Okay I have a lot of confusion in regards to the way likelihood functions are defined in the context of different machine learning algorithms. For the context of this discussion, I will reference Andrew Ng 229 lecture notes.
Here is my understanding thus far.
In the context of classification, we have two different types of algorithms: discriminative and generative. The goal in both of these cases is to determine the posterior probability, that is p(C_k|x;w), where w is parameter vector and x is feature vector and C_k is kth class. The approaches are different as in discriminative we are trying to solve for the posterior probability directly given x. And in the generative case, we are determining the conditional distributions p(x|C_k), and prior classes p(C_k), and using Bayes theorem to determine P(C_k|x;w).
From my understanding Bayes theorem takes the form: p(parameters|data) = p(data|parameters)p(parameters)/p(data) where the likelihood function is p(data|parameters), posterior is p(parameters|data) and prior is p(parameters).
Now in the context of linear regression, we have the likelihood function:
p(y|X;w) where y is the vector of target values, X is design matrix.
This makes sense in according to how we defined the likelihood function above.
Now moving over to classification, the likelihood is defined still as p(y|X;w). Will the likelihood always be defined as such ?
The posterior probability we want is p(y_i|x;w) for each class which is very weird since this is apparently the likelihood function as well.
When reading through a text, it just seems the likelihood is always defined to different ways, which just confuses me profusely. Is there a difference in how the likelihood function should be interpreted for regression vs classification or say generative vs discriminative. I.e the way the likelihood is defined in Gaussian discriminant analysis looks very different.
If anyone can recommend resources that go over this in detail I would appreciate this.
A quick answer is that the likelihood function is a function proportional to the probability of seeing the data conditional on all the parameters in your model. As you said in linear regression it is p(y|X,w) where w is your vector of regression coefficients and X is your design matrix.
In a classification context, your likelihood would be proportional to P(y|X,w) where y is your vector of observed class labels. You do not have a y_i for each class, because your training data was observed to be in one particular class. Given your model specification and your model parameters, for each observed data point you should be able to calculate the probability of seeing the observed class. This is your likelihood.
The posterior predictive distribution, p(y_new_i|X,y), is the probability you want in paragraph 4. This is distinct from the likelihood because it is the probability for some unobserved case, rather than the likelihood, which relates to your training data. Note that I removed w because typically you would want to marginalize over it rather than condition on it because there is still uncertainty in the estimate after training your model and you would want your predictions to marginalize over that rather than condition on one particular value.
As an aside, the goal of all classification methods is not to find a posterior distribution, only Bayesian methods are really concerned with a posterior and those methods are necessarily generative. There are plenty of non-Bayesian methods and plenty of non-probabilistic discriminative models out there.
Any function proportional to p(a|b) where a is fixed is a likelihood function for b. Note that p(a|b) might be called something else, depending on what's interesting at the moment. For example, p(a|b) can also be called the posterior for a given b. The names don't really matter.
Related
We know we need 4 things for building a machine learning algorithm:
A Dataset
A Model
A cost function
An optimization procedure
Taking the example of linear regression (y = m*x +q) we have two most common way of finding the best parameters: using ML or MSE as cost functions.
We hypotize data are Gaussian-distributed, using ML.
Is this assumption part of the model, also?
It it's not, why? Is it part of the cost function?
I can't see the "edge" of the model, in this case.
Is this assumption part of the model, also?
Yes it is. The ideas of different loss functions derived from the nature of the problem, consequently the nature of the model.
MSE by definition calculates for the mean of the squares of the errors (error means the difference between real y and predicted y) which in its turn will be high if the data is not Gaussian-Like distributed. Just imagine a few extreme values among the data, what will happen to the line slope and consequently the residual error?
It is worth mentioning the assumptions of Linear Regression:
Linear relationship
Multivariate normality
No or little multicollinearity
No auto-correlation
Homoscedasticity
If it's not, why? Is it part of the cost function?
As far I have seen, the assumption is not directly related to the cost function itself, rather related -as above-mentioned- to the model itself.
For example, Support Vector Machine idea is separation of classes. That’s finding out a line/ hyper-plane (in multidimensional space that separate outs classes), thus its cost function is Hinge Loss to "maximum-margin" of classification.
