关于少样本学习、零样本学习、单样本学习中的support set和query set的概念解析

梳理一下关于少样本学习(如few-shot learning)、零样本学习(zero-shot learning)、单样本学习(one-shot learning)中的support set和query set的概念,以及为何测试集中会涉及到“unseen”(未见过)的新类。

关键概念解析

  1. Support Set:在少样本学习任务中,support set是用来训练模型的一组有标签的样本,这些样本代表了任务中的类别。在训练过程中,每个类别只有少量样本(例如,在N-way K-shot任务中,每个类别有K个样本,总共有N个类别)。

  2. Query Set:query set包含了用来测试或调整模型的样本,这些样本同样来自于任务中的类别,但不包含在support set中。模型需要利用在support set上学到的知识来正确分类query set中的样本。

训练与测试过程

  • 训练过程:在训练阶段,模型通过学习support set中的样本来尝试理解每个类别的特征,并通过query set来调整和改善模型的性能。这一过程可能会重复多次,每次都使用不同的类别和样本来构成新的support set和query set。

  • 测试过程:在测试阶段,模型面对的是完全未见过的新类别(unseen classes)。这意味着这些类别既没有出现在训练阶段的support set中,也没有出现在query set中。测试过程中,模型需要使用在训练阶段学到的知识,来分类这些新类别的样本。

为何称之为"unseen"

当我们提到测试集中的类别是“unseen”的,意味着这些类别在训练阶段完全未被模型见过。这是少样本学习特别设计的挑战,目的是检验模型学习到的知识是否能够泛化到新的、未知的情况。即使support set中每个类别的样本非常少,模型也需要能够抓住这些样本的关键特征,并将这种能力迁移到全新的类别上。

例子中的概念应用

在您的例子中,训练过程选择了数据集中的一部分类别和样本来训练模型,使用剩下的样本作为query set来测试和调整模型。测试过程则针对完全新的类别进行,这些新类别之前未参与模型的训练,因此被称为“unseen”。模型需要利用之前学到的知识,通过与support set的样本进行比较,来正确识别query set中样本的类别。

