I am new to ML and going through that i found difficulty in choosing the optimizer there are lots of optimizer, but I am confused when to use the Stochastic Gradient Descent optimizer and Why?
Here is my understanding: the modern, sophisticated gradient descent algorithms like ADAM use more computational power than SGD. They make trial moves before committing to a step down the gradient. This extra sophistication is quite helpful when working with deeper neural network architectures. Such architectures may have very complex loss hypersurfaces, with multiple saddle points and local minima. ADAM and its relatives do a good job finding the solution in these situations.
If your model architectures are simpler, apparently that the extra effort is not worth it. Stochastic gradient descent works well for simple models, and if you add Nesterov momentum and weight decay, you can often converge on a solution faster than you can with ADAM.
Another place that SGD is used is as a substitute for solving very large linear regressions by ordinary least-squares. In theory, any invertible matrix can be solved exactly. But it's entirely possible in the modern age to have a matrix inversion problem whose intermediate terms require more memory than you have in RAM. In that case, SGD is the more practical way to solve the problem.
I am trying to use machine learning to predict a dataset. It is a regression problem with 180 input features and 1 continuously-valued output. I try to compare deep neural networks, random forest regression, and linear regression.
As I expect, 3-hidden-layer deep neural networks outperform other two approaches with a root mean square error (RMSE) of 0.1. However, I unexpected to see that random forest even performs worse than linear regression (RMSE 0.29 vs. 0.27). In my expectation, the random forest can discover more complex dependencies between features to decrease error. I have tried to tune the parameters of random forest (number of trees, maximum features, max_depth, etc.). I also tried different K-cross validation, but the performance is still less than linear regression.
I searched online, and one answer says linear regression may perform better if features have a smooth, nearly linear dependence on the covariates. I do not fully get the point because if that is the case, should not deep neural networks give much performance gain?
I am struggling to give an explanation. Under what situation, random forest is worse than linear regression, but deep neural networks can perform much better?
If your features explain linear relation to the target variable then a Linear Model usually performs well than a Random Forest Model. It totally depends on the linear relations between your features.
That said, Linear models are not superior or the Random Forest is any inferior one.
Try scaling and transforming the data using MinMaxScaler() from scikit-learn to see if the linear model improves further
Pro Tips
If linear model is working like a charm you need to ask your self Why? and How? And get into the basics of both the models to understand why it worked on your data. These questions will lead you to feature engineer better. And as a matter of fact, Kaggle Grand Masters do use Linear Models in stacking to get that top 1% score by capturing the linear relations in the dataset.
So at the end of the day, linear models could wonders too.
My colleague and I are trying to wrap our heads around the difference between logistic regression and an SVM. Clearly they are optimizing different objective functions. Is an SVM as simple as saying it's a discriminative classifier that simply optimizes the hinge loss? Or is it more complex than that? How do the support vectors come into play? What about the slack variables? Why can't you have deep SVM's the way you can't you have a deep neural network with sigmoid activation functions?
I will answer one thing at at time
Is an SVM as simple as saying it's a discriminative classifier that simply optimizes the hinge loss?
SVM is simply a linear classifier, optimizing hinge loss with L2 regularization.
Or is it more complex than that?
No, it is "just" that, however there are different ways of looking at this model leading to complex, interesting conclusions. In particular, this specific choice of loss function leads to extremely efficient kernelization, which is not true for log loss (logistic regression) nor mse (linear regression). Furthermore you can show very important theoretical properties, such as those related to Vapnik-Chervonenkis dimension reduction leading to smaller chance of overfitting.
Intuitively look at these three common losses:
hinge: max(0, 1-py)
log: y log p
mse: (p-y)^2
Only the first one has the property that once something is classified correctly - it has 0 penalty. All the remaining ones still penalize your linear model even if it classifies samples correctly. Why? Because they are more related to regression than classification they want a perfect prediction, not just correct.
How do the support vectors come into play?
Support vectors are simply samples placed near the decision boundary (losely speaking). For linear case it does not change much, but as most of the power of SVM lies in its kernelization - there SVs are extremely important. Once you introduce kernel, due to hinge loss, SVM solution can be obtained efficiently, and support vectors are the only samples remembered from the training set, thus building a non-linear decision boundary with the subset of the training data.
What about the slack variables?
This is just another definition of the hinge loss, more usefull when you want to kernelize the solution and show the convexivity.
Why can't you have deep SVM's the way you can't you have a deep neural network with sigmoid activation functions?
You can, however as SVM is not a probabilistic model, its training might be a bit tricky. Furthermore whole strength of SVM comes from efficiency and global solution, both would be lost once you create a deep network. However there are such models, in particular SVM (with squared hinge loss) is nowadays often choice for the topmost layer of deep networks - thus the whole optimization is actually a deep SVM. Adding more layers in between has nothing to do with SVM or other cost - they are defined completely by their activations, and you can for example use RBF activation function, simply it has been shown numerous times that it leads to weak models (to local features are detected).
