Examples

The following examples show the SignalPop AI Designer in action.

Detecting Gas Leaks from UAV Video

Using Single-Shot Multi-Box (SSD) object detection, described by [1], we were able to detect gas leaks in a video shot from an unmanned aircraft vehicle (UAV).

Original “IR Optical Gas Imaging videos provided by Sierra Olympic Technologies”.

Creating New and Unique Art

Using the new Neural Style Transfer feature, the SignalPop AI Designer easily creates new art by applying the style learned from a piece of art to the content provided by any photograph.  For example, below Vincent Van Gogh’s Starry Night was used to paint the picture of a train track.

Neural Style Transfer

In another example, we painted the same train track, but this time used Edvard Munch’s The Scream.

Edvard Munch’s The Scream

And finally, we had to let Claude Monet get into the action, so we painted the same train track with Claude Monet’s Water Lilies 1916.

Claude Monet’s Water Lilies 1916

All images used by permission under the CC0 license or as public domain.

For more cool examples of Neural Style Transfer using the SignalPop AI Designer, see the Zorodervoncodier Instagram.

Creating Shakespeare Sonnets with Recurrent LEARNING

Using an LSTM based Recurrent Learning model, we were able to quickly train the model to generate a new Shakespeare sonnet.

Beating ATARI Pong with Policy Gradients

Using a simple Policy Gradient Reinforcement Learning model, we were able to quickly train the model on an ATARI gym (that uses the AleControl) and beat the ATARI Pong game using the SignalPop AI Designer. The new MyCaffeTrainerRL seamlessly trains the model to learn the game dynamics using MyCaffe.

Balancing Cart-Pole using Policy Gradients

Using a simple Policy Gradient Reinforcement Learning model, we were able to quickly train the model on a Cart-Pole gym to balance a pole using the SignalPop AI Designer.  The new MyCaffeTrainerRL seamlessly trains the model to learn the balancing dynamics using MyCaffe.

Deep Draw – CIFAR-10

We had so much fun making the first DeepDraw video that we had to create another.  With this one, we again used the GoogleNet model and imported it to boot-strap our training of the CIFAR-10 dataset up to 88% in about 30 minutes with the SignalPop AI Designer.  And then with the SignalPop AI Designer’s Dream Evaluator, we generated the images from the 10 CIFAR-10 classes using the MyCaffe implementation of the Deep Draw algorithm (inspired by the Deep Dream algorithm).  We then synchronized the deep draw images with great electronic dance music by CLOV15.

Deep Draw – IMAGENET

Using the SignalPop AI Designer and MyCaffe‘s implementation of the Deep Draw algorithm (inspired by the Deep Dream algorithm) with the GoogleNet model, we synchronized the deep draw images with electronic dance music by CLOV15.

T-SNE

Using the SignalPop AI Designer’s implementation of the T-SNE algorithm, we animated the results of the algorithm as it converged on several different datasets including the MNIST dataset of handwritten digits. We then synchronized the animation to the electronic dance beat provided by CLOV15.

UnPooling

Using the MyCaffe implementation of Unpooling employed by the SignalPop AI Designer for debugging deep convolutional networks, we synchronized various visualizations of the CIFAR-10 dataset, with electronic dance music by CLOV15.


[1] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg, SSD: Single Shot MultiBox Detector, arXiv:1512.02325, 2016.