Deep Learning

What’s New in MATLAB for Deep Learning?

MATLAB makes deep learning easy and accessible for everyone, even if you’re not an expert. Check out the latest features for designing and building your own models, network training and visualization, and deployment.

Data Preparation and Labeling

  • Video Labeler: Label ground-truth data in a video or image sequences
  • Audio Labeler: Interactively define and visualize ground-truth labels for audio datasets
  • New Signal Labeler: Visualize and label signals interactively
  • New Pixel label datastore: Store pixel information for 2D and 3D semantic segmentation data
  • Augmented image datastore: Create more training samples to augment deep learning training data
  • New Audio datastore: Manage large collections of audio recordings

Network Architectures

  • Regression and bidirectional LSTMs for continuous, time-series outputs
  • New Train a “you-only-look-once” (YOLO) v2 deep learning object detector and generate CUDA code
  • Deep Network Designer: Graphically design and analyze deep networks and generate MATLAB code
  • New Custom layers support: Define new layers with multiple inputs and outputs, and specify loss functions for classification and regression
  • New Combine LSTM and convolutional layers for video classification and gesture recognition

Deep Learning Interoperability

  • Import and export models with other deep learning frameworks using the ONNX model format
  • Ability to work with MobileNet-v2, ResNet-101, Inception-v3, SqueezeNet, and NASNet
  • New Import TensorFlow-Keras models and generate CUDA code
  • New Import DAG networks in Caffe model importer

See a comprehensive list of pretrained models supported in MATLAB.

Network Training

  • Automatically validate network performance, and stop training when the validation metrics stop improving
  • New Train deep learning networks on 3-D image data  
  • Perform hyperparameter tuning using Bayesian optimization
  • Additional optimizers for training: Adam and RMSProp
  • Train DAG networks in parallel and on multiple GPUs
  • New Train deep learning models on NVIDIA DGX and cloud platforms

Debugging and Visualization

  • DAG activations: Visualize intermediate activations for networks like ResNet-50, ResNet-101, GoogLeNet, and Inception-v3
  • Monitor training progress with plots for accuracy, loss, and validation metrics
  • Network Analyzer: Visualize, analyze, and find problems in network architectures before training

Deployment

  • Integrate generated CUDA code with NVIDIA®TensorRT that takes advantage of FP16 optimization
  • Support for DAG networks including GoogLeNet, ResNet-50, ResNet-101, and SegNet
  • Generate code from trained deep learning models for Intel® Xeon and ARM® Cortex-A® processors
  • Automated deployment to NVIDIA Jetson and DRIVE platforms
  • Deep learning optimization: Improved performance through auto-tuning, layer fusion, and Thrust library support
  • New Apply CUDA optimized transposes using shared memory for improved performance 

Reinforcement Learning

  • Reinforcement Learning Algorithms: Train deep neural network policies using DQN, DDPG, A2C, and other algorithms
  • Environment Modeling: Create MATLAB and Simulink models to represent environments and provide observation and reward signals for training policies
  • Training Acceleration: Parallelize policy training on GPUs and multicore CPUs
  • Reference Examples: Implement controllers using reinforcement learning for automated driving and robotics applications

Get a Free Trial

30 days of exploration at your fingertips.

Have Questions?

Talk to a deep learning expert.