Ask YC: Real World Applications of Machine Learning?
15 comments
Basically any topic where you know what hypothesis look like, and there is lots of data floating around.
Medical diagnosis is an area where ML should be used, but isn't. It's been discovered that a decision tree beats a cardiologist for heart attack triage, but no one will touch it (1). In research, it is often used to discover genetic regulatory networks, and other such things. Basically, ML is replacing what would be a boring job for a grad student.
Also look at marketing. Thanks to their ML systems, Walmart knows that people buy poptarts to prepare for a hurricane.
(1) If a cardiologist kills 10/1000 people due to human error (whoops, forgot the aspirin), it happens. If a computer program kills 5/1000 people due to classification error it's 5 lawsuits waiting to happen, provided you can find a cardiologist to say "In hindsight, I would have gotten that right."
Medical diagnosis is an area where ML should be used, but isn't. It's been discovered that a decision tree beats a cardiologist for heart attack triage, but no one will touch it (1). In research, it is often used to discover genetic regulatory networks, and other such things. Basically, ML is replacing what would be a boring job for a grad student.
Also look at marketing. Thanks to their ML systems, Walmart knows that people buy poptarts to prepare for a hurricane.
(1) If a cardiologist kills 10/1000 people due to human error (whoops, forgot the aspirin), it happens. If a computer program kills 5/1000 people due to classification error it's 5 lawsuits waiting to happen, provided you can find a cardiologist to say "In hindsight, I would have gotten that right."
I met a guy at a university commercialization seminar that was trying to make ML/decision tree stuff for medical diagnoses and such. I don't think he was thinking of triage but much more basic decisions. I hope he got some funding, but this was also in Pittsburgh.
many thanks, these are good ideas and your examples are awesome
An out-of-left-field option, if your goal is only to make some money working with ML: go into independent game design and create a game around a ML system.
It's not something mainstream gaming would be interested in using because the goals are different - there the development process is enslaved to some well-defined intellectual property, and (with rare exceptions) does not roam freely enough to allow more than a "good-enough AI opponent" - but if you developed a concept and brought it to market yourself, making sure that the result is fun to play and easy to learn, as well as a good use of tech, you have a shot at a decent revenue stream. Probably not enough to live on, but the beauty of independent gaming is that you can keep revising and expanding on an innovative concept, gradually growing your own market over a period of many years.
It's not something mainstream gaming would be interested in using because the goals are different - there the development process is enslaved to some well-defined intellectual property, and (with rare exceptions) does not roam freely enough to allow more than a "good-enough AI opponent" - but if you developed a concept and brought it to market yourself, making sure that the result is fun to play and easy to learn, as well as a good use of tech, you have a shot at a decent revenue stream. Probably not enough to live on, but the beauty of independent gaming is that you can keep revising and expanding on an innovative concept, gradually growing your own market over a period of many years.
Here's a ML game that I think would be cool:
The idea is genetic programming with human interaction based around a Pokemon kind of game. There would be pre-made and player made pets, and a "wild" area where creatures could evolve.
how about an online "hot-or-not" type of game for docs where you are given the same diagnostic information that would be given to a cardiologist and you can see if you can beat the decision tree?
I think someone made something like this not too long ago for stock prices vs. random walks..
I think someone made something like this not too long ago for stock prices vs. random walks..
Actually we're doing that... If you hear about our company in 5 - 6 years, that means it worked!
On 2, we are a company founded by a combination of CS/Physics people. We have found that pairing a physics person with a CS person leads to the best outcomes.
I think you should emphasize the things you are good at (probably modeling and data analysis) and not worry too much about the CS gaps There are a ton of opportunities out there for your skill-set.
I think you should emphasize the things you are good at (probably modeling and data analysis) and not worry too much about the CS gaps There are a ton of opportunities out there for your skill-set.
(1) All sorts of information and collaborative filtering. Examples: Netflix, Amazon, Pandora, etc. Lots on ML in defense these days too -- i.e. here's everyone's call data, who's the terrorist?
(2) If you just want to implement or tweak existing algorithms, it doesn't really matter what your background is -- the qualifications are the same as any software engineering job (so CS degree helps a lot but not mandatory). If you want to be in a research type position, you'll need a PhD in CS with a dissertation on ML.
(2) If you just want to implement or tweak existing algorithms, it doesn't really matter what your background is -- the qualifications are the same as any software engineering job (so CS degree helps a lot but not mandatory). If you want to be in a research type position, you'll need a PhD in CS with a dissertation on ML.
> If you want to be in a research type position, you'll need a PhD in CS with a dissertation on ML.
Comp Bio works fine too. Same difference.
Comp Bio works fine too. Same difference.
I suggest you get out and talk to friends and their friends at your school, about the problems in their fields, that you might be able to help with. They might not identify their problem as needing ML, but you can connect those dots yourself.
I think your chance of success is much higher if you work directly with someone who knows the problem domain in depth. Cross-domain work is where its at.
I think your chance of success is much higher if you work directly with someone who knows the problem domain in depth. Cross-domain work is where its at.
Tag grouping for smarter search.
I just read an article about the guys who won the spock challenge
http://www.spock.com/do/pages/pr_spock_challenge_winner
check out www.zillow.com
Sebastian Seung's research group at MIT is using machine learning to automatically segment the profiles of sectioned axons and dendrites in electron microscopical images of 60 nm serial sections. A very hard problem, and currently in high demand among neuroscience labs.
The purpose: to reconstruct all neuronal connections of a piece of brain tissue.
http://hebb.mit.edu/
The purpose: to reconstruct all neuronal connections of a piece of brain tissue.
http://hebb.mit.edu/
(1) What are some specific commercial applications of ML technology (other than banking/finance which I'm not interested in for personal reasons) I've been looking at medical devices / biometrics, but any other suggestions would be helpful.
(2) Would someone with a physics background be taken seriously? I have a decent amount of experience doing numerical work and data analysis with python / c++, but I suspect that these types of jobs are hyper-competitive even for people with hard-core CS backgrounds.