结构增强的增量神经学习：图像分类中的子图提取

45 浏览量更新于2024-08-28 收藏 680KB PDF 举报

"这篇研究论文‘STRUCTURALLY ENHANCED INCREMENTAL NEURAL LEARNING FOR IMAGE CLASSIFICATION WITH SUBGRAPH EXTRACTION’主要探讨了一种结构增强的增量式神经学习技术，用于图像分类，通过子图提取优化了判别码本表示。文章发表在2014年《国际神经系统杂志》第24卷第7期，由南京大学新软件技术国家重点实验室的研究人员与曼彻斯特大学电气与电子工程学院的学者共同撰写。" 在该论文中，作者提出了一种创新的结构增强的增量式神经学习方法，旨在构建具有区分性的图像码本表示，从而提高图像分类的效率。传统的神经网络学习方法可能忽视了视觉词汇之间的结构和分布关系，而该方法则尝试将这些关系纳入到码本学习过程中。为了实现这一目标，他们开发了一种基于新型结构增强增量学习技术的在线码本图学习方法，称为“可视化诱导自组织增量神经网络（ViSOINN）”。ViSOINN网络通过自我组织过程，不仅能够学习和更新码本，还能捕获图像数据中的结构信息，使得学习过程更加动态且适应性强。隐藏层节点在ViSOINN中扮演了关键角色，它们负责编码和解码图像特征，同时通过子图提取来揭示图像的局部结构信息。这种子图提取策略有助于识别图像中的关键特征，提高分类精度。此外，通过增量式学习，网络能够逐步适应新的类别，无需重新训练整个模型，这在处理大规模、不断增长的数据集时尤其重要。论文进一步阐述了ViSOINN的实现细节，包括学习算法的步骤、参数调整以及性能评估。实验结果证明，该方法在多个图像分类基准数据集上的表现优于其他现有的增量学习和非增量学习方法，展示了其在处理不断变化的数据集和复杂图像分类任务方面的潜力。这篇研究论文为图像分类领域的神经网络学习提供了一个新的视角，通过结构增强和增量学习，有效地提高了图像识别的准确性和模型的适应性。这对于未来智能系统的发展，特别是在视觉识别和机器学习应用中，有着重要的理论和实践意义。

September 26, 2014 15:48 1450024

Structurally Enhance d Incr emental Neural Le arning

the Self-Organizing Map (SOM),

has been used

to deal with this problem as its topology preserv-

ing property is favorable for VQ to provide the map-

ping with enhanced tolerance to faults and noises.

SOM is trained in an unsupervised learning manner

to produce a low-dimensional space and it can pro-

vide topologically preserved mapping from input to

output space using a neighborhood-based function

between neurons.

17–20

However, the distributions of

data points are often distorted on the map.

For better scaling and visualization, a direct and

faithful display of data structure and data distribu-

tion is highly desirable. An extended SOM called

“Visualization-induced SOM” (ViSOM),

has been

proposed to constrain the lateral contraction force

between the neurons in SOM to regularize the inter-

neuron distances according to a scalable parameter,

used to determine and control the resolution of the

map. ViSOM can preserve the data structure and the

topology as faithfully as possible. However, neither

SOM nor ViSOM takes the classiﬁcation border into

account. To design more eﬃcient and robust learn-

ing algorithms for classiﬁcation, various incremental

learning algorithms have been proposed and applied

on classiﬁcation and clustering tasks.

The well-known examples of incremental or com-

petitive neural networks include the “Growing Cell

Structures” (GCS)

and the “Growing Neural Gas”

(GNG).

It is known that the requirement of a pre-

determined neuron number and a pre-ﬁxed struc-

ture makes SOM unpractical. GCS was developed to

allow projections onto a nonlinear, discretely sam-

pled subspace whose dimensionality has to be cho-

sen as apriori. However, some information on the

topological characteristics of the input data may be

lost in this process by simply considering the rela-

tionship between the inherent data and the target-

ing space. GNG is a further modiﬁcation to GCS,

in which the dimensionality of topological structure

is not pre-deﬁned but discovered during the training

process. Hamker

proposed an extension to GNG

in order to perform some supervised incremental

learning tasks, but it is not suitable for unsuper-

vised learning. In recent years, further variations of

GNG have been proposed to perform unsupervised

clustering tasks, such as SOINN.

SOINN works as

an incremental neural learning technique to pro-

cess online dynamic data. It represents a topological

structure of the input distribution. Empirical results

show that SOINN is capable of learning the necessary

number of nodes, using fewer nodes but obtain-

ing better results than GNG. Furthermore, sev-

eral extensions of SOINN have been developed and

achieved improved performance on many databases,

such as adjust SOINN

and ESOINN.

They have

been used in incremental acquisition of language

for humanoid robot,

associative memory,

ﬁnding

typical vector for k-nearest neighbor classiﬁcation,

and pattern-based incremental reasoning.

However, there are still four problems remaining

unsolved with SOINN and its extensions: (1) SOINN

is not stable. The learning results rely heavily on the

order of the input data, that is, the number and the

position of diﬀerent nodes will be diﬀerent even if

one repeats the training under the same environment

by only changing the order of the input; (2) SOINN

uses a two-layer network. During online learning, the

user must set clearly the time points to stop the ﬁrst-

layer’s learning and to begin the second-layer’s learn-

ing; (3) SOINN only preserves topological structures,

not the inter-neuron distances, of the input data on

the output space, which fails to preserve the metric

on the mapped space; (4) SOINN only updates the

nearest node incrementally. It ignores the informa-

tion of other representative nodes such as the sec-

ond nearest node and the farthest node, losing some

important information in the learning process.

Moreover, most of the above methods work

oﬄine. Nowadays, online learning is proven to be

able to ﬁt large, or slowly varying datasets better

as it learns one instance at a time, and thus has

been widely used in image classiﬁcation tasks due to

its high learning eﬃciency. Examples are as follows.

Mairal et al.

proposed an iterative online code-

book learning algorithm to minimize the expected

cost instead of the empirical cost for inﬁnite size

of training set. Based on the ﬁrst-order stochastic

gradient descent, the algorithm scaled up gracefully

to large-scale datasets and was suitable for a wide

range of learning problems. To capture salient prop-

erties of images in real time, Zhang et al.

proposed

an online sparse learning algorithm that utilized the

reconstruction error to update the current codebook.

The method had greatly sped up the computation,

saving storage space and memory.

In the most recent research,

the codebook has

been represented as a graph, in which the nodes are

visual words and the edges describe the relationships

1450024-3

Int. J. Neur. Syst. 2014.24. Downloaded from www.worldscientific.com

by NANJING UNIVERSITY on 12/08/14. For personal use only.

剩余12页未读，继续阅读

weixin_38687505

粉丝: 10
资源: 969

结构增强的增量神经学习：图像分类中的子图提取

最新资源