Unsupervised visual learning of three-dimensional objects using a modular network architecture

Ando, Hiroshi; Suzuki, Satoshi; Fujita, Toshifumi

doi:10.1016/s0893-6080(99)00052-0

Cited by 14 publications

(5 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…[9] showed that mixtures of autoencoders are able to specialize in different handwritten digit classes. [12] used the reconstruction error of modular autoencoders in the weight updates to force the output of at least one module to fit the input, and to force the rest of the modules to increase the error between the input and the output. [13] used a neural gas mixture of local principal component analyses.…”

Section: Related Workmentioning

confidence: 99%

Modular deep belief networks that do not forget

Pape

Gomez

Ring

et al. 2011

The 2011 International Joint Conference on Neural Networks

View full text Add to dashboard Cite

Abstract-Deep belief networks (DBNs) are popular for learning compact representations of highdimensional data. However, most approaches so far rely on having a single, complete training set. If the distribution of relevant features changes during subsequent training stages, the features learned in earlier stages are gradually forgotten. Often it is desirable for learning algorithms to retain what they have previously learned, even if the input distribution temporarily changes. This paper introduces the M-DBN, an unsupervised modular DBN that addresses the forgetting problem. M-DBNs are composed of a number of modules that are trained only on samples they best reconstruct. While modularization by itself does not prevent forgetting, the M-DBN additionally uses a learning method that adjusts each module's learning rate proportionally to the fraction of best reconstructed samples. On the MNIST handwritten digit dataset module specialization largely corresponds to the digits discerned by humans. Furthermore, in several learning tasks with changing MNIST digits, MDBNs retain learned features even after those features are removed from the training data, while monolithic DBNs of comparable size forget feature mappings learned before.

show abstract

Section: Related Workmentioning

confidence: 99%

Modular deep belief networks that do not forget

Pape

Gomez

Ring

et al. 2011

The 2011 International Joint Conference on Neural Networks

View full text Add to dashboard Cite

show abstract

“…This learning process is based on on-line processing without a teacher, and is also additive learning in which new objects are provided from time to time. Attempting the partial realization of these functions on a computer, one of the authors adopted a competitive modular learning scheme [1,2], and proposed a model by which some three-dimensional objects are recognized and classified without a teacher based on their two-dimensional images projected from various viewpoints [3,4]. This method realized unsupervised clustering in which continuously distributed data are regarded as a class.…”

Section: Introductionmentioning

confidence: 99%

“…In this work we attempt to achieve the above goal by enhancing the learning procedure of the past 3D object recognition model [3,4].…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Adaptive clustering using modular learning architecture

Suzuki

2003

Systems & Computers in Japan

View full text Add to dashboard Cite

SUMMARYWe present a clustering method for data that are continuously distributed within each class and where there is no overlap between the classes. With the method the data are clustered by on-line processing, starting at a point where the number of classes is unknown. The proposed method uses ideas derived from statistical mechanics and applies them to a modular learning architecture: learning progresses by regarding the data as particles and regarding each module as a space in which the particles are distributed. The learning method estimates the energy state that matches the particle distribution. As a result of learning, continuously distributed data are clustered in the same module. At the same time, certain modules appear that contain no data. Consequently, the number of classes is estimated by defining each module containing data as a class. This paper describes the proposed method in detail and shows the results of clustering experiments undertaken using two-dimensional artificial data and facial images.

show abstract

“…There have been many studies on technology of recognizing a 3D object from its 2D image. As example, parametric eigenspace method [7]; neural networks [8,9]; self-organizing neural architecture called VIEWNET [2]; and modular neural networks [10]. Section 2 describes a proposed 3D model to recognize 3D objects based on the Hu and Zernike moment invariants.…”

Section: Introductionmentioning

confidence: 99%

Lecturer, Dpt. of Transportation Engineering, Faculty of Engineering, Tanta University, Tanta, Egypt.

Salama

Bahgat

Elmhrez

2003

International Conference on Aerospace Sciences and Aviation Tec

View full text Add to dashboard Cite

In this paper, we present a new model to recognize a three-dimensional (3D) object. which uses six Hu moment invariants and six Zernike moments as a feature vector. It extends our former work on 3D-object recognition based on Hu moment invariants [1]. The classification of the desired 3D object using Hu-moment model [1] is based on selecting the object that corresponding the minimum distance of the six minima obtained from six reference libraries and the desired object. The classification may be correct if we selected the object that corresponding the second minimum distance value instead of the first value. To select the correct one, this paper describes another model to recognize a 3D object based on Hu and Zernike moment invariants as a feature vector. Zernike moment invariants are used to find the pose of the aircraft at first to know from which library we can make the decision. The proposed model differs significantly from many recent 3D recognition models, which emphasize on stereo reconstruction and structured light analysis. Several trajectories for 3, 6, and 9 aircraft are generated, using 3D Studio Max software. To study the performance of the proposed model, the aircraft patterns were chosen to test the proposed model because the views of these patterns are relatively difficult due to the similarity of their images. Finally, we show that the proposed model achieves high recognition rates compared with that of the View Information Encoded With Network model (VIEWNET) [2] and the Hu moment invariants [1] using the approximately the same set of objects and using the same decision rule.

show abstract

Unsupervised visual learning of three-dimensional objects using a modular network architecture

Cited by 14 publications

References 27 publications

Modular deep belief networks that do not forget

Modular deep belief networks that do not forget

Adaptive clustering using modular learning architecture

Lecturer, Dpt. of Transportation Engineering, Faculty of Engineering, Tanta University, Tanta, Egypt.

Contact Info

Product

Resources

About