On the other hand, Logistic Regression uses Log-Loss (related to cross-entropy) because the model is binary and works on the probability of the output (0 or 1). And the list goes on...
The assumption that the data is Gaussian-distributed is part of the model in the sense that, for Gaussian distributed data the minimal Mean Squared Error also yields the maximum liklelihood solution for the data, given the model parameters. (Common proof, you can look it up if you are interested).
So you could say that the Gaussian distribution assumption justifies the choice of least squares as the loss function.
Most examples of neural networks for classification tasks I've seen use the a softmax layer as output activation function. Normally, the other hidden units use a sigmoid, tanh, or ReLu function as activation function. Using the softmax function here would - as far as I know - work out mathematically too.
What are the theoretical justifications for not using the softmax function as hidden layer activation functions?
Are there any publications about this, something to quote?
I haven't found any publications about why using softmax as an activation in a hidden layer is not the best idea (except Quora question which you probably have already read) but I will try to explain why it is not the best idea to use it in this case :
1. Variables independence : a lot of regularization and effort is put to keep your variables independent, uncorrelated and quite sparse. If you use softmax layer as a hidden layer - then you will keep all your nodes (hidden variables) linearly dependent which may result in many problems and poor generalization.
2. Training issues : try to imagine that to make your network working better you have to make a part of activations from your hidden layer a little bit lower. Then - automaticaly you are making rest of them to have mean activation on a higher level which might in fact increase the error and harm your training phase.
3. Mathematical issues : by creating constrains on activations of your model you decrease the expressive power of your model without any logical explaination. The strive for having all activations the same is not worth it in my opinion.
4. Batch normalization does it better : one may consider the fact that constant mean output from a network may be useful for training. But on the other hand a technique called Batch Normalization has been already proven to work better, whereas it was reported that setting softmax as activation function in hidden layer may decrease the accuracy and the speed of learning.
Actually, Softmax functions are already used deep within neural networks, in certain cases, when dealing with differentiable memory and with attention mechanisms!
Softmax layers can be used within neural networks such as in Neural Turing Machines (NTM) and an improvement of those which are Differentiable Neural Computer (DNC).
To summarize, those architectures are RNNs/LSTMs which have been modified to contain a differentiable (neural) memory matrix which is possible to write and access through time steps.
Quickly explained, the softmax function here enables a normalization of a fetch of the memory and other similar quirks for content-based addressing of the memory. About that, I really liked this article which illustrates the operations in an NTM and other recent RNN architectures with interactive figures.
Moreover, Softmax is used in attention mechanisms for, say, machine translation, such as in this paper. There, the Softmax enables a normalization of the places to where attention is distributed in order to "softly" retain the maximal place to pay attention to: that is, to also pay a little bit of attention to elsewhere in a soft manner. However, this could be considered like to be a mini-neural network that deals with attention, within the big one, as explained in the paper. Therefore, it could be debated whether or not Softmax is used only at the end of neural networks.
Hope it helps!
Edit - More recently, it's even possible to see Neural Machine Translation (NMT) models where only attention (with softmax) is used, without any RNN nor CNN: http://nlp.seas.harvard.edu/2018/04/03/attention.html
Use a softmax activation wherever you want to model a multinomial distribution. This may be (usually) an output layer y, but can also be an intermediate layer, say a multinomial latent variable z. As mentioned in this thread for outputs {o_i}, sum({o_i}) = 1 is a linear dependency, which is intentional at this layer. Additional layers may provide desired sparsity and/or feature independence downstream.
Page 198 of Deep Learning (Goodfellow, Bengio, Courville)
Any time we wish to represent a probability distribution over a discrete variable with n possible values, we may use the softmax function. This can be seen as a generalization of the sigmoid function which was used to represent a probability
distribution over a binary variable.
Softmax functions are most often used as the output of a classifier, to represent the probability distribution over n different classes. More rarely, softmax functions can be used inside the model itself, if we wish the model to choose between one of n different options for some internal variable.
Softmax function is used for the output layer only (at least in most cases) to ensure that the sum of the components of output vector is equal to 1 (for clarity see the formula of softmax cost function). This also implies what is the probability of occurrence of each component (class) of the output and hence sum of the probabilities(or output components) is equal to 1.
