Breakthroughs in Machine Learning

New System Can Recognize Objects from a Verbal Description

Recently, computer scientists at MIT revealed a major breakthrough in machine learning.  They created a system which learned to recognize objects in photographs based upon provided audio descriptions in real-time.  The success of the project illustrates new potential for artificial intelligence due to the ability of the system to utilize voice and object recognition in conjunction.

Low-level speech recognition technology, like the technology presently used by smart phones and voice assistants, are programmed through an arduous training process of analysis, filtering, and digitization.  To train these systems, countless hours of recordings and their associated transcriptions must be processed so that the language can be mapped.  The “training” process becomes more complex when there is a necessity to learn different accents or dialects.

The system that MIT has developed learns more like a human child. Similar to the board books used to teach language and word association to toddlers, the system is given words in an audio format and objects in the form of raw images, which it learns to associate with one-another.