Microsoft has announced the formation of Microsoft Research AI, a research and development division within the company that is specifically devoted to finding out what the biggest obstacles in AI are, overcoming those, and applying the lessons learned and principles developed to more traditional AI work. Things that AI can’t typically do well by nature, such as multi-faceted comprehension and sorting work, are set to be tackled through an integrated approach that entails multiple teams working together to reach breakthroughs in different AI specializations, then bringing those areas together to enhance or enable AI tasks. Along with an AI experiment lab, Microsoft will also be putting together an AI ethics board that will work both independently and with other prominent AI research entities to ensure that AI research is being pursued and applied in a safe and ethical manner.
One of the simplest examples of this approach in action is Microsoft’s work in machine reading. The field requires both normal machine learning to enhance comprehension, natural language processing for putting words together into ideas and topics, and machine vision to identify words. AI can surpass humans in one-dimensional rote processing tasks, but something multifaceted like this can trip up most purpose-built AIs by virtue of the need to combine AI disciplines, and the vast number of extra possibilities and potential inputs and outputs that it adds. The practically endless options mean that more than one AI must be built, and they must work together as one. This means that the research teams behind them have to work together to closely integrate them. This is the sort of approach that Microsoft Research AI plans to take in getting over some of AI’s biggest issues.
This approach is not entirely new; Google’s DeepMind is one example of an AI research outfit aimed squarely at getting past AI’s biggest issues, but its method is a bit different. Whereas Microsoft Research AI will hand-pick the absolute hardest problems in order to troubleshoot them and reap the benefits across applications, DeepMind tends to make AI do incredibly complicated things in a task-oriented and goal-oriented way, then apply learnings outward.