Softmax function is one of the most important output function used in deep learning within the neural networks (see Understanding Softmax in minute by Uniqtech). The Softmax function is apply where there are three or more classes of outcomes. The softmax formula takes the e raised to the exponent score of each value score and devide it by the sum of e raised the exponent scores values. For example, if I know the Logit scores of these four classes to be: [3.00, 2.0, 1.00, 0.10], in order to obtain the probabilities outputs, the softmax function can be apply as follows:
import numpy as np
def softmax(x):
z = np.exp(x - np.max(x))
return z / z.sum()
scores = [3.00, 2.0, 1.00, 0.10]
print(softmax(scores))
Output: probabilities (p) = 0.642 0.236 0.087 0.035
The sum of all probabilities (p) = 0.642 + 0.236 + 0.087 + 0.035 = 1.00. You can try to substitute any value you know in the above scores, and you will get a different values. The sum of all the values or probabilities will be equal to one. That’s makes sense, because the sum of all probability is equal to one, thereby turning Logit scores to probability scores, so that we can predict better. Finally, the softmax output, can help us to understand and interpret Multinomial Logit Model. If you like the thoughts, please leave your comments below.
Here in SO I found the following explanation of generative and discriminitive algorithms:
"A generative algorithm models how the data was generated in order to categorize a signal. It asks the question: based on my generation assumptions, which category is most likely to generate this signal?
A discriminative algorithm does not care about how the data was generated, it simply categorizes a given signal."
And here is the definition for parametric and nonparametric algorithms
"Parametric: data are drawn from a probability distribution of specific form up to unknown parameters.
Nonparametric: data are drawn from a certain unspecified probability distribution.
"
So essentially can we say that generative and parametric algorithms assume underlying model whereas discriminitve and nonparametric algorithms dont assume any model?
thanks.
Say you have inputs X (probably a vector) and output Y (probably univariate). Your goal is to predict Y given X.
A generative method uses a model of the joint probability p(X,Y) to determine P(Y|X). It is thus possible given a generative model with known parameters to sample jointly from the distribution p(X,Y) to produce new samples of both input X and output Y (note they are distributed according to the assumed, not true, distribution if you do this). Contrast this to discriminative approaches which only have a model of the form p(Y|X). Thus provided with input X they can sample Y; however, they cannot sample new X.
Both assume a model. However, discriminative approaches assume only a model of how Y depends on X, not on X. Generative approaches model both. Thus given a fixed number of parameters you might argue (and many have) that it's easier to use them to model the thing you care about, p(Y|X), than the distribution of X since you'll always be provided with the X for which you wish to know Y.
Useful references: this (very short) paper by Tom Minka. This seminal paper by Andrew Ng and Michael Jordan.
The distinction between parametric and non-parametric models is probably going to be harder to grasp until you have more stats experience. A parametric model has a fixed and finite number of parameters regardless of how many data points are observed. Most probability distributions are parametric: consider a variable z which is the height of people, assumed to be normally distributed. As you observe more people, your estimate for the parameters \mu and \sigma, the mean and standard deviation of z, become more accurate but you still only have two parameters.
In contrast, the number of parameters in a non-parametric model can grow with the amount of data. Consider an induced distribution over peoples' heights which places a normal distribution over each observed sample, with mean given by the measurement and fixed standard deviation. The marginal distribution over new heights is then a mixture of normal distributions, and the number of mixture components increases with each new data point. This is a non-parametric model of people's height. This specific example is called a kernel density estimator. Popular (but more complicated) non parametric models include Gaussian Processes for regression and Dirichlet Processes.
A pretty good tutorial on non-parametrics can be found here, which constructs the Chinese Restaurant Process as the limit of a finite mixture model.
I don't think you can say it. E.g. linear regression is a discriminative algorithm - you make an assumption about P(Y|X), and then estimate paramenters directly from the data, without making any assumption about P(X) or P(X|Y), as you would do in case of generative models. But at the same time, aby inference based on linear regression, including the properties of the paramenters, is a parametric estimation, as there is an assumption about behaviour of unobserved errors.
Here I'm only talking about parametric/non-parametric. Generative/ discriminative is a separate concept.
