Slashdot Mirror
DARPA Tackles Machine Learning
coondoggie writes "Researchers at DARPA want to take the science of machine learning — teaching computers to automatically understand data, manage results and surmise insights — up a couple notches. Machine learning, DARPA says, is already at the heart of many cutting edge technologies today, like email spam filters, smartphone personal assistants and self-driving cars. 'Unfortunately, even as the demand for these capabilities is accelerating, every new application requires a Herculean effort. Even a team of specially-trained machine learning experts makes only painfully slow progress due to the lack of tools to build these systems,' DARPA says."
Even a team of specially-trained machine learning experts makes only painfully slow progress due to the lack of tools to build these systems
Why not just teach a machine to do it?
sudo ergo sum
Defense agency investing in Machine Learning technology? What could possibly go wrong?!
a headline for future 2030.
They've been trying it since the 50s without it has to be said, too much success given the amount of effort thats been put in. I suspect until we REALLY understand how boligical brains do it (not , "meh, some sort of neural back propagation", yeah , we know that , but what propagation and how exactly?) then machine learning will still remain at the bottom rung of the intelligence ladder.
Personally I think at the moment pre programmed intelligence is still a more successful route to go down. Though hopefully that will change.
That ad is super annoying. But as an answer to parent, I actually leave ads *on* for slashdot - only for their gesture to let me turn them off. You know the setting up in the corner on slashdot. This alone made me keep them on, to support them.
Waiting for you by the bridge
They're hard coded and use massively parallel depth searching. The brute force approach has been the best for chess computers for decades.
And google search and translate isn't really learning, they're just statistical systems that given the best result based on the data they've gathered. They don't "think" about it in any meaningful way.
There are a ton of off-the-shelf machine learning toolkits that are sufficient for 90% of possible use cases. The problem is getting annotated data to feed into these tools so they can learn the appropriate patterns. But all that requires is a host of annotators (i.e. undergrads and interns), not machine learning experts.
Its a form of A.I. for sure, but the skill shown has more to do with the volume of data it uses than it has to do with a skill at learning.
Machine Learning is a very particular subset of A.I, often characterized by one or more training phases which build of model of the training set that is smaller than the set itself.
"His name was James Damore."
Jeopardy, and the machine is called Watson
Sounds like the 1990s fetish for making programming languages so simple that even your boss could make reports and do other stuff for himself. Unfortunately, programming language syntax wasn't the primary hurdle: I've had bosses request reports that would add pounds of product and shipping costs.
For ML, it takes a good bit of training just to know what kinds of problems you can apply it to. A cookbook toolkit isn't going to reduce the need for expertise very much.
Sheesh, evil *and* a jerk. -- Jade
I was just talking with someone about this the other day. Machine learning is going to be the SQL database of the next generation. In 15 years it will be hard to find basic apps that don't use it. The tools will reach a point that it's so easy to include them in your program, people will assume to include them even though they may not really be the most appropriate method to solve the problem. This is how SQL is today. Go to any SMB and try to find a non-trivial application that doesn't use a SQL database. It's difficult.
.net or Java, they won't use it. Weka can be used for Java, but it's a difficult library for a machine learning novice to use. The developer has to know some internals of machine learning to know which algorithm to use and their pros and cons. Meta learners complicate the issue even more. Modern RDBMS have been sugar coated so much a developer can use a RAD IDE and not understand a single line of SQL. I'm not saying that's really a good thing, but it definitely has made SQL databases very common and improved the state of the industry for everyone.
However, the state of current tools is not good. We currently have really good algorithms for machine learning. The gap is in actually getting a developer to use them. If it's not branded and blessed by Oracle or Microsoft, many businesses won't use it. If you search for implementations on the internet you can usually find an implementation of R or Matlab. However, people are weary of including R and Matlab in their programs to begin with. If it's not in
Raw data need to be cleaned up and organized to feed into the ML algorithm.
The results of the ML algorithm need to be cleaned up and organized so that they can be used by the rest of the system.
No one (currently) can tell you which ML algorithm will work best on your problem and how its parameters should be chosen without a lot of study. Preconceived bias (e.g., that it should be biologically based, blah, blah) can be a killer here.
The best results typically come from combinations of ML algorithms through some kind of ensemble learning, so now your have the problem of choosing a good combination and choosing a lot more parameters.
All of the above need to work together in concert.
Certainly, it's not a bad idea to try to make this process better, but I wouldn't be expecting miracles too soon.