Is there a way to ‘teach’ common sense? Microsoft co-founder Paul Allen certainly believes so, at least as far as machines are concerned—which is why Allen is putting $125 million into research that will imbue artificial intelligence with common sense. Allen had earlier founded the Seattle-based non-profit, Allen Institute for Artificial Intelligence (AI2), that aims to be at the forefront of breakthroughs in developing AI capability. AI2’s Project Alexandria will be focussing on creating an AI system that exhibits good judgement. For perspective on how far we are from that, commonly used AI today can respond to basic commands, tell apart objects, identify words and languages and even translate. AI has also beaten humans at complex human designed games. But it still can’t exhibit the judgement that an average 10-year-old is capable of. It relies on big data pools and learning—which is why, as New Scientist points out, AI can be taught to play a video-game and even excel at it, to the extent that it beats human players, but try and get it to use what it has learnt to play a different game based on similar fundamentals, and it can’t. It has to learn to play the second game from scratch. AI systems may be able to learn without our active control over the process but what they acquire is of little consequence beyond the context in which they gathered the learning. Neural networks, of course, have eased this a bit by helping AI systems recognise what is salient for solving a new problem. But this is not free from glitches. In many cases, the manifestation of neural network processing is incomprehensible to humans and can’t be passed on to another neural network.
AI2’s focus is on developing systems that can learn fast, without requiring the volume of information that they currently need to arrive at a comparable level of judgement—it is like learning to avoid putting your finger into fire without having stuck it into open flame, once or ten times. Once AI is capable of unstructured problem-solving or handling unanticipated solutions, it would have unlocked machine-common-sense. Common sense is both a precondition and marker of overall intelligence. For AI to be truly intelligent, it must acquire this.