Women in Big Data Global

Donate
×

Blogs

A Night of Deep Learning: Women In Big Data Come Together at AI Connect

Women in Big Data

By Renee Yao,

October 23, 2017

icon_AI

NVIDIA and SAP hosted their first Women In Big Data community event, AI Connect, on October 18, 2017.  More than 100 participants met up at the NVIDIA campus in Santa Clara to learn about the  latest advancements in deep learning

Talk 1:  Deep Learning Use Cases

Renee Yao kicked things off, presenting seven deep learning use cases that highlight how different companies use different deep learning technologies to solve their customers’ problem.

  • Capital One uses deep learning to recreate customer service, drive interactions with a new generation of consumers via chatbots and provide enhanced security.
  • Cape Analytics uses NVIDIA GPUs to train deep learning models on frequently updated geospatial imagery to detect property features at a massive scale, creating a more accurate and up-to-date data source.
  • Athelas uses Keras and TensorFlow to train deep learning models and provide a portable diagnostic device that enables rapid analytics from a drop of blood.
  • Blue River Technology, recently acquired by John Deere, uses Caffe trained on NVIDIA TITAN X and inferenced on NVIDIA Jetson TX2, coupled with computer vision and machine learning, to reduce chemical usage and increase food production.
  • Sadako uses Caffe to train deep neural networks and create the Max AI robot for better identification and recognition of a wide range of objects in complex waste streams.
  • Maketime uses NVIDIA GPUs to train neural network models to streamline procurement and production management, especially in part pricing and supplier matching use cases.
  • Heuritech uses Keras and TensorFlow to train 40K in-house images on 100 NVIDIA GPUs to help luxury brands spot future trends.

Click here to view the complete evening’s video. (Renee’s portion begins at 4:32)
Click here to view the deck for the presentation.


Talk 2: Deep Learning Workflows 

Kari Briski, director of Deep Learning Software Products at NVIDIA, talked about deep learning workflows. She went into detail on the NVIDIA SDK, deep learning training stack, and specific GPU benchmark results, and she laid out a step-by-step guide to deep learning workflows.

TensorRT, which maximizes inference throughput for latency critical services, was one of the highlights of her talk. Customers are thrilled about TensorRT performance:

  • WRNCH body-tracking software uses AI and deep learning to track and process human motion in real time and teach humans to read body language. Paul Kruzewski, CEO of WRNCH, said, “On average, TensorRT has doubled the speed of our inferencing, which is pretty amazing.”
  • Dre Gray at Uber said “Self-driving car’s having real-time execution is obviously very important. Developing state of the art perception algorithms always requires a painful trade-off between speed and accuracy. With our ResNet101 network, TensorRT brought our inference time down from 250ms to 89ms.”
  • Clarifai CEO Matthew Zieler stated that “on average we see around a 10x speedup, with between 3-70x speedups depending on the scenarios.”

The audience in particular enjoyed Kari’s live demo, where she showed how we used transfer learning to train a model and deploy the edge in under a minute. Kari used DIGITS and Volta Tensor Core accelerated training on DGX Station (an AI personal supercomputer) to re-train GoogleNet (with 7 classes) in 15 epochs in 35 seconds and immediately deployed a new model to the edge into the Jetson TX2 developer kit.

Click here to view the complete evening’s video. (Kari’s portion begins at 26:03)
Click here to view the deck for the presentation.


Talk 3: Deep Learning In Enterprise 

Wrapping up the evening, SAP Data Scientist Nazanin Zaker spoke about Deep Learning in the Enterprise. She took us through how deep learning works by breaking down what features, classifiers and CNNs are. She also introduced SAP Leonardo Digital innovation System for machine learning applications.

One of the demos she showed was FaceAPI (a rest API available in the SAP Leonardo Machine Learning foundation), which uses CNNs to detect the position of faces in input images. This API is already available for developers to add to their code and can be called using a post request. An example research paper related to this topic is provided in the slides for more information (see below).

Following that demo, she talked about Text API, which is another functional service available in SAP Leonardo Machine Learning and accessible by a post request. This API can do sentiment analysis on an input text and return the result as positive, negative or neutral. CNN models were applied in this API as well. An example of how CNNs can be used for text processing can be found in the slides.

Nazanin concluded with a few examples, one of which illustrated how Swarovski leverages machine learning and computer vision to repair broken products. Swarovski handles more than 100 service tickets and images a day, and previously spent manual effort going through a catalog with more than 40,000 products. Now, with SAP’s solution, they can compare the broken image with images in the product catalog in seconds.

In another example, SAP Brand Impact uses deep learning to detect logos in an input video, as well as to compare detected logos based on the their frequency and latency in an input video. This application can establish brand awareness within seconds instead of days or weeks.

Click here to view the complete evening’s video. (Nazanin’s portion begins at 56:02)
Click here to view the deck for the presentation.


Overall, feedback on the event was extremely positive. If you’re eager to get started, we shared some resources at the end of each talk. Here are a few:

Click here to view the Meetup announcement.

Thanks to all who made the event possible, and we look forward to seeing everyone at out next Meetup.