. A challenge of this approach is to reduce the flexibility of the student's representations inducing inaccurate learning of the teacher's knowledge. To resolve it in transferring, we investigate distillation of structures of representations specified to three types: intra-feature, local inter-feature, global inter-feature structures. To transfer them, we introduce feature structure distillation methods based on the Centered Kernel Alignment, which assigns a consistent value to similar features structures and reveals more informative relations. In particular, a memory-augmented transfer method with clustering is implemented for the global structures. The methods are empirically analyzed on the nine tasks for language understanding of the GLUE dataset with Bidirectional Encoder Representations from Transformers (BERT), which is a representative neural language model. In the results, the proposed methods effectively transfer the three types of structures and improve performance compared to state-of-the-art distillation methods. Indeed, the code for the methods is available in https://github.com/maroo-sky/FSD.

Hoyong Kim, Minchan Kwon, Kangil Kim

In replay-based methods for continual learning, replaying input samples in episodic memory has shown its effectiveness in alleviating catastrophic forgetting. However, the potential key factor of cross-entropy loss with softmax in causing catastrophic forgetting has been underexplored. In this paper, we analyze the effect of softmax and revisit softmax masking with negative infinity to shed light on its ability to mitigate catastrophic forgetting. Based on the analyses, it is found that negative infinity masked softmax is not always compatible with dark knowledge. To improve the compatibility, we propose a general masked softmax that controls the stability by adjusting the gradient scale to old and new classes. We demonstrate that utilizing our method on other replay-based methods results in better performance, primarily by enhancing model stability in continual learning benchmarks, even when the buffer size is set to an extremely small value.

Sangmin Woo, Junhyug Noh, Kangil Kim

Identifying relations between objects is central to understanding the scene. While several works have been proposed for relation modeling in the image domain, there have been many constraints in the video domain due to challenging dynamics of spatio-temporal interactions (e.g., between which objects are there an interaction? when do relations start and end?). To date, two representative methods have been proposed to tackle Video Visual Relation Detection (VidVRD): segment-based and window-based. We first point out limitations of these methods and propose a novel approach named Temporal Span Proposal Network (TSPN). TSPN tells what to look: it sparsifies relation search space by scoring relationness of object pair, i.e., measuring how probable a relation exist. TSPN tells when to look: it simultaneously predicts start-end timestamps (i.e., temporal spans) and categories of the all possible relations by utilizing full video context. These two designs enable a win-win scenario: it accelerates training by 2X or more than existing methods and achieves competitive performance on two VidVRD benchmarks (ImageNet-VidVDR and VidOR). Moreover, comprehensive ablative experiments demonstrate the effectiveness of our approach. Codes are available at https://github.com/sangminwoo/Temporal-Span-Proposal-Network-VidVRD.

Jinwook Park, Kangil Kim

Neural parameterization has significantly advanced unsupervised grammar induction. However, training these models with a traditional likelihood loss for all possible parses exacerbates two issues: 1) $\textit{structural optimization ambiguity}$ that arbitrarily selects one among structurally ambiguous optimal grammars despite the specific preference of gold parses, and 2) $\textit{structural simplicity bias}$ that leads a model to underutilize rules to compose parse trees. These challenges subject unsupervised neural grammar induction (UNGI) to inevitable prediction errors, high variance, and the necessity for extensive grammars to achieve accurate predictions. This paper tackles these issues, offering a comprehensive analysis of their origins. As a solution, we introduce $\textit{sentence-wise parse-focusing}$ to reduce the parse pool per sentence for loss evaluation, using the structural bias from pre-trained parsers on the same dataset. In unsupervised parsing benchmark tests, our method significantly improves performance while effectively reducing variance and bias toward overly simplistic parses. Our research promotes learning more compact, accurate, and consistent explicit grammars, facilitating better interpretability.

Jinwook Park, Kangil Kim

Unsupervised neural grammar induction aims to learn interpretable hierarchical structures from language data. However, existing models face an expressiveness bottleneck, often resulting in unnecessarily large yet underperforming grammars. We identify a core issue, $\textit{probability distribution collapse}$, as the underlying cause of this limitation. We analyze when and how the collapse emerges across key components of neural parameterization and introduce a targeted solution, $\textit{collapse-relaxing neural parameterization}$, to mitigate it. Our approach substantially improves parsing performance while enabling the use of significantly more compact grammars across a wide range of languages, as demonstrated through extensive empirical analysis.

