认知深化模型驱动的感知进化网络：应对新感官输入

需积分: 5 96 浏览量更新于2024-08-26 收藏 2.02MB PDF 举报

"这篇研究论文探讨了一种名为感知进化网络（Perception Evolution Network，PEN）的模型，该模型基于认知加深理论，旨在适应新感觉受体的出现。PEN是一种生物启发式的无监督学习和在线增量学习神经网络模型，能够以增量方式自动从学习数据中学习合适的原型，并且不需要预先定义的原型数量或相似性阈值。此外，PEN比现有的无监督神经网络模型更先进，允许网络的感知域中出现新的感知维度。当引入新的感知维度时，PEN能够将新的多维感官输入与已学习的原型整合，即将原型映射到一个高维空间，这个空间包含原始维度和新引入的感官输入维度。实验使用人工数据和真实世界数据验证了该模型的性能和有效性。" 这篇论文主要关注的是在神经网络模型中如何处理和适应新的感官输入，这是生物系统中一个重要的适应性特征。感知进化网络（PEN）是作者提出的一种创新模型，它模拟了生物系统如何随环境变化而进化其感知能力。PEN的核心特点是它的无监督学习能力和在线增量学习能力，这使得它能够在数据流不断变化的情况下自我调整和优化。无监督学习是机器学习的一个分支，通常用于发现数据中的隐藏结构或模式，而不依赖于预先标记的类别信息。PEN在此基础上进一步发展，不仅能够自动学习数据集中的代表原型，而且无需预设原型数量，这意味着模型可以根据数据本身的特性自我调整。在线增量学习是指模型能够处理新数据流，并实时更新其参数和结构的过程。在PEN中，这种特性使得模型能够应对新感觉受体的出现，即新的感官输入维度。当网络遇到新的感知维度时，它会将这些新信息集成到现有原型中，通过映射到更高维度的空间来适应变化，从而保持其对环境的适应性。实验部分，PEN在人工数据和现实世界数据上进行了测试，以验证其在处理新感知维度和持续学习的能力。这些实验结果对于理解PEN模型的性能和实际应用潜力提供了关键证据。这篇论文提出的PEN模型为神经网络在适应环境变化，特别是处理新感官输入方面的研究提供了一个新的视角，对于未来在机器人、人工智能和生物启发式计算等领域有潜在的应用价值。

XING et al.: PEN BASED ON COGNITION DEEPENING MODEL 609

Fig. 4. Abstract structure of the perception-prototype ﬁeld limited neural

networks. The number of prototypes in the prototype ﬁeld is ﬁxed. Lines in

the prototype ﬁeld represent the topology relationship between prototypes.

In some algorithms, topology relationship is not predeﬁned but learned by

the algorithm itself such as the TRN.

Fig. 5. Abstract structure of the prototype ﬁeld open-ended neural networks.

The prototypes with dashed edge in the prototype ﬁeld are added or pruned

during learning. Some algorithms do not deﬁne the topology relationship

between the prototypes such as the ART network.

To solve this problem, many growing neural networks

or incremental networks are designed. Growing

SOM (GSOM) [19], growing cell structure [20], and

growing NG (GNG) [21] insert new prototype(s) for every

λ samples learned, where λ is a constant parameter. Life-long

learning cell structure [22] introduces an insertion criterion to

decide whether to insert a new prototype for every λ patterns

learned; meanwhile, it also deletes prototypes to avoid

overﬁtting. However, in these methods, during each λ period,

the input sample is forced to merge with a prototype no

matter how big the gap between them. Considering the

physical meaning, it is unreasonable to merge two patterns

with signiﬁcant difference.

There is another problem, which is well known as the

stability-plasticity dilemma [9], i.e., making the system

quickly learn about new knowledge (e.g., new objects) without

just as quickly being forced to forget previously learned, but

still useful, memories. The adaptive resonance theory (ART)

network [9], fuzzy ART [23], evolving SOMs [24], and

TopoART [25] (ART family) will create a new prototype

when no match occurs between the current input sample

and the current category set. The degree of matching is

controlled by a parameter known as the vigilance parameter.

