4 ms·

For a novice, it's hard to assess how important PGMs are currently. Investing my time to learn Deep Learning (CNNs, RNNs) vs. Random Forest vs. PGMs vs. Reinf

For a novice, it's hard to assess how important PGMs are currently. Investing my time to learn Deep Learning (CNNs, RNNs) vs. Random Forest vs. PGMs vs. Reinforcement Learning well enough to be able to apply the chosen approach, it seems that PGMs are not high in the list, is that correct? Are there Kaggle competitions, in which PGMs have been the best approach? What are the real-world problem areas that PGM's currently excel at compared to other methods?

eachro 9y ago
Random Forests is not really at all in the same class of depth(no pun intended) as the other models you've mentioned; if you understand decision trees, random forests are just a way of combining independently trained decision trees that is less prone to overfitting. I'd probably rank these areas in order of importance as follows: deep learning, pgms, reinforcement learning. Deep learning as a framework is pretty general. PGMs, as I have seen them, don't really have any one killer domain area - maybe robotics and areas where you want to explicitly model causality? Applications for reinforcement learning seem the more niche, but maybe that's because they haven't been adequately explored to the extent that DL/CNNs/RNNs/PGMs have been.
- digitalzombie 9y ago
  Tree based algorithm have it's place. I dunno what your definition of depth is but it all depends on your data. I'm doing a thesis on tree based algorithm and it works great for medical data. Granted I have little exposer to NN, you can't do that with NN when clinical trials data is small as hell. It's all base on the type data and your resource and criteria. NN is really hype and people tend to over look other algorithms but NN is not a silver bullet. I don't get how you rank those importance either, it could be bias.
  - eachro 9y ago
    Oh there's no question that tree based methods are effective - random forests/gradient boosted trees routinely win kaggle competitions. But I was more referring to how random forests is learnable in a day, whereas deep learning would probably take at least a few weeks to learn properly.
- jhbadger 9y ago
  In analysis of scientific data, causality is important. Typically in science you aren't really making a model to predict things (although that can be a good way to test the model) but rather to understand what is going on, and PGMs are great for that.
mannigfaltig 9y ago
https://www.quora.com/What-are-some-applications-of-Probabilistic-Graphical-Models https://www.quora.com/What-are-some-applications-of-Probabil... I'd say for people with interest in ML and DL, the main insight from PGMs is intuitions about probability theory, especially conditional probabilities and the concept of (un)conditional independence. State of the art in which PGMs would be relevant is currently mostly brute-force learning via stochastic gradient descent (with momentum) in deep directed models/function approximators, but function approximation likely does not solve all problems as adversarial examples show. There are certainly very important applications of PGMs in all kinds of domains as listed in the link above (perhaps I would have added FastSLAM [1]), and there are also variational autoencoders/Helmholz machines and (Deep) Boltzmann Machines, which are especially interesting for research perspective. [1] http://robots.stanford.edu/papers/montemerlo.fastslam-tr.pdf http://robots.stanford.edu/papers/montemerlo.fastslam-tr.pdf

eachro 9y ago

Random Forests is not really at all in the same class of depth(no pun intended) as the other models you've mentioned; if you understand decision trees, random forests are just a way of combining independently trained decision trees that is less prone to overfitting. I'd probably rank these areas in order of importance as follows: deep learning, pgms, reinforcement learning. Deep learning as a framework is pretty general. PGMs, as I have seen them, don't really have any one killer domain area - maybe robotics and areas where you want to explicitly model causality? Applications for reinforcement learning seem the more niche, but maybe that's because they haven't been adequately explored to the extent that DL/CNNs/RNNs/PGMs have been.

digitalzombie 9y ago
Tree based algorithm have it's place. I dunno what your definition of depth is but it all depends on your data. I'm doing a thesis on tree based algorithm and it works great for medical data. Granted I have little exposer to NN, you can't do that with NN when clinical trials data is small as hell. It's all base on the type data and your resource and criteria. NN is really hype and people tend to over look other algorithms but NN is not a silver bullet. I don't get how you rank those importance either, it could be bias.
- eachro 9y ago
  Oh there's no question that tree based methods are effective - random forests/gradient boosted trees routinely win kaggle competitions. But I was more referring to how random forests is learnable in a day, whereas deep learning would probably take at least a few weeks to learn properly.
jhbadger 9y ago
In analysis of scientific data, causality is important. Typically in science you aren't really making a model to predict things (although that can be a good way to test the model) but rather to understand what is going on, and PGMs are great for that.

mannigfaltig 9y ago

https://www.quora.com/What-are-some-applications-of-Probabilistic-Graphical-Models https://www.quora.com/What-are-some-applications-of-Probabil... I'd say for people with interest in ML and DL, the main insight from PGMs is intuitions about probability theory, especially conditional probabilities and the concept of (un)conditional independence. State of the art in which PGMs would be relevant is currently mostly brute-force learning via stochastic gradient descent (with momentum) in deep directed models/function approximators, but function approximation likely does not solve all problems as adversarial examples show. There are certainly very important applications of PGMs in all kinds of domains as listed in the link above (perhaps I would have added FastSLAM [1]), and there are also variational autoencoders/Helmholz machines and (Deep) Boltzmann Machines, which are especially interesting for research perspective. [1] http://robots.stanford.edu/papers/montemerlo.fastslam-tr.pdf http://robots.stanford.edu/papers/montemerlo.fastslam-tr.pdf