Neural Networks Mathematica Version 11 introduces a high-performance neural network framework with both CPU and GPU training support. A full complement of vision-oriented layers is included, as well as encoders and decoders to make trained networks interoperate seamlessly with the rest of the language. Constructing and training networks often requires only a few lines of code, putting deep learning in the hands of even nonexpert users. Key Features Use a wide range of image-oriented layer types to implement cutting-edge computer-vision algorithms. » Define network topologies with multiple inputs, outputs, and arbitrary directed acyclic graph connectivity structure. » Work with image, categorical, and numeric inputs and outputs. » Define networks with multiple loss functions to perform multitask learning. » Easily evaluate trained networks using a variety of built-in classifier metrics. » Train on out-of-core image datasets. » Train networks on either CPUs or NVIDIA GPUs.» Take advantage of the NVIDIA CUDA® Deep Neural Network library (cuDNN) for optimal GPU performance. » Import and export trained networks as "WLNet" files. » Employ automatic tensor shape inference to write succinct network definitions. » Related Examples Digit Classification » Unsupervised Learning with Autoencoders » Object Classification» Learn a Parameterization of a Manifold » Learn to Classify Points from Different Clusters » Out-of-Core Image Classification » Multi-task Learning » Use a Validation Set to Minimize Overfitting » Measure Classification Performance » Accelerate Training Using a GPU » Generate Random Images » Model an Image as a Function »
