Knowledge based or (and?) Machine Learning Based AI?
Since its origins, AI (Artificial Intelligence) has swung between formal reasoning techniques based on knowledge (such as rule bases systems and Bayesian Networks ) and machine learning based techniques.
With the recent renaissance in deep learning, the pendulum has swung once again. This has reached a point where deep learning is now so synonymous with AI, that many people consider any solution which does not include deep learning to not really be AI. Moreover, world leading scientists like Turing award winner Prof. Judea Pearl feel the need to justify the importance of other methods (see Prof. Pearl's The Seven Tools of Causal Inference, with Reflections on Machine Learning) .
Some currently believe that we should only utilize deep learning based techniques, relying solely on computational power and the availability of data (see, for example http://www.incompleteideas.net/IncIdeas/BitterLesson.html) while others advocate an approach combining the two techniques (see https://www.pcmag.com/news/368508/the-ai-breakthrough-will-require-researchers-burying-their-h).
I, for one, am an advocate for using deep learning together with the formal reasoning techniques and believe that the next big leap in AI can be made only by finding the right trade-offs in combining the two methodologies.
What do you think? Should we indeed be investing efforts in trying to combine the two approaches? Or should we be focusing solely on deep learning, and stop wasting our time on knowledge based techniques which may introduce serious misconceptions we may have both about the world we live in and the way we think and reason about it?
The input of humanistic ways of thinking ,is very important to the output provided by AI. There is a whole lot going on in the more "Shallows" of due to so much of our thinking processes being there. We are only privy to 10-14% of the conscious mental abilities of our brains, that's being generous. Computers will think faster than we can imagine and Deep Learning is something we ,,should be doing as Humans. The philosophical concepts of "Thought" show us an element ,,,without limits,,,,, in a universe full of limitations,...Making a more sentient machine, is like making a bigger fish. I like being at the top of the food chain. We also know ,that we do not know if this has been a problem for things,or if AI has happened before,.The whole machine vs organics could be a thing already , with all the time that the universe appears to have had, there are a whole lot of potentials that are possible. Our progression in logical thinking as Human needs a real update to just catch up . Here is an example of how backwards we are,: We encourage and pay for "Football" in our schools, a game proven to have a 99% brain damage rate,,, at school, a place we send our children to learn.... This is an obvious objective truth. Now try explaining our logic ,of just this one example,,,, to a Deep Thinking logical sentient machine. What will it think of us? The computer Created and should considered be as a Great Library of us ,of our thoughts of ,our past thoughts ,while also playing a role as a Great Librarian, to assist us in a logical progression as humans. AI is not the first" tool ",that we have made that can be dangerous if you don't understand ,,,how it works....A" library "is a tool to provide a source of knowledge. a record keeping expanding the connection and Memory of each and every Human Being who will use it,...That is a safe way to innovate it... It is a safe way to conceptualize it. That's just my thoughts and I enjoy thinking ,for myself..
I would say, at least in decision making, symbolic models can be built with the help of machine learning techniques, but solved with formal methods, that provide guarantees on solutions. I see the future in such an approach.
Segev Wasserkrug. I cannot agree more on trying to combine the two techniques. Deep learning, and in general Machine Learning, is just one of the AI techniques. Each AI technique addresses well specific problems, on the other hand only through a 'smart' combination of different AI techniques we can solve better these problems and address new ones.
Segev Wasserkrug - I agree with your thoughts on a combination of both principle approaches. Machine learning is a strong mechanism on information that we as a human are not able to express in formal facts and relationships, like pictures, voice, mass data correlations, but it lacks of being able to explain why a pictures shows a ball or a car. On the other hand, knowledge based systems as you mentioned in your article has been left aside for a long time. But knowldge based systems can explain why the result was chosen. The downside of todays solutions...building up the knowledge. And here I personally see the challenge for AI in the next wave, how does a system gain knowledge, build abstractions, create relationships between them. ML has strengths, formal reasoning has its strengths, maybe a combination of both principles would open the door for real learning / building up knowledge.
Lior Limonad - I definitely agree with your thoughts. I believe the pendulum in recent years may have swung too far towards the deep learning only approach (which is probably a natural reaction to it swinging towards the knowledge/model based approaches in previous years). I hope where it ultimately lands will be a place that overall, better advances AI