arXiv subjects

Jingsong Sun

Publications and source records attributed to Jingsong Sun.

3 recordsLinked to original sources

arXiv2024

Exploring the Robustness of In-Context Learning with Noisy Labels

Recently, the mysterious In-Context Learning (ICL) ability exhibited by Transformer architectures, especially in large language models (LLMs), has sparked significant research interest. However, the resilience of Transformers' in-context learning capabilities in the presence of noisy samples, prevalent in both training corpora and prompt demonstrations, remains underexplored. In this paper, inspired by prior research that studies ICL ability using simple function classes, we take a closer look at this problem by investigating the robustness of Transformers against noisy labels. Specifically, we first conduct a thorough evaluation and analysis of the robustness of Transformers against noisy labels during in-context learning and show that they exhibit notable resilience against diverse types of noise in demonstration labels. Furthermore, we delve deeper into this problem by exploring whether introducing noise into the training set, akin to a form of data augmentation, enhances such robustness during inference, and find that such noise can indeed improve the robustness of ICL. Overall, our fruitful analysis and findings provide a comprehensive understanding of the resilience of Transformer models against label noises during ICL and provide valuable insights into the research on Transformers in natural language processing. Our code is available at https://github.com/InezYu0928/in-context-learning.

cs.CL↗

arXiv2022

Weak Hardy Spaces Associated with Ball Quasi-Banach Function Spaces on Spaces of Homogeneous Type: Decompositions, Real Interpolation, and Calderón--Zygmund Operators

Let $(\mathbb{X},d,μ)$ be a space of homogeneous type in the sense of R. R. Coifman and G. Weiss, and $X(\mathbb{X})$ a ball quasi-Banach function space on $\mathbb{X}$. In this article, the authors introduce the weak Hardy space $WH_X(\mathbb{X})$ associated with $X(\mathbb{X})$ via the grand maximal function, and characterize $WH_X(\mathbb{X})$ by other maximal functions and atoms. The authors then apply these characterizations to obtain the real interpolation and the boundedness of Calderón--Zygmund operators in the critical case. The main novelties of this article exist in that the authors use the Aoki--Rolewicz theorem and both the dyadic system and the exponential decay of approximations of the identity on $\mathbb{X}$, which closely connect with the geometrical properties of $\mathbb{X}$, to overcome the difficulties caused by the absence of both the triangle inequality of $\|\cdot\|_{X(\mathbb{X})}$ and the reverse doubling assumption of the measure $μ$ under consideration, and also use the relation between the convexification of $X(\mathbb{X})$ and the weak space $WX(\mathbb{X})$ associated with $X(\mathbb{X})$ to prove that the infinite summation of atoms converges in the space of distributions on $\mathbb{X}$. Moreover, all these results have a wide range of generality and, particularly, even when they are applied to the weighted Lebesgue space, the Orlicz space, and the variable Lebesgue space, the obtained results are also new and, actually, some of them are new even on RD-spaces (namely, spaces of homogeneous type satisfying the additional reverse doubling condition).

math.FA↗

arXiv2019

Localized John--Nirenberg--Campanato Spaces

Let $p\in(1,\infty)$, $q\in[1,\infty)$, $s\in{\mathbb Z}_{+}$, $α\in[0,\infty)$ and $\mathcal{X}$ be $\mathbb R^n$ or a cube $Q_0\subsetneqq\mathbb R^n$. In this article, the authors first introduce the localized John--Nirenberg--Campanato space $jn_{(p,q,s)_α}(\mathcal{X})$ and show that the localized Campanato space is the limit case of $jn_{(p,q,s)_α}(\mathcal{X})$ as $p\to\infty$. By means of local atoms and the weak-$*$ topology, the authors then introduce the localized Hardy-kind space $hk_{(p',q',s)_α}(\mathcal{X})$ which proves the predual space of $jn_{(p,q,s)_α}(\mathcal{X})$. Moreover, the authors prove that $hk_{(p',q',s)_α}(\mathcal{X})$ is invariant when $1<q<p$, where $p'$ or $q'$ denotes the conjugate number of $p$ or $q$, respectively. All these results are new even for the localized John--Nirenberg space.

math.CA↗