MSU Technologies

060116: Multilayer In-place Learning Networks

Description:

Creating a neural network that can be fully and automatically trained for classification or regression analysis has been a great challenge. The well-known methods such as feed-forward networks with back-propagation learning, radial-basis functions, support vector machines, cascade-correlation learning architecture, and independent component analysis do not consider optimal statistical efficiency, and therefore suffer from a variety of problems such as local minima and unnecessarily large memory requirements.

Description

This technology is a network design for multi-layer neural-network representations suited for classification and regression analysis. It introduces a new, recurrent network architecture that includes bottom-up, lateral, top-down, and out-of-network projections, a new near-optimal in-place learning method, and the integration of unsupervised learning and supervised learning through every layer of the network.

In-place learning is a biological concept where each neuron is fully responsible for its own learning in its environment and there is no need for an external learning network. This results in a simple overall network architecture. Computationally, in-place learning provides unusually efficient learning algorithms whose simplicity, low computational complexity, and generality are set apart from typical conventional learning algorithms.

This technology provides fully automatic internal self-organization that enables it to autonomously learn skills or tasks through autonomous interactions with its environment. After being trained on a set of samples, the network balances between two conflicting criteria afforded by the training set: global within-class invariance and global between-class discrimination.

Benefits

Broad applicability: Applicable to wide range of applications.
Optimal statistical efficiency: Unlike established methods that suffer from architectural-specific problems, the subject invention considers statistical efficiency and almost completely eliminates the local minima problem, which is common in high dimensional networks for classification or regression analysis.
Low computational complexity: This is expected to lead to improved performance and reduced memory requirements.
Ease of use: This network is easy to use, with few user-selected parameters.

Applications

This technology can serve as a core engine for a wide variety of applications such as face, object, character, or biometric data recognition, image analysis, stock value prediction, financial data analysis such as for automated trading systems, and intelligent robots. It can be implemented in software, hardware, or a combination thereof.