Differential Transformer has recently gained significant attention for its impressive empirical performance, often attributed to its ability to perform noise canceled attention. However, precisely how differential attention achieves its empirical benefits remains poorly understood. Moreover, Differential Transformer architecture demands large-scale training from scratch, hindering utilization of open pretrained weights. In this work, we conduct an in-depth investigation of Differential Transformer, uncovering three key factors behind its success: (1) enhanced expressivity via negative attention, (2) reduced redundancy among attention heads, and (3) improved learning dynamics. Based on these findings, we propose DEX, a novel method to efficiently integrate the advantages of differential attention into pretrained language models. By reusing the softmax attention scores and adding a lightweight differential operation on the output value matrix, DEX effectively incorporates the key advantages of differential attention while remaining lightweight in both training and inference. Evaluations confirm that DEX substantially improves the pretrained LLMs across diverse benchmarks, achieving significant performance gains with minimal adaptation data (< 0.01%).

Contrastive learning is a form of distance learning that aims to learn invariant features from two related representations. In this work, we explore the hypothesis that an image and caption can be regarded as two different views of the underlying mutual information, and train a model to learn a unified vision-language representation space that encodes both modalities at once in a modality-agnostic manner. We first identify difficulties in learning a one-tower model for vision-language pretraining (VLP), and propose One Representation (OneR) as a simple yet effective framework for our goal. We discover intriguing properties that distinguish OneR from the previous works that have modality-specific representation spaces such as zero-shot localization, text-guided visual reasoning and multi-modal retrieval, and present analyses to provide insights into this new form of multi-modal representation learning. Thorough evaluations demonstrate the potential of a unified modality-agnostic VLP framework.

State-of-the-art frameworks in self-supervised learning have recently shown that fully utilizing transformer-based models can lead to performance boost compared to conventional CNN models. Striving to maximize the mutual information of two views of an image, existing works apply a contrastive loss to the final representations. Motivated by self-distillation in the supervised regime, we further exploit this by allowing the intermediate representations to learn from the final layer via the contrastive loss. Through self-distillation, the intermediate layers are better suited for instance discrimination, making the performance of an early-exited sub-network not much degraded from that of the full network. This renders the pretext task easier also for the final layer, leading to better representations. Our method, Self-Distilled Self-Supervised Learning (SDSSL), outperforms competitive baselines (SimCLR, BYOL and MoCo v3) using ViT on various tasks and datasets. In the linear evaluation and k-NN protocol, SDSSL not only leads to superior performance in the final layers, but also in most of the lower layers. Furthermore, qualitative and quantitative analyses show how representations are formed more effectively along the transformer layers. Code is available at https://github.com/hagiss/SDSSL.

Fashion attribute editing is a task that aims to convert the semantic attributes of a given fashion image while preserving the irrelevant regions. Previous works typically employ conditional GANs where the generator explicitly learns the target attributes and directly execute the conversion. These approaches, however, are neither scalable nor generic as they operate only with few limited attributes and a separate generator is required for each dataset or attribute set. Inspired by the recent advancement of diffusion models, we explore the classifier-guided diffusion that leverages the off-the-shelf diffusion model pretrained on general visual semantics such as Imagenet. In order to achieve a generic editing pipeline, we pose this as multi-attribute image manipulation task, where the attribute ranges from item category, fabric, pattern to collar and neckline. We empirically show that conventional methods fail in our challenging setting, and study efficient adaptation scheme that involves recently introduced attention-pooling technique to obtain a multi-attribute classifier guidance. Based on this, we present a mask-free fashion attribute editing framework that leverages the classifier logits and the cross-attention map for manipulation. We empirically demonstrate that our framework achieves convincing sample quality and attribute alignments.