The Penn State researchers were recently awarded a grant to teach computers to generate original design ideas and to determine their technical feasibility. This initiative has been envisaged in the interest of national security, to give the United States a competitive advantage in technology-related sectors.
The 18-month project titled, “Generative Adversarial Networks for Design Exploration and Refinement,” or GANDER has the support of the Defense Advanced Research Projects Agency (DARPA). The project will try to determine whether a computer can be trained to do multiple things. Initially, a computer would generate novel engineering design ideas and then those generated ideas would be evaluated for feasibility in the real world.
This approach to machine learning is novel in that the researchers are moving away from merely instructing a computer on how to classify the difference between items in a given environment, say the difference between a car or a bicycle. This is more of an attempt to teach computers creativity — having them generate new concepts from scratch, which is of utmost importance in aspects of engineering design.
Researchers are exploring a domain of deep learning that is referred to as generative adversarial networks (GAN) which incorporate two competing neural networks: one to generate ideas and the other to dissect and determine whether the idea is of practical use.
Demonstration of the function to train multiple networks simultaneously in a virtual environment can help with the generation of innovative design concepts. With this, the discriminator component of GAN will be aware of the embedded environment in its model. In short, an iterative process till the idea generator network can fool the discriminator network into believing that a generated idea is really viable.
To create the discriminator network, the research team plans to make use of simulation environments (i.e., not unlike those used in virtual reality) to incorporate information about physics and physical attributes of the universe so that the designs generated are grounded in the physical laws that govern the real world.
As these two paradigms come together and turn successful, it would empower a wide range of industries to explore novel designs in an accelerated manner. While anyone can generate new ideas, the challenge lies in filtering the feasible ideas from the infeasible, especially those relevant to the physical laws of nature. Any machine-generated ideas will consecutively be guided onto design feasibility after the idea-generation phase.
There seems to be many applications that DARPA is likely interested in, not just in the area of defense. What’s crucial from an artificial intelligence (AI) point of view is being able to significantly reduce the time an AI system takes to learn about the physical properties of its environment. This will have a tremendous impact on autonomous systems in general.
Training a computer to become an expert designer is something that takes human beings several years to perfect. It also requires embedding physical characteristics of the universe into a system and training it to understand such properties. If successful, the significant strength of this project will be the synergy of a novel idea with it being grounded in the reality of the physical law.
Related Post: What You Need To Know About CPU & GPU Miner Trojans