Naming conventions for neural networks - machine-learning

Are there any standard naming conventions for neural networks? I am new to programming neural networks but have noticed a lot of consistency across tutorials and books and would like to get familiar with the naming conventions but can't find any sources. An example is that 'X' is usually capitalised and 'y' is lowercase for datasets. Some other terminology below.
NB_EPOCH = 200
BATCH_SIZE = 128
VERBOSE = 1
NB_CLASSES = 10
N_HIDDEN

If you are talking about naming conventions of neural network frameworks and libraries (e.g. Keras, PyTorch, Caffe etc), the answer is generally no.
Keras itself is a good example case, where the parameter name for the number of epochs changed from nb_epoch in Keras 1 to epochs in Keras 2 (causing enough headaches in the transition...). MXNet uses num_epoch, while lower level libraries, such as Tensorflow and PyTorch, do not use epochs at all (notice that in PyTorch there are still complaints for confusion between using 'epoch' and 'iteration' in the tutorial texts).
It is true that, say, batch_size (in lowercase), tends to be standard (verbose too, but this is a much more general parameter, and not confined to neural networks).
So, in principle, you should not expect cross-platform compatibility in the naming conventions of the parameters.

Related

How to Customize Metric for GridSearchCV in Scikit Learn to tune for specific class?

I have a use case in ML where I have 2 classes, 0 and 1 for a given text.
Class-0: Can afford some misclassifications
Class-1: Very Important, can't afford any misclassifications
There's a huge imbalance in samples for both classes,
about 30000 for class-0, and only 1000 for class-1
While doing the train-test split, I'm stratifying the split based on the labels, such that, the ratio of 70% train and 30% test is maintained for each label class.
I want to tune parameters in such a way that Precision or Recall for class-1 is improved. I tried using 'f1_macro', 'precision', 'recall' as individual metrics and all combined as well to tune using GridSearchCV, but it's less helpful due to majority samples being Class-0.
I'm exploring the safer ways to reduce class 0 data, although, there's only small degree we can reduce, anyways even without tuning, or with any parameters, class-0 always have above 98% f1-score.
So all I care about tuning is for class-1.
Can you please suggest, perhaps a customized callable metric such that it only focuses on Class-1's Precision, Recall or F1-Score?
I'm using scikit-learn latest stable version.
Similar Problem here, the author is trying to Tune Class-1's F1 Score using Neural Networks (MLP) in Keras
Its been suggested to try customizing metric, just didn't mention how.
The one who can answer here for Scikit-Learn, can also answer below link for Keras.
Hyperparameter tuning in Keras (MLP) via RandomizedSearchCV
Using class_weight='balanced' is helping here.
I referred these articles in Scikit-Learn's official documentation pages.
Understanding how parameter class_weights works:
https://scikit-learn.org/stable/modules/svm.html#unbalanced-problems
https://stackoverflow.com/a/30982811/3149277
Understanding what parameters to use for class_weights:
https://scikit-learn.org/stable/modules/svm.html#tips-on-practical-use
How does the class_weight parameter in scikit-learn work?
Although, due to time limits, I didn't bother defining the custom function as this seemed working close to my expectations.

Sigmoid activation for multi-class classification?