To sum up:
there are deep SVMs, simply this is a typical deep neural network with SVM layer on top.
there is no such thing as putting SVM layer "in the middle", as the training criterion is actually only applied to the output of the network.
using of "typical" SVM kernels as activation functions is not popular in deep networks due to their locality (as opposed to very global relu or sigmoid)
The motivating idea behind neural nets seems to be that they learn the "right" features to apply logistic regression to. Is there a similar approach for linear regression? (or just regression problems in general?)
Would doing the obvious thing of removing the application of a sigmoid function for all neurons (ie, including the hidden layers) make sense/work? (ie, each neuron is performing linear regression instead of logistic regression).
Alternatively, would doing the (maybe even more obvious) thing of just scaling output values to [0,1] work? (intuitively I would think not, as the sigmoid function seems like it would cause the net to arbitrarily favor extreme values) (edit: though I was just searching around some more, and saw that one technique is to scale based on mean and variance, which seems like it might deal with this issue -- so maybe this is more viable than I thought).
Or is there some other technique for doing "feature learning" for regression problems?
Check out this applet. Try to learn different functions. When you dictate linear activation functions at both hidden and output layers, it even fails to learn the quadratic function. At least one layer needs to be set to sigmoid function, see figures below.
There are different kinds of scaling. Standard scaling, as you mentioned, eliminates the impact of mean and standard deviation of the training sample, is most often used in machine learning. Just make sure you are using the same mean and std value from training sample in the test sample.
The reason why scaling is required is because the output of sigmoid function ranges at (0,1). I didn't try, but I think it is better to scale the output even if you select linear function at output layer. Otherwise large input at hidden layer (with sigmoid) won't lead to drastic output (the sigmoid function is approximately linear when the input is at a small range, out of such range will make the output changes much slowly). You can try this by yourself in your own data.
Besides, if you have various features, the feature normalization that makes different features in the same scale is also recommended. The scaling speeds up gradient descent by avoiding many extra iterations that are required when one or more features take on much larger values than the rest.
As #Ray mentioned, deep learning that many levels of features are involved can help you with the feature learning, it's not all linear combinations though.
I know SVMs are supposedly 'ANN killers' in that they automatically select representation complexity and find a global optimum (see here for some SVM praising quotes).
But here is where I'm unclear -- do all of these claims of superiority hold for just the case of a 2 class decision problem or do they go further? (I assume they hold for non-linearly separable classes or else no-one would care)
So a sample of some of the cases I'd like to be cleared up:
Are SVMs better than ANNs with many classes?
in an online setting?
What about in a semi-supervised case like reinforcement learning?
Is there a better unsupervised version of SVMs?
I don't expect someone to answer all of these lil' subquestions, but rather to give some general bounds for when SVMs are better than the common ANN equivalents (e.g. FFBP, recurrent BP, Boltzmann machines, SOMs, etc.) in practice, and preferably, in theory as well.
Are SVMs better than ANN with many classes? You are probably referring to the fact that SVMs are in essence, either either one-class or two-class classifiers. Indeed they are and there's no way to modify a SVM algorithm to classify more than two classes.
The fundamental feature of a SVM is the separating maximum-margin hyperplane whose position is determined by maximizing its distance from the support vectors. And yet SVMs are routinely used for multi-class classification, which is accomplished with a processing wrapper around multiple SVM classifiers that work in a "one against many" pattern--i.e., the training data is shown to the first SVM which classifies those instances as "Class I" or "not Class I". The data in the second class, is then shown to a second SVM which classifies this data as "Class II" or "not Class II", and so on. In practice, this works quite well. So as you would expect, the superior resolution of SVMs compared to other classifiers is not limited to two-class data.
As far as i can tell, the studies reported in the literature confirm this, e.g., In the provocatively titled paper Sex with Support Vector Machines substantially better resolution for sex identification (Male/Female) in 12-square pixel images, was reported for SVM compared with that of a group of traditional linear classifiers; SVM also outperformed RBF NN, as well as large ensemble RBF NN). But there seem to be plenty of similar evidence for the superior performance of SVM in multi-class problems: e.g., SVM outperformed NN in protein-fold recognition, and in time-series forecasting.
My impression from reading this literature over the past decade or so, is that the majority of the carefully designed studies--by persons skilled at configuring and using both techniques, and using data sufficiently resistant to classification to provoke some meaningful difference in resolution--report the superior performance of SVM relative to NN. But as your Question suggests, that performance delta seems to be, to a degree, domain specific.
For instance, NN outperformed SVM in a comparative study of author identification from texts in Arabic script; In a study comparing credit rating prediction, there was no discernible difference in resolution by the two classifiers; a similar result was reported in a study of high-energy particle classification.
I have read, from more than one source in the academic literature, that SVM outperforms NN as the size of the training data decreases.
Finally, the extent to which one can generalize from the results of these comparative studies is probably quite limited. For instance, in one study comparing the accuracy of SVM and NN in time series forecasting, the investigators reported that SVM did indeed outperform a conventional (back-propagating over layered nodes) NN but performance of the SVM was about the same as that of an RBF (radial basis function) NN.
[Are SVMs better than ANN] In an Online setting? SVMs are not used in an online setting (i.e., incremental training). The essence of SVMs is the separating hyperplane whose position is determined by a small number of support vectors. So even a single additional data point could in principle significantly influence the position of this hyperplane.