Minchan Kwon, Junwon Ko, Kangil Kim et al.

Direct Preference Optimization (DPO) is a powerful paradigm to align language models with human preferences using pairwise comparisons. However, its binary win-or-loss supervision often proves insufficient for training small models with limited capacity. Prior works attempt to distill information from large teacher models using behavior cloning or KL divergence. These methods often focus on mimicking current behavior and overlook distilling reward modeling. To address this issue, we propose \textit{Teacher Value-based Knowledge Distillation} (TVKD), which introduces an auxiliary reward from the value function of the teacher model to provide a soft guide. This auxiliary reward is formulated to satisfy potential-based reward shaping, ensuring that the global reward structure and optimal policy of DPO are preserved. TVKD can be integrated into the standard DPO training framework and does not require additional rollouts. Our experimental results show that TVKD consistently improves performance across various benchmarks and model sizes.

Junsik Kim, Jinwook Park, Kangil Kim

In knowledge graph embedding, leveraging relation specific entity transformation has markedly enhanced performance. However, the consistency of embedding differences before and after transformation remains unaddressed, risking the loss of valuable inductive bias inherent in the embeddings. This inconsistency stems from two problems. First, transformation representations are specified for relations in a disconnected manner, allowing dissimilar transformations and corresponding entity embeddings for similar relations. Second, a generalized plug-in approach as a SFBR (Semantic Filter Based on Relations) disrupts this consistency through excessive concentration of entity embeddings under entity-based regularization, generating indistinguishable score distributions among relations. In this paper, we introduce a plug-in KGE method, Relation-Semantics Consistent Filter (RSCF). Its entity transformation has three features for enhancing semantic consistency: 1) shared affine transformation of relation embeddings across all relations, 2) rooted entity transformation that adds an entity embedding to its change represented by the transformed vector, and 3) normalization of the change to prevent scale reduction. To amplify the advantages of consistency that preserve semantics on embeddings, RSCF adds relation transformation and prediction modules for enhancing the semantics. In knowledge graph completion tasks with distance-based and tensor decomposition models, RSCF significantly outperforms state-of-the-art KGE methods, showing robustness across all relations and their frequencies.

Hee-Jun Jung, Jaehyoung Jeong, Kangil Kim

Disentanglement learning is a core issue for understanding and re-using trained information in Variational AutoEncoder (VAE), and effective inductive bias has been reported as a key factor. However, the actual implementation of such bias is still vague. In this paper, we propose a novel method, called Multiple Invertible and partial-equivariant transformation (MIPE-transformation), to inject inductive bias by 1) guaranteeing the invertibility of latent-to-latent vector transformation while preserving a certain portion of equivariance of input-to-latent vector transformation, called Invertible and partial-equivariant transformation (IPE-transformation), 2) extending the form of prior and posterior in VAE frameworks to an unrestricted form through a learnable conversion to an approximated exponential family, called Exponential Family conversion (EF-conversion), and 3) integrating multiple units of IPE-transformation and EF-conversion, and their training. In experiments on 3D Cars, 3D Shapes, and dSprites datasets, MIPE-transformation improves the disentanglement performance of state-of-the-art VAEs.

Hoyong Kim, Semi Lee, Kangil Kim

In the feature space, the collapse between features invokes critical problems in representation learning by remaining the features undistinguished. Interpolation-based augmentation methods such as mixup have shown their effectiveness in relieving the collapse problem between different classes, called inter-class collapse. However, intra-class collapse raised in coarse-to-fine transfer learning has not been discussed in the augmentation approach. To address them, we propose a better feature augmentation method, asymptotic midpoint mixup. The method generates augmented features by interpolation but gradually moves them toward the midpoint of inter-class feature pairs. As a result, the method induces two effects: 1) balancing the margin for all classes and 2) only moderately broadening the margin until it holds maximal confidence. We empirically analyze the collapse effects by measuring alignment and uniformity with visualizing representations. Then, we validate the intra-class collapse effects in coarse-to-fine transfer learning and the inter-class collapse effects in imbalanced learning on long-tailed datasets. In both tasks, our method shows better performance than other augmentation methods.

