Google Develops Evolutionary Techniques for Advanced AI Video Understanding Models

2019-10-18

Video understanding is a subfield of artificial intelligence that not only supports systems capable of automatically recognizing semantics, such as network video classifiers and sports activity recognizers, but also serves as the foundation for robot perception and navigation systems. In this field, designing machine learning models that utilize spatiotemporal information from videos is particularly challenging, and developing models that require minimal computation is also difficult. To address this, Google researchers have conducted a series of studies on automatically searching for the best computer vision algorithms, and they have detailed their findings in a blog post today. The team reports that the performance of the architectures determined by their three methods (EvaNet, assembly enet, and TinyVideoNet) has been proven to be 10 to 100 times faster than existing manual systems on multiple public datasets, marking a significant breakthrough. In the blog post, the researchers write, "To the best of our knowledge, this is the first work on neural architecture search for video understanding. The video architectures generated by our new evolutionary algorithm significantly outperform the most famous hand-designed CNN architectures on public datasets." The researchers believe that the first model, EvaNet, is a module-level architecture search that discovers the optimal configuration. The evolutionary algorithm iteratively updates candidate AI models, while EvaNet modifies modules within each model to generate entirely new architectures, making them fundamentally different. Assemblenet is a method that combines different sub-models with different input patterns and temporal resolutions. In this approach, an architecture can learn the feature representation relationship between patterns through evolution. Google claims that the Assemblenet architecture, trained for 50 to 150 rounds, has achieved state-of-the-art results on popular video recognition datasets, such as Charades and Moments in Time (MIT). Lastly, TinyVideoNets, which can automatically design networks, provide cutting-edge performance overall. It achieves high accuracy and operates efficiently on processors at speeds ranging from 37ms to 100ms, and on graphics chips at a speed of 10ms. Additionally, it offers excellent cost-effectiveness by considering models and simultaneously changing algorithms during the architecture evolution process to reduce computation and improve overall benefits. In conclusion, the Google researchers proudly state in their paper, "Our research opens up new directions for video understanding and demonstrates the prospects of machine-evolved CNNs."

RECENT AI NEWS

RECENT AI TOOLS