4 ms·
That makes it sound like the problem is lack of data, which isn't true. The problem is that the kind of ML that involves downloading a framework from github an
by otabdeveloper4 5y ago
That makes it sound like the problem is lack of data, which isn't true.
The problem is that the kind of ML that involves downloading a framework from github and tweaking features until the percent goes up is actually built on certain statistical models under the hood that people don't understand and that don't fit the process they're trying to model.
When the statistical model is correct you don't need loads of data. E.g., you don't need more than a thousand respondents to make valid inferences about millions of people in a sociological survey.
- raverbashing 5y agoFor real. People are downloading ready-made models from repositories to try and solve minor problems. Guess what, your problem might be a simple linear regression. Yes you can solve it with a DNN (one level, one neuron - but hey, don't keep that from putting it into your CV) but you don't need to.
- MaximumYComb 5y agoAt university, I generated Markov chains of the solution space from a single neuron that was being used as a binary classifier. You take n samples, average them out and look at the decision boundary. The decision boundary itself is linear but the margin of error is not. It was really cool. Attempting to implent Hamiltonian MCMC on a single neuron really forced you to learn what a gradient is in regards to NN.
- nerdponx 5y agoBut the problem really is lack of data in many cases. Not necessarily lack of it in quantity, but lack of it in quantity in any usable/trustworthy form.
- zaphar 5y agoThere point of that statement is that you need data to train a valid model for your usecase. And you may need a lot of it if you are trying to train a neural network. Using a pre-trained model only works if the usecase it was trained for matches you're closely enough.