One of a series of six white papers for my client on the subject of AI and the different types of machine learning. Each acted as a companion piece for the others, ensuring that a highly complicated subject could be explained in easily digestible chunks.
AI and ML challenges
There are many testing everyday situations that AI and ML continue to face:
When humans make decisions, they not only take the immediate facts into account, they also apply common-sense filters. These tend to be intuitive and not definable by rules. However, ask a computer to generate a sentence using the words “dog, ball, catch, throw” and it could come up with “The dogs are throwing balls at each other”. We know that’s nonsensical, but the computer doesn’t.
This is closely related to common sense knowledge. Related to this is the qualification problem – the impossibility of listing all preconditions required for a real-world action to have its intended effect. Unexpected circumstances may prevent something from happening – a human would see this, or at least understand why, but a computer probably wouldn’t.
This refers to the exponential growth rate at which search problems grow. Chess is a good example here. A computer may be able to beat even a chess grandmaster, but what it can’t do is completely analyse a game from start to finish. There are simply far too many variables involved. The focus now is on heuristic search to effectively reduce the search parameters and make them manageable.
At least 10 English words have hundreds of definitions each – for example, “go” and “put”. The word “set” has 430 senses listed in the current Oxford English Dictionary, but that is likely to be overtaken in the next edition by the word “run”, with over 600 different senses. The challenge for AI is trying to understand the context of a word and use it correctly.
Complex planning scenarios have been handled by computers for decades – for example transport and staff scheduling. However, working out how efficient plans are, or if a new plan will work at all, still has some way to go. The AI and ML challenge is to improve feedback loops plus checks and balances to ensure that planning becomes more efficient over time.
The issue here is two-fold: learning the wrong things and failing to learn the right ones. A good example is the Tay chatbot mentioned above. Related to this is the issue of knowing for sure that improvements are occurring – that the learning is having the right results.