This strategy makes the network add new prototype, when

the input sample is not similar to the existing prototypes that

the network has learned. However, the vigilance parameter

should be predeﬁned. It is a difﬁcult job when we have little

prior knowledge about the learning task, especially for the

unsupervised learning task. Evolving vector quantization [26]

introduces an online split-and-merge strategy to overcome the

poor setting of the vigilance parameter.

Self-organizing incremental neural network (SOINN) [27],

enhanced-SOINN [28], adjusted-SOINN [29], and

load-balancing-SOINN [30] decide whether to create a

new prototype for the input sample according to the

prototype-distribution around the local region of the input

sample. These methods overcome the disadvantages of

GSOM, GNG, and ART family algorithm. Incremental

learning vector quantization [31], [32] introduces the idea of

the adjusted-SOINN to the learning vector quantization and

achieves very good results.

We summarize these unsupervised incremental learning

methods as prototype ﬁeld open-ended models, as shown

in Fig. 5. This type of network focuses on the prototype ﬁeld;

it makes the prototype ﬁeld open-ended for new categories

by adding new prototypes. In recent years, these methods are

applied to various domains, including reasoning system [33],

pattern recognition [34], and computer vision [35].

However, the perception-prototype ﬁeld limited and

prototype ﬁeld open-ended models are not able to expand

the perception ﬁeld, which we think is a very important

ability for unsupervised learning as mentioned in Section I.

For example, if we install new sensors to a robot as a new

information channel, and we want the robot to use the new

sensors effectively. The perception-prototype ﬁeld limited and

prototype ﬁeld open-ended models cannot deal with such

a task, and the proposed PEN is designed to solve such

problems.

III. P

ERCEPTION EVOLUTION NETWORK

A. Problem Formulation

Assume that the original neural network N has n neurons

in the perception ﬁeld, which receive n-dimensional external

data x = (x

, x

,...,x

) ∈ R

. After a period of learn-

ing, some prototypes are created in the prototype ﬁeld of

PEN. Then, m new sensory neurons emerge in the percep-

tion ﬁeld of PEN. In addition, the received data becomes

x = (x

, x

,...,x

, x

n+1

,...,x

n+m

) ∈ R

n+m

. The learned

prototypes will be mapped to a high-dimensional space, which

contains the dimension of these m new sensory neurons.

If the new sense brings some new distinguishable categories,

PEN will create prototypes for such new categories.

The entire workﬂow of the PEN is as follows. The prototype

ﬁeld of PEN is empty in the beginning, and learning samples

are fed into the network sequentially, i.e., in an online way.

The PEN will create two prototypes using the ﬁrst two input

samples. For the latter input sample, PEN ﬁrst conducts

prototype competition, then prototype learning and prototype

self-adaptive associating are conducted according to the result

of the competition step; meanwhile, the similarity threshold of

the activated prototypes will be updated. When all the steps

are done, PEN will process the next input sample. Prototype

pruning is conducted after every λ samples learned.

When some new sensory neurons are introduced,

PEN will ﬁnd some low-dimensional prototype to

map to high-dimensional space for each input sample.

Prototype self-adaptive associating and similarity threshold

updating are conducted similarly as the procedure before

剩余13页未读，继续阅读

weixin_38562085

粉丝: 6
资源: 964

认知深化模型驱动的感知进化网络：应对新感官输入

基于海马体位置细胞的空间定位神经网络模型.pdf

基于遗传神经网络的颜色恒常感知计算模型.pdf

基于老年人认知需求模型的资讯APP界面设计研究.pdf

基于视觉感知的复杂3D树模型的实时信息重组

基于虚拟现实触觉感知接口技术的研究与进展

基于认知心理学的电子商务网站设计研究.pptx

顾客对重庆精品民宿的体验感知研究——基于网络评价.doc

课堂交互产生学习结果的认知模型

基于35要素双压电晶片可变形镜的自适应光学视觉仿真和感知学习系统

波前传播增强的皮层网络环境认知模型

最新资源