Network modularization to learn high dimensional robot tasks

A method for solving a high dimensional pattern recognition problem using neural networks of lower input dimensions, said method comprising:

• dividing a multi-dimensional problem into two mutually exclusive sets of grouped dimensions;
• providing two neural networks concatenated in series running on a computer having a processor and memory, where a first neural network in the series encodes only its corresponding set of grouped dimensions and a second neural network in the series encodes both of the sets of grouped dimensions;
• independently training the neural networks using supervised learning, where the first neural network is trained using a maximal margin technique to find target values of its corresponding set of grouped dimensions having an optimal value of an output variable for any one combination of target values of the other set of grouped dimensions;
• running the neural networks in inference mode, where an original input defining an environment of the problem is provided, and each of the neural networks searches only its corresponding set of grouped dimensions to find the target values, and the target values from the first neural network are used as inputs by the second neural network;
• and
• combining outputs of the neural networks to provide a final output including the target values of all dimensions in both of the sets of grouped dimensions.

A method for solving a high dimensional pattern recognition problem using neural networks of lower input dimensions, said method comprising:

• dividing a multi-dimensional problem into first, second and third mutually exclusive sets of grouped dimensions;
• providing three neural networks concatenated in series running on a computer having a processor and memory, where a first neural network in the series encodes only the first set of grouped dimensions, a second neural network in the series encodes the first and second sets of grouped dimensions, and a third neural network in the series encodes all three sets of grouped dimensions;
• independently training the neural networks using supervised learning, where the first neural network is trained using a maximal margin technique to find target values of the first set of grouped dimensions having an optimal value of an output variable for any one combination of target values of the other sets of grouped dimensions, and the second neural network is trained using a maximal margin technique to find target values of the second set of grouped dimensions having the optimal value of the output variable for any one combination of target values of the third set of grouped dimensions;
• running the neural networks in inference mode, where an original input defining an environment of the problem is provided, and each of the neural networks searches only its corresponding set of grouped dimensions to find the target values, and the target values from the first neural network are used as inputs by the second neural network, and the target values from the first and second neural networks are used as inputs by the third neural network;
• and
• combining outputs of the neural networks to provide a final output including the target values of all dimensions in all three of the sets of grouped dimensions.