Hacker News Reader

Comment by indigochill | Hacker News Reader

A girl I went to high school with went on to apply machine learning to identify human trafficking victims and ads for law enforcement.

My perception of machine learning as a mere dabbler myself is that "machine learning" is just a sci-fi name for what's essentially applied statistics. In places where that is useful (e.g. clustering ads by feature similarity to highlight unclassified ads that appear similar to known trafficking ads), then machine learning is useful. It's not necessarily as one-size-fits-all as, say, networking or operating systems are, but in cases where you can identify a useful application of statistics, machine learning can be a useful tool.

View on HN

ML (neural nets) are useful because they automate much of parameter search. Instead of trying an SVM and then Random forests and then whatever, just throw all the data at a GPU and let it build an NN for you.

NNs come with many caveats and lack the theoretical guarantees and support of SVMs and RFs. They should never be the first approach to solving a problem unless a) you have good reason to support the hypothesis that other models don't cut it such as a non stationary problem e.g. RL, or b) other methods can't scale to the difficulty of the problem. NNs are also very expensive to train compared to everything else.

It's a rare case when a neural net will give you materially different results than a random forest or an svm. ie, if an svm is 80% accurate, a well tuned neural net might be 83% accurate. sometimes that's a huge difference. in a domain like medical image classification, maybe not so much

also you can just "throw all the data" at an svm or random forest, or any number of similar models. Automated parameter tuning can be convenient, but it's prone to overfitting and doesn't eliminate a lot of the actual work

The great tragedy of AI: a promise of creating synthetic sentience, a reality of mundane number crunching.

That's the question. If consciousness/intelligence is wholly in the brain, and the brain's interaction with the physical world, is that something that can be modeled with number crunching, i.e. computation? Perhaps we are still very early days.

a black hole is entirely a product of physical processes. how long until we can make one? or even fully understand them

> My perception of machine learning as a mere dabbler myself is that "machine learning" is just a sci-fi name for what's essentially applied statistics

As a physicist I used to say that ML is simply non-linear tensor calculus. (I’m not sure if I’m right though)

I don't think you are, if you have never heard of Reinforcement Learning, I think you are in for a treat.

Reply on news.ycombinator.com