Since deep learning began in earnest in 2012, AI technology has surpassed the performance of existing technologies in many fields. Although it is a limited environment, Atari games reached the human level in 2015, image recognition and Go in 2016, skin cancer diagnosis and voice recognition, and poker in 2017. AI is showing close to human performance.
However, that doesn't mean that AI has reached a level close to humans. Because today's AI technology is only specialized in any one of the many things humans can do, it is far from humans' comprehensive thinking skills.
An article posted on VentureBeat predicts that 2021 AI Next Steps will attempt to address these challenges:
In the shared article, four main topics are mentioned.
- Hybrid Artificial Intelligence
- Inspiration from evolution
- Reinforcement learning
- Integrating world knowledge and common sense into AI
Hybrid artificial intelligence is used with existing rule-based legacy technologies to solve the problems of current deep learning, such as the need for excessive data, that transfer learning is not easy, and inference and knowledge expression are unnatural. It refers to an attempt to converge deep learning technologies. Of course, in the future, deep learning technology alone may solve all the problems, but I think combining the two different approaches into one could create a more efficient system with a limited amount of training data. In addition, according to one study, rather than a huge system in which the human brain is also an integrated system, a partial network that performs specific functions is formed for each part of the brain, and an analysis result that each module is connected in a specific relationship. I don't think there is a need to stick to the idea that running alone should solve all problems.
Inspiration from evolution is not an injection-type learning based on a large amount of learning data, but a form that is gradually learned by interaction with the surrounding environment. It appears to be a concept similar to online learning, but in fact, most living organisms learn through this process and acquire incredible strength even in environments they have never encountered. These areas seem worth considering in terms of what current deep learning technologies do not have.
Reinforcement learning is a technique that is already widely used in the game field. It is not a learning method based on a combination of problems and correct answers, but a method of learning to give a given environment and goal and maximize the reward for achieving that goal. When I think about it, I think that if the way teachers convey knowledge is supervised learning, the learning process that takes place through social relationships with friends and colleagues is similar to reinforcement learning. In particular, from the point of view of learning decision-making, not knowledge, we believe that it is an essential process that must be taken to make AI close to humans.
Integrating world knowledge and common sense into AI means that the concept of common sense should be applied to AI, not just a way of retrieving facts from a lot of data, but being able to infer logical correlations from accumulated knowledge. , Which means you should be able to make logical decisions from this. Through this process, AI that is not included in the training data but can adapt and generalize to similar situations will be able to emerge.
In addition to the above, many new topics are emerging in the field of AI. However, if the goal is limited to artificial general intelligence (AGI), or AI in a form close to humans, rather than thinking about who solves problems in a limited environment with higher performance, like most AI tasks at the moment, it is comprehensive in the open domain. Research is indispensable to equip you with thinking ability, and I think it is one of the directions to go forward.