Noise-Based Neural Network Training for Bias Reduction

Instead of training the neural network to generate specific target outputs directly, the patent inverts the approach by training the network to predict noise (random values). The training process uses noise labels instead of meaningful labels, and the network learns to map inputs to noise outputs. This inversion eliminates the need for expensive manual labeling while still enabling the network to learn effective feature representations that improve downstream task performance.

The patent makes the training process self-service by automatically generating noise labels without requiring human annotators. The system itself creates the training labels through random noise generation, eliminating the need for external human resources. This self-service approach to label generation dramatically reduces both time and cost while maintaining training effectiveness.

The patent applies homogeneity by using uniform noise distribution across all training samples regardless of input content. Instead of selecting diverse labeled images that introduce bias, the system uses identical noise labeling strategy for all inputs, creating a homogeneous training regime that prevents the model from learning biased associations between specific inputs and specific outputs, thereby improving generalization.