The sample complexity of pattern classification with neural networks: the size of the weights is more important than the size of the network

Sample complexity results from computational learning theory, when applied to neural network learning for pattern classification problems, suggest that for good generalization performance the number of training examples should grow at least linearly with the number of adjustable parameters in the network. Results in this paper show that if a large neural network is used for a pattern classification problem and the learning algorithm finds a network with small weights that has small squared error on the training patterns, then the generalization performance depends on the size of the weights rather than the number of weights. For example, consider a two-layer feedforward network of sigmoid units, in which the sum of the magnitudes of the weights associated with each unit is bounded by A and the input dimension is n. We show that the misclassification probability is no more than a certain error estimate (that is related to squared error on the training set) plus A3 √((log n)/m) (ignoring log A and log m factors), where m is the number of training patterns. This may explain the generalization performance of neural networks, particularly when the number of training examples is considerably smaller than the number of weights. It also supports heuristics (such as weight decay and early stopping) that attempt to keep the weights small during training. The proof techniques appear to be useful for the analysis of other pattern classifiers: when the input domain is a totally bounded metric space, we use the same approach to give upper bounds on misclassification probability for classifiers with decision boundaries that are far from the training examples.

Face recognition using fractal codes

In this paper we propose a new method for face recognition using fractal codes. Fractal codes represent local contractive, affine transformations which when iteratively applied to range-domain pairs in an arbitrary initial image result in a fixed point close to a given image. The transformation parameters such as brightness offset, contrast factor, orientation and the address of the corresponding domain for each range are used directly as features in our method. Features of an unknown face image are compared with those pre-computed for images in a database. There is no need to iterate, use fractal neighbor distances or fractal dimensions for comparison in the proposed method. This method is robust to scale change, frame size change and rotations as well as to some noise, facial expressions and blur distortion in the image

Demonstration of closed-­loop airborne sense-­and-­avoid using machine vision

This paper describes a vision-based airborne collision avoidance system developed by the Australian Research Centre for Aerospace Automation (ARCAA) under its Dynamic Sense-and-Act (DSA) program. We outline the system architecture and the flight testing undertaken to validate the system performance under realistic collision course scenarios. The proposed system could be implemented in either manned or unmanned aircraft, and represents a step forward in the development of a “sense-and-avoid” capability equivalent to human “see-and-avoid”.

Face recognition across pose on video using eigen light-fields

We propose an approach to employ eigen light-fields for face recognition across pose on video. Faces of a subject are collected from video frames and combined based on the pose to obtain a set of probe light-fields. These probe data are then projected to the principal subspace of the eigen light-fields within which the classification takes place. We modify the original light-field projection and found that it is more robust in the proposed system. Evaluation on VidTIMIT dataset has demonstrated that the eigen light-fields method is able to take advantage of multiple observations contained in the video.

Patterns and transitions of query reformulation during web searching

Purpose – To investigate and identify the patterns of interaction between searchers and search engine during web searching. Design/methodology/approach – The authors examined 2,465,145 interactions from 534,507 users of Dogpile.com submitted on May 6, 2005, and compared query reformulation patterns. They investigated the type of query modifications and query modification transitions within sessions. Findings – The paper identifies three strong query reformulation transition patterns: between specialization and generalization; between video and audio, and between content change and system assistance. In addition, the findings show that web and images content were the most popular media collections. Originality/value – This research sheds light on the more complex aspects of web searching involving query modifications.

CAT-SLAM : probabilistic localisation and mapping using a continuous appearance-based trajectory

This paper describes a new system, dubbed Continuous Appearance-based Trajectory Simultaneous Localisation and Mapping (CAT-SLAM), which augments sequential appearance-based place recognition with local metric pose filtering to improve the frequency and reliability of appearance-based loop closure. As in other approaches to appearance-based mapping, loop closure is performed without calculating global feature geometry or performing 3D map construction. Loop-closure filtering uses a probabilistic distribution of possible loop closures along the robot’s previous trajectory, which is represented by a linked list of previously visited locations linked by odometric information. Sequential appearance-based place recognition and local metric pose filtering are evaluated simultaneously using a Rao–Blackwellised particle filter, which weights particles based on appearance matching over sequential frames and the similarity of robot motion along the trajectory. The particle filter explicitly models both the likelihood of revisiting previous locations and exploring new locations. A modified resampling scheme counters particle deprivation and allows loop-closure updates to be performed in constant time for a given environment. We compare the performance of CAT-SLAM with FAB-MAP (a state-of-the-art appearance-only SLAM algorithm) using multiple real-world datasets, demonstrating an increase in the number of correct loop closures detected by CAT-SLAM.