I am implementing a simple neural net from scratch, just for practice. I have got it working fine with sigmoid, tanh and ReLU activations for binary classification problems. I am now attempting to use it for multi-class, mutually exclusive problems. Of course, softmax is the best option for this.
Unfortunately, I have had a lot of trouble understanding how to implement softmax, cross-entropy loss and their derivatives in backprop. Even after asking a couple of questions here and on Cross Validated, I can't get any good guidance.
Before I try to go further with implementing softmax, is it possible to somehow use sigmoid for multi-class problems (I am trying to predict 1 of n characters, which are encoded as one-hot vectors)? And if so, which loss function would be best? I have been using the squared error for all binary classifications.
Your question is about the fundamentals of neural networks and therefore I strongly suggest you start here ( Michael Nielsen's book ).
It is python-oriented book with graphical, textual and formulated explanations - great for beginners. I am confident that you will find this book useful for your understanding. Look for chapters 2 and 3 to address your problems.
Addressing your question about the Sigmoids, it is possible to use it for multiclass predictions, but not recommended. Consider the following facts.
Sigmoids are activation functions of the form 1/(1+exp(-z)) where z is the scalar multiplication of the previous hidden layer (or inputs) and a row of the weights matrix, in addition to a bias (reminder: z=w_i . x + b where w_i is the i-th row of the weight matrix ). This activation is independent of the others rows of the matrix.
Classification tasks are regarding categories. Without any prior knowledge ,and even with, most of the times, categories have no order-value interpretation; predicting apple instead of orange is no worse than predicting banana instead of nuts. Therefore, one-hot encoding for categories usually performs better than predicting a category number using a single activation function.
To recap, we want an output layer with number of neurons equals to number of categories, and sigmoids are independent of each other, given the previous layer values. We also would like to predict the most probable category, which implies that we want the activations of the output layer to have a meaning of probability disribution. But Sigmoids are not guaranteed to sum to 1, while softmax activation does.
Using L2-loss function is also problematic due to vanishing gradients issue. Shortly, the derivative of the loss is (sigmoid(z)-y) . sigmoid'(z) (error times the derivative), that makes this quantity small, even more when the sigmoid is closed to saturation. You can choose cross entropy instead, or a log-loss.
EDIT:
Corrected phrasing about ordering the categories. To clarify, classification is a general term for many tasks related to what we used today as categorical predictions for definite finite sets of values. As of today, using softmax in deep models to predict these categories in a general "dog/cat/horse" classifier, one-hot-encoding and cross entropy is a very common practice. It is reasonable to use that if the aforementioned is correct. However, there are (many) cases it doesn't apply. For instance, when trying to balance the data. For some tasks, e.g. semantic segmentation tasks, categories can have ordering/distance between them (or their embeddings) with meaning. So please, choose wisely the tools for your applications, understanding what their doing mathematically and what their implications are.
What you ask is a very broad question.
As far as I know, when the class become 2, the softmax function will be the same as sigmoid, so yes they are related. Cross entropy maybe the best loss function.
For the backpropgation, it is not easy to find the formula...there
are many ways.Since the help of CUDA, I don't think it is necessary to spend much time on it if you just want to use the NN or CNN in the future. Maybe try some framework like Tensorflow or Keras(highly recommand for beginers) will help you.
There is also many other factors like methods of gradient descent, the setting of hyper parameters...
Like I said, the topic is very abroad. Why not trying the machine learning/deep learning courses on Coursera or Stanford online course?

Huge number of classes with Multinominal Naive Bayes (scikit-learn)

Whenever I start having a bigger number of classes (1000 and more) MultinominalNB gets super slow and takes Gigabytes of RAM. The same is true for all the scikit learn classification algorithms that support .partial_fit() (SGDClassifier, Perceptron).
When working with convolutional neural networks 10000 classes are no problem. But when I want to train MultinominalNB on the same data my 12GB of RAM are not enough and it is very very slow.
From my understanding of Naive Bayes, even with a lot of classes, it should be a lot faster.
Might this be a problem of the scikit-learn implementation (maybe of the .partial_fit() function) ? How can I train MultinominalNB/SGDClassifier/Perceptron on 10000+ classes (batchwise)?
Short answer without much information:
The MultinomialNB fits an independent model to each of the classes, thus, if you have C=10000+ classes it will fit C=10000+ models and therefore, only the model parameters will be [n_classes x n_features], which is quite a lot of memory if n_features is large.
The SGDClassifier of scikits-learn uses OVA (one-versus-all) strategy to train a multiclass model (as the SGDC is not inherently multiclass) and therefore, another C=10000+ models need to be trained.
And Perceptron, from the documentation of scikits-learn:
Perceptron and SGDClassifier share the same underlying implementation. In fact, Perceptron() is equivalent to SGDClassifier(loss=”perceptron”, eta0=1, learning_rate=”constant”, penalty=None).
So, all the 3 classifiers you mention don't work well with high number of classes, as an independent model needs to be trained for each of the classes. I would recommend you to try something that inherently support multiclass classification, such as RandomForestClassifier.