Hee-Jun Jung, Jaehyoung Jeong, Kangil Kim

Symmetries of input and latent vectors have provided valuable insights for disentanglement learning in VAEs. However, only a few works were proposed as an unsupervised method, and even these works require known factor information in the training data. We propose a novel method, Composite Factor-Aligned Symmetry Learning (CFASL), which is integrated into VAEs for learning symmetry-based disentanglement in unsupervised learning without any knowledge of the dataset factor information. CFASL incorporates three novel features for learning symmetry-based disentanglement: 1) Injecting inductive bias to align latent vector dimensions to factor-aligned symmetries within an explicit learnable symmetry code-book 2) Learning a composite symmetry to express unknown factors change between two random samples by learning factor-aligned symmetries within the codebook 3) Inducing a group equivariant encoder and decoder in training VAEs with the two conditions. In addition, we propose an extended evaluation metric for multi-factor changes in comparison to disentanglement evaluation in VAEs. In quantitative and in-depth qualitative analysis, CFASL demonstrates a significant improvement of disentanglement in single-factor change, and multi-factor change conditions compared to state-of-the-art methods.

Minchan Kwon, Kangil Kim

In real life, adversarial attack to deep learning models is a fatal security issue. However, the issue has been rarely discussed in a widely used class-incremental continual learning (CICL). In this paper, we address problems of applying adversarial training to CICL, which is well-known defense method against adversarial attack. A well-known problem of CICL is class-imbalance that biases a model to the current task by a few samples of previous tasks. Meeting with the adversarial training, the imbalance causes another imbalance of attack trials over tasks. Lacking clean data of a minority class by the class-imbalance and increasing of attack trials from a majority class by the secondary imbalance, adversarial training distorts optimal decision boundaries. The distortion eventually decreases both accuracy and robustness than adversarial training. To exclude the effects, we propose a straightforward but significantly effective method, External Adversarial Training (EAT) which can be applied to methods using experience replay. This method conduct adversarial training to an auxiliary external model for the current task data at each time step, and applies generated adversarial examples to train the target model. We verify the effects on a toy problem and show significance on CICL benchmarks of image classification. We expect that the results will be used as the first baseline for robustness research of CICL.

HeeSeung Jung, Kangil Kim, Hoyong Kim et al.

The flexibility of decision boundaries in neural networks that are unguided by training data is a well-known problem typically resolved with generalization methods. A surprising result from recent knowledge distillation (KD) literature is that random, untrained, and equally structured teacher networks can also vastly improve generalization performance. It raises the possibility of existence of undiscovered assumptions useful for generalization on an uncertain region. In this paper, we shed light on the assumptions by analyzing decision boundaries and confidence distributions of both simple and KD-based generalization methods. Assuming that a decision boundary exists to represent the most general tendency of distinction on an input sample space (i.e., the simplest hypothesis), we show the various limitations of methods when using the hypothesis. To resolve these limitations, we propose matured dumb teacher based KD, conservatively transferring the hypothesis for generalization of the student without massive destruction of trained information. In practical experiments on feed-forward and convolution neural networks for image classification tasks on MNIST, CIFAR-10, and CIFAR-100 datasets, the proposed method shows stable improvement to the best test performance in the grid search of hyperparameters. The analysis and results imply that the proposed method can provide finer generalization than existing methods.

Sangmin Woo, Junhyug Noh, Kangil Kim

In this work, we seek new insights into the underlying challenges of the Scene Graph Generation (SGG) task. Quantitative and qualitative analysis of the Visual Genome dataset implies -- 1) Ambiguity: even if inter-object relationship contains the same object (or predicate), they may not be visually or semantically similar, 2) Asymmetry: despite the nature of the relationship that embodied the direction, it was not well addressed in previous studies, and 3) Higher-order contexts: leveraging the identities of certain graph elements can help to generate accurate scene graphs. Motivated by the analysis, we design a novel SGG framework, Local-to-Global Interaction Networks (LOGIN). Locally, interactions extract the essence between three instances of subject, object, and background, while baking direction awareness into the network by explicitly constraining the input order of subject and object. Globally, interactions encode the contexts between every graph component (i.e., nodes and edges). Finally, Attract & Repel loss is utilized to fine-tune the distribution of predicate embeddings. By design, our framework enables predicting the scene graph in a bottom-up manner, leveraging the possible complementariness. To quantify how much LOGIN is aware of relational direction, a new diagnostic task called Bidirectional Relationship Classification (BRC) is also proposed. Experimental results demonstrate that LOGIN can successfully distinguish relational direction than existing methods (in BRC task), while showing state-of-the-art results on the Visual Genome benchmark (in SGG task).