Neural Network Patents
Neural networks have revolutionized many data processing operations. And even though the neural nets learn from training data sets, they can still be protected with patents. This is surprising to many inventors, who consider their solutions to be implementations of basic ideas. However, my regular review of recent neural network patents shows that many innovations using basic tools and ideas are patentable subject matter. For example, innovations with unique sources of inputs for training and operation have been recently patented.
We look for strategic choke points that are data agnostic but that are key to discerning trends and processing data efficiently. Even common elements such as data types, data orientation, data preprocessing, neuron types, convolution methods, cost functions, overfitting reduction, softmax functions, and neuron layer organizations can be protected if implemented in a novel way.