Deep Learning

Understanding and using deep learning networks

This is machine translation

Translated by Microsoft
Mouseover text to see original. Click the button below to return to the Original version of the page.

What’s new in R2019a: Features 1

Posted by Johanna Pingel,

Welcome again to R2019a! There’s a new release of MATLAB out right now. Last post, we covered a bunch of new 19a examples, and today's post will dive into the specifics of new deep learning features.
Please note: this is primarily for deep learning features, and this list isn’t comprehensive. (Think of this as “Johanna’s list of favorite new features.”). If you think I forgot something important, please leave a comment below!
New Deep Learning Features:
#1. Signal Labeling App
Signal Labeler is the latest app to join our collection of labeling apps, including ground truth labeler, video labeler, image labeler, and audio labeler.
#2. YOLO v2
Hurray! Yolo v2 is here! Train a “you-only-look-once” object detector AND generate CUDA code from the MATLAB code.*

I pulled this image directly from the documentation link above. It's a great page introducing the basics of this algorithm.

#3. Deep learning for videos
Combine LSTM and convolutional layers for video classification and gesture recognition.
#4. Deep Learning on NVIDIA GPU Cloud and DGX
Bring MATLAB with you to the cloud or GPU clusters using NVIDIA’s docker container. A video on how to get this set up can be found here, and documentation is below.
#5. New pretrained models
You can now use MobileNet-v2, Inception-ResNet-v2, Xception, and SqueezeNet, to name a few.

 A new documentation page visualizes pretrained networks' speed, size and accuracy

#6. Deep learning for 3-D image data
Over 20 deep learning layers support 3D data, semantic segmentation workflows work in 3D, and a new example walking through segmenting a brain in 3D is below.
#7. (advanced) Custom layers with multiple inputs and outputs
We’ve been investing in more advanced features for deep learning for the last few releases, so it feels right to add one new advanced feature. While defining your own custom layer is an advanced maneuver, it may be important to implement networks from research papers, since a lot of networks in literature become quite complex.  

Showing intermediate custom layer architecture

  • You can find more information on custom layers here
  • And custom layers with multiple inputs here

I'm always curious how you learn about new features and functionality. Do you watch Gabriel's video? Read the documentation? Let me know in the comments below.
*YOLO v2 CUDA Code generation is available with GPU Coder

1 CommentsOldest to Newest

Gareth Thomas replied on : 1 of 1
Johanna thanks for sharing. The list is indeed very impressive and just to think that this all came in 6 months of work. I wonder what will happen in R2019b:) I really like the idea of having more and more pre-trained models that we can just use.