Nvidia: Researchers train AI to reward dogs for responding to commands

Colorado State University researchers Jason Stock and Tom Cavey published a paper on an AI system that rewards dogs for doing tricks.

The computer science students trained image classification networks to determine whether a dog is sitting, standing or lying down. When a dog responds to a command by adopting the correct posture, the machine issues a reward.

The students used an Nvidia Jetson edge AI platform for real-time trick recognition and treats. Stock and Cavey see their prototype system as a tool for a dog trainer – it handles the treats – or a way to teach dogs how to behave better at home.

“We have shown the potential for a future product to get out of this,” Stock said in a statement.

Retrieve dog training data

The researchers needed dog images showing the three specified poses. They found the Stanford Dogs datasets, which contained more than 20,000 images of various sizes with dogs in many positions. The images had to be preprocessed, so they wrote a program to quickly label them.

In an email to VentureBeat, Nvidia said, “It doesn’t work remotely yet; it is currently for personal use. But that would be an easy setup to turn it into a remote system. Think of it as a system, or IP, to license devices such as the Furbo. The researchers see many possible applications, but have not yet committed themselves to anything. “

To refine the model, the researchers applied dog characteristics from ImageNet to enable transfer learning. They then applied post-training and optimization techniques to increase speed and reduce model size.

For optimizations, they used Nvidia’s Jetpack software development kit on Jetson, a lightweight AI platform for drones and other systems. It provides an easy way to get things up and running quickly and access the TensorRT and cuDNN libraries, Stock said. Nvidia TensorRT optimization libraries offered “significant improvements in speed,” he added.

Using the university’s computer system, Stock trained the model overnight on two 24GB Nvidia RTX 6000 graphics processing units (GPUs).

Models deployed on Henry

The researchers tested their models on Henry, Cavey’s Australian Shepherd. The model achieved an accuracy of up to 92% in tests and showed it was capable of inference in fractions of a second at nearly 40 frames per second.

Using the Jetson Nano, the system makes real-time decisions about dog behavior and amplifies positive actions with a treat, sending a signal to release a reward.

“We looked at Raspberry Pi and Coral, but neither was enough, and the obvious choice was to use Jetson Nano,” Cavey said.

Explainable AI helps provide transparency about the composition of neural networks. It is increasingly common in financial services as a way of understanding fintech models. Stock and Cavey have included model interpretation in their paper to provide explainable AI for the pet industry.

They do this with images of the videos that show the posture analysis. One set of images is based on GradCAM – a common technique for showing where a convolutional neural network model is focused. Another set of images explains the model by tapping on Integrated Gradients, which helps analyze pixels.

The researchers said it was important to create a reliable and ethical component of the AI ​​system for trainers and general users. Otherwise, there is no way to explain the methodology, should there be any doubt.

“We can explain what our model does, and that can be useful to certain stakeholders – how else can you back up what your model is really learning?” Cavey said.