3 Important 2016 Developments for AI in Business Operations

Common wisdom says that to "make AI work" you need lots of data. This is why AI in business operations is hard. There is just not a lot of data that exists on let's say processing Bills of Laden in trade finance. Even high-volume processes like OFAC sanction screening have data sets that are often much smaller than what Google uses to train its image recognition or translation algorithms.

Things are changing very quick though. Here is what happened in 2016:

1. Deep and symbolic learning are being combined. Symbolic learning has tons of enterprise-friendly advantages over deep learning: less data needed, less fragile, etc. Combining the 2 approaches is a great news for enterprise (read this for details). Garnelo published Towards deep symbolic reinforcement learning on this in 2016.
2. Portability of nets from one task to another is improving. Used to be the case that AI trained on one data set had to be retrained when it is applied to another. Looks like that will not hold up for much longer, which will help enterprise scale up AI automations. OpenAI published Third Person Imitation Learning on this in 2016.
3. Reinforcement learning is being automated. Neural nets are now used to tune training itself making it faster and easier to train on smaller data sets. In November, OpenAI published a paper on this called RL^2 Fast Reinforcement Learning Via Slow Reinforcement Learning, and DeepMind published Learning to Reinforcement Learn.

AI field is developing incredibly fast. Forrester confirms that AI investments will go up 300% in 2017.

Stay current by following me @alyashok

Originally published on LinkedIn.