Scalable or online out-of-core multi-label classifiers

I have been blowing my brains out over the past 2-3 weeks on this problem.
I have a multi-label (not multi-class) problem where each sample can belong to several of the labels.
I have around 4.5 million text documents as training data and around 1 million as test data. The labels are around 35K.
I am using scikit-learn. For feature extraction I was previously using TfidfVectorizer which didn't scale at all, now I am using HashVectorizer which is better but not that scalable given the number of documents that I have.
vect = HashingVectorizer(strip_accents='ascii', analyzer='word', stop_words='english', n_features=(2 ** 10))
SKlearn provides a OneVsRestClassifier into which I can feed any estimator. For multi-label I found LinearSVC & SGDClassifier only to be working correctly. Acc to my benchmarks SGD outperforms LinearSVC both in memory & time. So, I have something like this
clf = OneVsRestClassifier(SGDClassifier(loss='log', penalty='l2', n_jobs=-1), n_jobs=-1)
But this suffers from some serious issues:
OneVsRest does not have a partial_fit method which makes it impossible for out-of-core learning. Are there any alternatives for that?
HashingVectorizer/Tfidf both work on a single core and don't have any n_jobs parameter. It's taking too much time to hash the documents. Any alternatives/suggestions? Also is the value of n_features correct?
I tested on 1 million documents. The Hashing takes 15 minutes and when it comes to clf.fit(X, y), I receive a MemoryError because OvR internally uses LabelBinarizer and it tries to allocate a matrix of dimensions (y x classes) which is fairly impossible to allocate. What should I do?
Any other libraries out there which have reliable & scalable multi-label algorithms? I know of genism & mahout but both of them don't have anything for multi-label situations?
I would do the multi-label part by hand. The OneVsRestClassifier treats them as independent problems anyhow. You can just create the n_labels many classifiers and then call partial_fit on them. You can't use a pipeline if you only want to hash once (which I would advise), though.
Not sure about speeding up hashing vectorizer. You gotta ask #Larsmans and #ogrisel for that ;)
Having partial_fit on OneVsRestClassifier would be a nice addition, and I don't see a particular problem with it, actually. You could also try to implement that yourself and send a PR.
The algorithm that OneVsRestClassifier implements is very simple: it just fits K binary classifiers when there are K classes. You can do this in your own code instead of relying on OneVsRestClassifier. You can also do this on at most K cores in parallel: just run K processes. If you have more classes than processors in your machine, you can schedule training with a tool such as GNU parallel.
Multi-core support in scikit-learn is work in progress; fine-grained parallel programming in Python is quite tricky. There are potential optimizations for HashingVectorizer, but I (one of the hashing code's authors) haven't come round to it yet.
If you follow my (and Andreas') advice to do your own one-vs-rest, this shouldn't be a problem anymore.
The trick in (1.) applies to any classification algorithm.
As for the number of features, it depends on the problem, but for large scale text classification 2^10 = 1024 seems very small. I'd try something around 2^18 - 2^22. If you train a model with L1 penalty, you can call sparsify on the trained model to convert its weight matrix to a more space-efficient format.
My argument for scalability is that instead of using OneVsRest which is just a simplest of simplest baselines, you should use a more advanced ensemble of problem-transformation methods. In my paper I provide a scheme for dividing label space into subspaces and transforming the subproblems into multi-class single-label classifications using Label Powerset. To try this, just use the following code that utilizes a multi-label library built on top of scikit-learn - scikit-multilearn:
from skmultilearn.ensemble import LabelSpacePartitioningClassifier
from skmultilearn.cluster import IGraphLabelCooccurenceClusterer
from skmultilearn.problem_transform import LabelPowerset
from sklearn.linear_model import SGDClassifier
# base multi-class classifier SGD
base_classifier = SGDClassifier(loss='log', penalty='l2', n_jobs=-1)
# problem transformation from multi-label to single-label multi-class
transformation_classifier = LabelPowerset(base_classifier)
# clusterer dividing the label space using fast greedy modularity maximizing scheme
clusterer = IGraphLabelCooccurenceClusterer('fastgreedy', weighted=True, include_self_edges=True)
# ensemble
clf = LabelSpacePartitioningClassifier(transformation_classifier, clusterer)
clf.fit(x_train, y_train)
prediction = clf.predict(x_test)
The partial_fit() method was recently added to sklearn so hopefully it should be available in the upcoming release (it's in the master branch already).
The size of your problem makes it attractive to tackling it with neural networks. Have a look at magpie, it should give much better results than linear classifiers.

When should I use support vector machines as opposed to artificial neural networks?

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

Resources