A Method to Enable AI to Effortlessly Parse Videos

2019-10-16

MIT and IBM have developed an algorithm that can accurately identify actions in videos, consuming only a small fraction of the processing power previously required. This breakthrough could significantly improve the economic benefits of applying artificial intelligence to a large number of videos. The method utilizes AI techniques to process static images with a rough concept of time, which could help manage the vast amount of content generated on the internet every moment. Despite the hype surrounding the use of AI by tech companies and their dazzling achievements, the application of AI in video analysis remains limited. YouTube, Facebook, and TikTok use machine learning algorithms to classify and recommend video clips, but they seem to rely mainly on video-related metadata such as descriptions, tags, and upload time and location. Everyone is researching methods to analyze video content, but all these methods invariably require more computational power. Companies are looking to use AI to automatically generate detailed descriptions for videos, enabling users to discover uncommented video clips, and also to sell video content-based advertisements. Facebook and Google also hope to use AI to automatically detect and filter illegal or malicious content. Although this seems like an ongoing cat-and-mouse game, it is actually quite challenging to accomplish without significantly increasing the AI carbon footprint. In recent years, as companies have utilized AI to analyze videos, there have been significant advancements in image recognition technology, largely thanks to deep learning. Deep learning algorithms can detect objects based on the pixels displayed in an image, but they are not very good at interpreting videos. Analyzing a single frame of a video does not reveal what has happened unless it is compared to the previous and subsequent frames, which involves the issue of coherence. For example, if a person hurriedly leaves a place, it could be because they have an urgent matter to attend to, or it could be due to an earthquake, or perhaps the person is a thief who quickly leaves after stealing something. There is always a cause and effect, deducing the process based on the cause or inferring the cause based on the result. This is a logical way of thinking. Cause and effect are like raw materials, and the researchers at MIT have cleverly utilized this fundamental point by simplifying the materials, such as converting 3D images into 2D. This undoubtedly saves a lot of data processing power, and by doing so, it significantly reduces the computational workload, making the AI "work" much easier than before. Currently, this method has not yet achieved the ideal results and requires further refinement. Once successful, the benefits it brings will be enormous. In today's world where computational power is increasingly important, the significance of methods that can free up computational power is self-evident.

RECENT AI NEWS

RECENT AI TOOLS