Non-parametric model means you don't make any assumptions on the distribution of your data. For example, in the real world, data will not 100% follow theoretical distributions like Gaussian, beta, Poisson, Weibull, etc. Those distributions are developed for our need's to model the data.
On the other hand, parametric models try to completely explain our data using parameters. In practice, this way is preferred because it makes easier to define how the model should behave in different circumstances (for example, we already know the derivative/gradients of the model, what happens when we set the rate too high/too low in Poisson, etc.)
When we train a training set using decision tree classifier, we will get a tree model. And this model can be converted to rules and can be incorporated into a java code.
Now if I train the training set using Naive Bayes, in what form is the model? And how can I incorporated the model into my java code?
If there is no model resulted from the training, then what is the difference between Naive Bayes and lazy learner (ex. kNN)?
Thanks in advance.
Naive Bayes constructs estimations of conditional probabilities P(f_1,...,f_n|C_j), where f_i are features and C_j are classes, which, using bayes rule and estimation of priors (P(C_j)) and evidence (P(f_i)) can be translated into x=P(C_j|f_1,...,f_n), which can be roughly read as "Given features f_i I think, that their describe object of class C_j and my certainty is x". In fact, NB assumes that festures are independent, and so it actualy uses simple propabilities in form of x=P(f_i|C_j), so "given f_i I think that it is C_j with probability x".
So the form of the model is set of probabilities:
Conditional probabilities P(f_i|C_j) for each feature f_i and each class C_j
priors P(C_j) for each class
KNN on the other hand is something completely different. It actually is not a "learned model" in a strict sense, as you don't tune any parameters. It is rather a classification algorithm, which given training set and number k simply answers question "For given point x, what is the major class of k nearest points in the training set?".
The main difference is in the input data - Naive Bayes works on objects that are "observations", so you simply need some features which are present in classified object or absent. It does not matter if it is a color, object on the photo, word in the sentence or an abstract concept in the highly complex topological object. While KNN is a distance-based classifier which requires you to classify object which you can measure distance between. So in order to classify abstract objects you have to first come up with some metric, distance measure, which describes their similarity and the result will be highly dependent on those definitions. Naive Bayes on the other hand is a simple probabilistic model, which does not use the concept of distance at all. It treats all objects in the same way - they are there or they aren't, end of story (of course it can be generalised to the continuous variables with given density function, but it is not the point here).
The Naive Bayes will construct/estimate the probability distribution from which your training samples have been generated.
Now, given this probability distribution for all your output classes, you take a test sample, and depending on which class has the highest probability of generating this sample, you assign the test sample to that class.
In short, you take the test sample and run it through all the probability distributions (one for each class) and calculate the probability of generating this test sample for that particular distribution.
I have a set of images. I would like to learn a one class SVM (OC-SVM) to model the distribution of a particular class (positive) as I dont have enough examples to represent the other classes (negative). What I understood about OC-SVM is that it tries to separate the data from the origin or in other words it tries to learn a hyper sphere to fit the one class data.
My questions are,
If I want to use the output of the OC-SVM as a probability estimate, how can I do it?
What is the difference between the OC-SVM and any clustering algorithm (e.g. k-means)?
If you want a probability estimate, don't use a one-class SVM. This is not what they were designed for. You want something like kernel density estimation, which provides a non-parametric density estimate given some positive examples.
The difference between a one-class SVM and clustering is that in clustering, you're given points from several classes but you don't know which points correspond to which classes: this is the goal of inference (and you may also end up with density estimates for the classes and the marginal density over all of feature space too). The one-class SVM is given points only from one class, and expected to learn a separation between members of that class and anything else.
EDIT: Clustering is not the same as density estimation. Clustering is concerned with determining which instances belong to which classes (clusters), when the assignments are not given, and does not necessarily result in a similarity score between the supplied examples and any point in input space.
If the goal is to say, how similar is this new instance to the positive training examples I've seen, then what you do is fit a probability distribution to your training examples, then evaluate the density function at the new point. If this density falls below a threshold, you say the new point is outside of the class defined by the supplied examples.
You can build a parametric model of the class if you like, but this is usually tricky unless you either know something about the problem or are willing to take a standard distribution (multi-variate normal or Naive Bayes being the two obvious ones). So, the alternative is to use a non-parametric density estimate. This is the kernel density estimation I mentioned.