Enhanced deep learning capabilities simplify model design, training, and deployment.

MathWorks today announced Release 2017b (R2017b). R2017b includes new features in MATLAB and Simulink, six new products, and updates and bug fixes to 86 other products. R2017b also adds key new deep learning capabilities that simplify how engineers, researchers, and other domain experts design, train, and deploy models.

The detailed deep learning features and products included in R2017b with deep learning support are as follows.

▲Neural Network Toolbox adds support for complex architectures, including directed acyclic graphs (DAGs) and long short-term memory (LSTM) networks, and provides access to well-known pre-trained models such as GoogLeNet.

▲The Image Labeler app in the Computer Vision System Toolbox provides a convenient, interactive way to label ground truth data from a sequence of images. Beyond object detection workflows, the Computer Vision System Toolbox also supports semantic segmentation, which uses deep learning to classify pixel regions in an image and evaluate and visualize the segmentation results.

▲The new product, GPU Coder, automatically converts deep learning models into CUDA code for NVIDIA GPUs. According to internal benchmarks, the generated code for deep learning inference outperforms Caffe2 by up to 4.5x and TensorFlow by up to 7x on deployed models.

In addition to features introduced in R2017a, pretrained models can be used for transfer learning, including convolutional neural network (CNN) models such as AlexNet, VGG-16, and VGG-19, as well as Caffe models from the Caffe Model Zoo. Furthermore, various models, such as image classification, object detection, and regression, can be developed from scratch using CNNs.

“R2017b enables engineers and system integrators to use MATLAB’s deep learning capabilities to gain greater control over their design processes and accelerate time-to-market,” said David Rich, MATLAB marketing director at MathWorks. “It also enables automation of ground truth data labeling in MATLAB.”