diff --git a/_bibliography/papers.bib b/_bibliography/papers.bib index 1b58a61..2217bc3 100644 --- a/_bibliography/papers.bib +++ b/_bibliography/papers.bib @@ -1377,7 +1377,8 @@ @inproceedings{Agarwal_2012 abbr = {PODS}, selected = {true}, field = {Streaming Algorithms}, - abstract = {We study the mergeability of data summaries and propose methods to make summaries mergeable while preserving error and size guarantees.} + abstract = {We study the mergeability of data summaries and propose methods to make summaries mergeable while preserving error and size guarantees.}, + award = {Test of Time Award} } diff --git a/_projects/sketch4ML.md b/_projects/sketch4ML.md index 6b5fc3e..b868e0b 100644 --- a/_projects/sketch4ML.md +++ b/_projects/sketch4ML.md @@ -6,11 +6,16 @@ importance: 3 category: Projects related_publications: true related_posts: false +tabs: true mermaid: enabled: true zoomable: false --- +{% tabs group-name %} + +{% tab group-name Background %} + > This project focuses on the streaming data algorithms, which maintain a small data structure in memory, and the applications of these algorithms in machine learning, which are called Sketch4ML. The streaming algorithms can be used to optimize the machine learning algorithms, which are usually memory-intensive and time-consuming. {: .block-tip } @@ -24,6 +29,23 @@ In computer science, **streaming algorithms** are algorithms for processing data As a result of these constraints, streaming algorithms often produce approximate answers based on a **sketch** of the data stream. +### Books and Papers + + + ## Sketch for Machine Learning (Sketch4ML) **Sketch for Machine Learning (Sketch4ML)** is a technique that uses streaming algorithms to optimize machine learning algorithms. For example, matrix sketching has been employed to accelerate **second-order online gradient descent** (SON, [Luo et al., 2016](https://papers.nips.cc/paper_files/paper/2016/hash/15de21c670ae7c3f6f3f1f37029303c9-Abstract.html); RFD-ONS, [Luo et al., 2019](https://www.jmlr.org/papers/v20/17-773.html)), **online kernel learning** ([Calandriello et al., 2017](https://proceedings.neurips.cc/paper/2017/hash/366f0bc7bd1d4bf414073cabbadfdfcd-Abstract.html)), and **linear contextual bandits** (SOFUL, [Kuzborskij et al., 2019](https://proceedings.mlr.press/v89/kuzborskij19a.html); CBSCFD [Chen et al., 2020](https://www.ijcai.org/Proceedings/2020/0588.pdf); DBSLinUCB), as shown in the following figure. @@ -42,7 +64,9 @@ graph LR D --> D3["DBSLinUCB"] ``` -### Optimal Matrix Sketching over Sliding Windows [[VLDB 2024](https://vldb.org/2024/)] +{% endtab %} + +{% tab group-name our work:
optimal matrix sketch
over sliding windows %}

@@ -54,9 +78,6 @@ graph LR arxiv - - dimensions - stars @@ -67,6 +88,17 @@ graph LR In this work, we proposes the optimal matrix sketch algorithm DS-FD on sliding windows, which achives the lower bound of space complexity for solving the matrix sketching problem over sliding windows. The paper addressed the open question of the lower bounds of space complexity for any deterministic algorithms solving the matrix sketching problem over sliding windows. The answer to this open problem confirms that our DS-FD algorithm is optimal in terms of space complexity. +The paper {% cite yin2024optimal %} is accepted by VLDB 2024 and nominated for the Best Research Paper. If you are interested in the details, please refer to the [paper](https://doi.org/10.14778/3665844.3665847), [arxiv](https://arxiv.org/abs/2405.07792) or the [code](https://github.com/yinhanyan/DS-FD). + +### Problem Definition + + +### Method + + +### Results + +

{% include figure.liquid loading="eager" path="assets/img/optimal2024yin.png" title="example image" class="img-fluid rounded z-depth-1" %}
@@ -74,20 +106,12 @@ In this work, we proposes the optimal matrix sketch algorithm DS-FD on sliding w The space upper bound of DS-FD and lower bound in various senarios. -The paper {% cite yin2024optimal %} is accepted by VLDB 2024 and nominated for the Best Research Paper. If you are interested in the details, please refer to the [paper](https://doi.org/10.14778/3665844.3665847), [arxiv](https://arxiv.org/abs/2405.07792) or the [code](https://github.com/yinhanyan/DS-FD). +{% endtab %} - -## Sketch4ML - -### Dyadic Block Sketching +{% tab group-name our work:
Dyadic Sketched Bandit %} TBD +{% endtab %} - -## Roadmap - -- [x] Streaming Algorithms - - [x] Frequent Directions over Sliding Windows -- [ ] Sketch4ML - - [ ] Dyadic Block Sketching \ No newline at end of file +{% endtabs %} \ No newline at end of file