这种设计模仿了人类面对新情况时的学习和适应能力,即用有限的信息快速学习新概念,并将学到的知识应用于识别未知类别的能力。希望这能帮助您更好地理解这些概念。

  • 16
    点赞
  • 10
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
PL/SQL Developer 11.0 - New Features Below you can find a brief description of the most important new features in PL/SQL Developer 11.0. New features SQL Window enhancements The result set toolbar has a new Compare records button to easily compare records in the result set: By default this will highlight identical field values in the result set when compared to the current record. If you move to a different row, the highlighted fields will be updated accordingly. Pressing the Compare records button again will clear the highlights. The arrow button next to the Compare records button controls whether identical or different values are highlighted, whether only the current column or all columns of the records are highlighted, and whether the field values are compare case sensitive. In Single Record mode, the SQL Window will now also display the column comments in the result set grid: Other enhancements include: A Refresh result set button has been added to refresh (re-execute) the result set of the currently selected result tab page Column totals (sum, count, min, max, avg) are now persistent after re-executing a query Added Excel exporter (both xls and xlsx support) Support has been added for Oracle12c extended strings (> 4000 bytes) The align button has a new Hide editor option to show only the result set The align button has a new Window width option to resize the SQL Window width to match the result set width You can now use an AltRowColor=[RRGGBB | name] directive in a comment section of the SQL Window to control the alternate row color You can now use a Records=[all|page|<n>] directive in a comment section of the SQL Window to control the number of initially fetched records You can now use a Totals=[mode:column] directive in a comment section of the SQL Window to display column totals
Chapter 1 overviews open research problems concerning building data warehouses for integrating and analyzing various complex types of data, dealing with temporal aspects of data, handling imprecise data, and ensuring privacy in DWs. Chapter 2 discusses challenges in designing ETL processes for real-time (or near real-time) data warehouses and proposes an architecture of a real-time DW system. Chapter 3 discusses data warehouse modeling techniques, based on multidimensional modeling. In particular, the chapter covers conceptual, logical, and physical modeling. Chapter 4 proposes an approach to personalizing a multidimensional database. The authors present a model and a language that allow to define user preferences on schema elements. User preferences are expressed by means of weights associated with schema elements and they express user interest in data. Chapter 5 covers designing spatial (geographical) data warehouses. It proposes a metamodel for the support of the design of spatial dimensional schemas. Chapter 6 presents a technique for approximate answering range-sum queries on data cubes. To this end, the authors propose tree-based data structures for separately storing sampled data and outliers data. The proposed technique assures a good quality of approximate answers. Chapter 7 addresses a problem of summarizing multidimensional search spaces, called data cubes. The authors propose the concept of the so-called closed cube, which is a cover for a data cube. The authors show that the closed cube is smaller than its competitor, i.e. a quotient cube, and it can be used for deriving a quotient cube. Chapter 8 analyzes multiple index structures for multiversion data warehouses. In particular, the paper describes how to extend index structures designed for data with linear evolution in order to handle data with branched evolution and it provides an analytical model for comparing various index structures for multiversion DWs. Chapter 9 discusses the application of WAH compressed bitmap indexes to indexing text data for full-text search. The chapter also presents performance characteristics of the proposed indexing technique in three systems, namely MySQL, FastBit, and MonetDB. Chapter 10 proposes the optimization of OLAP queries by means of applying horizontal partitioning of tables and bitmap join indexes. The partitioning schema viii Preface and the set of bitmap indexes are selected by means of genetic and greedy algorithms. The proposed optimization techniques are validated experimentally. Chapter 11 discusses the application of the x-BR-tree index to spatial data. The chapter provides an analytical cost model of spatial queries with the support of the x-BR-tree index, followed by its experimental evaluation. Chapter 12 proposes a formal model for representing spatio-temporal and nonspatial data about moving objects. Based on the model, the authors propose a query language allowing to query such data. The main idea is based on replacing object trajectories by sequences of object stops and moves. Chapter 13 addresses the problem of data mining in a multidimensional space in a data warehouse. The authors propose a compact representation of sequential patterns, called closed multidimensional sequential patterns, which allows to reduce the search space. The proposed representation and mining algorithms are followed by experimental evaluation. Chapter 14 presents issues on modeling and querying temporal semistructured data warehouses. Such a DW is modeled as a graph with labeled nodes and edges. The temporal aspect is added to the graph by means of labels denoting validity times. The model is supported with a query language based on path expressions. Chapter 15 contributes a multidimensional data model of a data warehouse, called "galaxy", that supports the analysis of XML documents. The authors also propose a technique and a tool for integrating XML documents and loading them into a DW.
IRPTrace2.00.002.zip ,IRP跟踪工具,支持Windows7。 IRPTrace 2.00.002, Build Date September 23, 2015 File name = README.TXT ====================================================================== CONTENTS ========= 1) IRPTrace Components 2) Release Notes 3) Known problems & limitations 4) Known bugs 5) Update to the documentation 6) List of supported I/O requests We strongly recommend that you read the following information about this release. 1) IRPTrace Components ===================== README.TXT - This file HOWTOREG.TXT - How to register IRPTrace and contact APSoft IRPTrace.EXE - Main application module IRPDRV.SYS - Driver for Windows NT/2000/XP/Server 2003 TERMINAL - Terminal log file TERMINAL.DLL - IRPTrace library UNINSDRV.DLL - IRPTrace library TSCUST.DLL - IRPTrace library IRPTRACE.CHM - Help file TIPS.TXT - 'Tip of the Day' tips GUID.INI - GUID database IRPTRACE.INI - Driver uninstallation information UNINST.ISU - Installation/uninstallation log file 2) Release Notes ================ Release 2.00.002 1. Fixed processing of IRPs at elevated IRQLs 2. Added support for Windows 10 Release (Build 10240) 3. Corrected processing of USB IOCTLs 4. Release 2.00.001 1. Added support for x64 Windows. 2. Added support for Windows Vista, 7 and 8.x. 3. Added decoding of all USB kernel-mode messages. 4. Added decoding of all CDB messages. 5. Revised Terminal tracing. 6. Revised on-line help file. 7. Revised list of known IOCTLs. 8. Revised list of known GUIDs. Release 1.00.007 1. Added support for Microsoft Windows Vista (?). 2. Redesign right panel view. 3. Fix bug with processing of Image hooks. 4. Several minor bugs were fixed. Release 1.00.006 1. A BSOD (blue screen) was fixed, which appeared o

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值