CES 2016: NVIDIA DRIVENet Demo - Visualizing a Self-Driving Future (part 5)

TL;DR

Nvidia demonstrates the ability to quickly train neural networks for object detection using large datasets and GPU acceleration.

Transcript

and so just to show it to you let's now see what it can do let me introduce Mike Houston everybody give them a round of applause Mike Mike heads up heads up the development of our neural net and Mike what are you going to show it's thick so when we start with kitty datasets this was developed with Karl's repeat and Toyota and so it's one of the sta... Read More

Key Insights

• 👾 Randomization techniques, such as cropping, warping, and manipulating color spaces, are used to enhance the robustness of neural networks during training.
• 🌥️ The initial training of the network is performed with a large dataset to develop a feature descriptor, which can then be retargeted for different applications.
• 🚂 GPU acceleration significantly reduces training time, making it feasible to train the network on large datasets in a practical timeframe.
• 🏂 Neural networks trained with large datasets and GPU acceleration show remarkable results in object detection, even in challenging conditions such as snow and road spray.
• 👻 The ability to quickly train and retarget neural networks allows for rapid improvements in object detection capabilities.

Questions & Answers

Q: How is NV drive net trained to improve object detection capabilities?

By training with a large dataset of 1.2 million images and manipulating it to look like 120 million objects, NV drive net develops a robust feature descriptor, enabling enhanced detection abilities.

Q: How long does it take to train the original network without GPU acceleration?

Without GPU acceleration, training the original network with the image net dataset would have taken a couple of years, but GPU acceleration reduces the training time to a month.

Q: Can the neural network be retargeted for other applications?

Yes, by using a feature descriptor obtained from one application, the network can be retargeted for similar applications, such as detection and segmentation, with minor modifications.

Q: How quickly can Nvidia experiment and iterate with neural networks?

Nvidia can turn around experiments quickly by using compute time, which is relatively cheap compared to engineering time, allowing for fast iterations and improvements.

Summary & Key Takeaways

• Nvidia's neural network, NV drive net, is trained to detect objects using the Kitty dataset, resulting in stable and accurate detections.

• The training set used for NV drive net consists of 7,400 images and 33,000 unique cars, with randomization techniques used to enhance the network's robustness.

• The network is trained using a large dataset of 1.2 million images, taking advantage of GPU acceleration to reduce training time to a month.