Circuits and method for shaping the influence field of neurons and neural networks resulting therefrom

• US 6,347,309 B1
• Filed: 12/30/1998
• Issued: 02/12/2002
• Est. Priority Date: 12/30/1997
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Tannhof, Pascal, De Tremiolles, Ghislain Imbert
• Primary Examiner(s): Teska, Kevin J.
• Assistant Examiner(s): STARKS, WILBERT L
• Application Number:
  US09/223,478
• Time in Patent Office: 1,140 Days
• Field of Search: 706/15, 706/18, 706/20, 706/22, 706/26, 706/27, 706/30, 706/31, 706/33, 706/34, 706/38, 706/39, 382/157
• US Class Current: 706/20
• CPC Class Codes: G06F 18/24133 Distances to prototypes
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

The improved neural network of the present invention results from the combination of a dedicated logic block with a conventional neural network based upon a mapping of the input space usually employed to classify an input data by computing the distance between said input data and prototypes memorized therein. The improved neural network is able to classify an input data, for instance, represented by a vector A even when some of its components are noisy or unknown during either the learning or the recognition phase. To that end, influence fields of various and different shapes are created for each neuron of the conventional neural network. The logic block transforms at least some of the n components (A1, . . . , An) of the input vector A into the m components (V1, . . . , Vm) of a network input vector V according to a linear or non-linear transform function F. In turn, vector V is applied as the input data to said conventional neural network. The transform function F is such that certain components of vector V are not modified, e.g. Vk=Aj, while other components are transformed as mentioned above, e.g. Vi=Fi(A1, . . . , An). In addition, one (or more) component of vector V can be used to compensate an offset that is present in the distance evaluation of vector V. Because, the logic block is placed in front of the said conventional neural network any modification thereof is avoided.

Claims

1. An improved neural network for classifying and identifying an input vector A having n components, comprising:

a) encoding means responsive to said input vector A for encoding the n components of vector A into m components of an output vector V, such that at least one of said m components of the output vector V is a linear or non linear function of some of the n components of vector A and at least another one of said m components of the output vector V is identical to a corresponding component of the input vector A, for shaping the influence field of a first neuron differently than the influence field of a second neuron; and

b) a neural network based upon a mapping of the input space having at least one input terminal for receiving the components of said vector V, said neural network including p neurons to store p prototypes, said neural network comprising;

p memorization means, each being adapted to store m weights for each one of the p prototypes that have been previously loaded during a learning phase;

q computing means, each being adapted to calculate a distance between vector V and a corresponding one of said prototypes; and

,

decision means that processes the p distances to elaborate the global response of the improved neural network based on the p prototype distances.

2. A method for classifying and identifying an input vector A having n components in an improved neural network having means for receiving said input vector and including a neural network based upon a mapping of the input space, said method comprising the steps of:

a) encoding the n components of vector A into m components of an output vector V, such that at least one of said m components of the output vector V is a linear or non linear function of some of the n components of vector A and at least another one of said m components of the output vector V is identical to a corresponding component of the input vector A, for shaping the influence field of a first neuron differently than the influence field of a second neuron;

b) applying the components of said output vector V to said neural network based upon a mapping of the input space having at least one input terminal for receiving the m components of said vector V and including p neurons to store m weights for each one of the p prototypes that have been previously loaded therein during a learning phase;

each said neuron being adapted to compute a distance between output vector V and the prototype memorized therein; and

,

c) elaborating a global response of said improved neural network based on the p prototype distances.