What problems will “AI” solve over the next 5–10 years? An AI/Machine Learning primer for entrepreneurs without PhDs
The process to create this guide: 1) (2015) become frustrated that everyone is talking about AI and I’m not sure what they mean 2) start asking around, turns out most people don’t know either 3) learn how to apply basic machine learning methods 4) read interviews with Stanford PhDs 5) talk to Stanford PhDs myself 6) simplify it so that me back at step 1 would be able to understand.
Artificial Intelligence (AI) vs Machine Learning (ML)
Artificial intelligence is anything that involves teaching computer systems to do tasks that normally only people can do.
Machine learning is a software technique that uses data to make predictions on data.
Many people that talk about AI in 2016 are referring to ML.
What is machine learning (“ML”) successfully being used for so far?
- Speech to text
- Text to speech
- Language translation
- Ranking search results on Google
- Handwriting recognition
- Self-driving cars
- Finance (algorithmic trading to predict market movements, valuing companies, valuing houses and other assets)
What can we infer about what problems are suited to machine learning?
They provide very large amounts of training data. Training data is data that we can feed into a computer that allows it to learn what normally happens so that it can better make predictions.
The data is available in a computer readable form (opportunities exist in digitizing data).
The output is representable by a machine (e.g. self-driving cars the output is “turn left, accelerate, etc”, easily representable by a machine).
Is “problem X” likely well suited to machine learning?
Is the problem represented in data?
Can you get large amounts of computer readable training data — i.e. data that is “correctly labeled”. e.g. if you wanted to convert between languages, training data might be like this: { [english: “I like you”, french: “je t’aime”], with many more examples}. Then the machine can look at all of this training data, and then when you feed it an English sentence it hasn’t seen before, it can make a guess at what the French equivalent would be based on all of the other data. A prediction.
Likely and/or hoped for progress in the next 5 years or so
Language translation.
Voice recognition.
Dialogue systems. Avoiding verbal dialogue systems that ask highly repetitive questions and don’t allow for quick clarifications is a challenge in this area. For example, if you ask SIRI to find a flight to Bermuda and it returns you a flight to New Jersey, you can’t fix it by saying, “No, I meant Bermuda.” SIRI looks differentially for actions (go to the grocery story), times (tomorrow), and places (San Francisco), but the system that does this is fairly brittle.
Text summary.
Self-driving cars.
In-principle capability to automate a significant fraction of factory work (though it would likely require additional translational work for specific types of factory work that would be automated).
Online education — for example, it would probably be possible to give people targeted explanations based on the problems they are getting wrong.
Online dating (algorithms for assigning people to likely matches could be improved).
Medicine — e.g. diagnosis and treatment selection.
Credit — e.g. deciding whom to give loans to. There is some regulation that makes this hard, so many of the challenges are legal rather than technical.
Energy — e.g. one could imagine monitoring cell phone transmissions and figuring out what factors affect energy use, and then use nudges to affect those factors in a way that decreases energy use. More ambitiously, a machine learning system might manage a power grid for a city.
Science — representing scientific knowledge; automating steps in science (especially biology); structured search (e.g. identifying all of the proteins relevant to a given process).
Customer service call centers.