What about in a semi-supervised case like reinforcement learning? Until the OP's comment to this answer, i was not aware of either Neural Networks or SVMs used in this way--but they are.
The most widely used- semi-supervised variant of SVM is named Transductive SVM (TSVM), first mentioned by Vladimir Vapnick (the same guy who discovered/invented conventional SVM). I know almost nothing about this technique other than what's it is called and that is follows the principles of transduction (roughly lateral reasoning--i.e., reasoning from training data to test data). Apparently TSV is a preferred technique in the field of text classification.
Is there a better unsupervised version of SVMs? I don't believe SVMs are suitable for unsupervised learning. Separation is based on the position of the maximum-margin hyperplane determined by support vectors. This could easily be my own limited understanding, but i don't see how that would happen if those support vectors were unlabeled (i.e., if you didn't know before-hand what you were trying to separate). One crucial use case of unsupervised algorithms is when you don't have labeled data or you do and it's badly unbalanced. E.g., online fraud; here you might have in your training data, only a few data points labeled as "fraudulent accounts" (and usually with questionable accuracy) versus the remaining >99% labeled "not fraud." In this scenario, a one-class classifier, a typical configuration for SVMs, is the a good option. In particular, the training data consists of instances labeled "not fraud" and "unk" (or some other label to indicate they are not in the class)--in other words, "inside the decision boundary" and "outside the decision boundary."
I wanted to conclude by mentioning that, 20 years after their "discovery", the SVM is a firmly entrenched member in the ML library. And indeed, the consistently superior resolution compared with other state-of-the-art classifiers is well documented.
Their pedigree is both a function of their superior performance documented in numerous rigorously controlled studies as well as their conceptual elegance. W/r/t the latter point, consider that multi-layer perceptrons (MLP), though they are often excellent classifiers, are driven by a numerical optimization routine, which in practice rarely finds the global minimum; moreover, that solution has no conceptual significance. On the other hand, the numerical optimization at the heart of building an SVM classifier does in fact find the global minimum. What's more that solution is the actual decision boundary.
Still, i think SVM reputation has declined a little during the past few years.
The primary reason i suspect is the NetFlix competition. NetFlix emphasized the resolving power of fundamental techniques of matrix decomposition and even more significantly t*he power of combining classifiers. People combined classifiers long before NetFlix, but more as a contingent technique than as an attribute of classifier design. Moreover, many of the techniques for combining classifiers are extraordinarily simple to understand and also to implement. By contrast, SVMs are not only very difficult to code (in my opinion, by far the most difficult ML algorithm to implement in code) but also difficult to configure and implement as a pre-compiled library--e.g., a kernel must be selected, the results are very sensitive to how the data is re-scaled/normalized, etc.
I loved Doug's answer. I would like to add two comments.
1) Vladimir Vapnick also co-invented the VC dimension which is important in learning theory.
2) I think that SVMs were the best overall classifiers from 2000 to 2009, but after 2009, I am not sure. I think that neural nets have improved very significantly recently due to the work in Deep Learning and Sparse Denoising Auto-Encoders. I thought I saw a number of benchmarks where they outperformed SVMs. See, for example, slide 31 of
http://deeplearningworkshopnips2010.files.wordpress.com/2010/09/nips10-workshop-tutorial-final.pdf
A few of my friends have been using the sparse auto encoder technique. The neural nets build with that technique significantly outperformed the older back propagation neural networks. I will try to post some experimental results at artent.net if I get some time.
I'd expect SVM's to be better when you have good features to start with. IE, your features succinctly capture all the necessary information. You can see if your features are good if instances of the same class "clump together" in the feature space. Then SVM with Euclidian kernel should do the trick. Essentially you can view SVM as a supercharged nearest neighbor classifier, so whenever NN does well, SVM should do even better, by adding automatic quality control over the examples in your set. On the converse -- if it's a dataset where nearest neighbor (in feature space) is expected to do badly, SVM will do badly as well.
- Is there a better unsupervised version of SVMs?
Just answering only this question here. Unsupervised learning can be done by so-called one-class support vector machines. Again, similar to normal SVMs, there is an element that promotes sparsity. In normal SVMs only a few points are considered important, the support vectors. In one-class SVMs again only a few points can be used to either:
"separate" a dataset as far from the origin as possible, or
define a radius as small as possible.
The advantages of normal SVMs carry over to this case. Compared to density estimation only a few points need to be considered. The disadvantages carry over as well.
Are SVMs better than ANNs with many classes?
SVMs have been designated for discrete classification. Before moving to ANNs, try ensemble methods like Random Forest , Gradient Boosting, Gaussian Probability Classification etc
What about in a semi-supervised case like reinforcement learning?
Deep Q learning provides better alternatives.
Is there a better unsupervised version of SVMs?
SVM is not suited for unsupervised learning. You have other alternatives for unsupervised learning : K-Means, Hierarchical clustering, TSNE clustering etc
From ANN perspective, you can try Autoencoder, General adversarial network
Few more useful links:
towardsdatascience
wikipedia