Batch normalization: fixed samples or different samples by dimension? - multidimensional-array

Some questions came to me as I read a paper 'Batch Normalization : Accelerating Deep Network Training by Reducing Internal Covariate Shift'.
In the paper, it says:
Since m examples from training data can estimate mean and variance of
all training data, we use mini-batch to train batch normalization
parameters.
My question is :
Are they choosing m examples and then fitting batch norm parameters concurrently, or choosing different set of m examples for each input dimension?
E.g. training set is composed of x(i) = (x1,x2,...,xn) : n-dimension
for fixed batch M = {x(1),x(2),...,x(N)}, perform fitting all gamma1~gamman and beta1~betan.
vs
For gamma_i, beta_i picking different batch M_i = {x(1)_i,...,x(m)_i}

I haven't found this question on cross-validated and data-science, so I can only answer it here. Feel free to migrate if necessary.
The mean and variance are computed for all dimensions in each mini-batch at once, using moving averages. Here's how it looks like in code in TF:
mean, variance = tf.nn.moments(incoming, axis)
update_moving_mean = moving_averages.assign_moving_average(moving_mean, mean, decay)
update_moving_variance = moving_averages.assign_moving_average(moving_variance, variance, decay)
with tf.control_dependencies([update_moving_mean, update_moving_variance]):
return tf.identity(mean), tf.identity(variance)
You shouldn't worry about technical details, here's what's going on:
First the mean and variance of the whole batch incoming are computed, along batch axis. Both of them are vectors (more precisely, tensors).
Then current values moving_mean and moving_variance are updated by an assign_moving_average call, which basically computes this: variable * decay + value * (1 - decay).
Every time batchnorm gets executed, it knows one current batch and some statistic of previous batches.

Related

Code syntax in calculating posterior distribution in WinBUGS

Recently I read "The BUGS Book – A Practical Introduction to Bayesian Analysis" to learn WinBUGS. The way WinBUGS describes the derivation of posterior distribution makes me feel confused.
Let's take Example 4.1.1 in this book to illustrae:
Suppose we observe the number of deaths y in a given hospital for a
high-risk operation. Let n denote the total number of such
operations performed and suppose we wish to make inferences regarding
the underlying true mortality rate, $\theta$.
The code of WinBUGS is:
y <- 10 # the number of deaths
n <- 100 # the total number of such operations
#########################
y ~ dbin(theta,n) # likelihood, also a parametric sampling distribution
logit(theta) <- logit.theta # normal prior for the logistic transform of theta
logit.theta ~ dnorm(0,0.368) # precision = 1/2.71
The author said that:
The software knows how to derive the posterior distribution and
subsequently sample from it.
My question is:
Which code reflects the logic structure to tell WinBUGS about "which parameter that I want to calculate its posterior distribution"?
This question seems silly, but if I do not read the background first, I truly cannot find directly in the code above about which parameter is focused on (e.g., theta, or y?).
Below are some of my thoughts (as a beginner of WinBUGS):
I think the following three attributions of the code style in WinBUGS makes me confused:
(1) the code does not follow "a specific sequence". For example, why is logit.theta ~ dnorm(0,0.368) not in front of logit(theta) <- logit.theta?
(2) repeated variable. Foe example, why did the last two lines not be reduced into one line: logit(theta) ~ dnorm(0,0.368)?
(3) variables are defined in more than one place. For example, y is defined two times: y <- 10 and y ~ dbin(theta, n). This one has been explained in Appendix A of the book (i.e., However, a check has been built in so that when finding a logical node that also features as a stochastic node, a stochastic node is created with the calculated values as fixed data), yet I still cannot catch its meaning.
BUGS is a declarative language. For the most part, statements aren't executed in sequence, they define different parts of the model. BUGS works on models that can be represented by directed acyclic graphs, i.e. those where you put a prior on some components, then conditional distributions on other components given the earlier ones.
It's a fairly simple language, so I think logit(theta) ~ dnorm(0, 0.368) is just too complicated for it.
The language lets you define a complicated probability model, and declare observations of certain components in it. Once you declare an observation, the model that BUGS samples from is the the original full model conditioned on that observation. y <- 10 defines observed data. y ~ dbin(theta,n) is part of the model.
The statement n <- 100 could be either: for fixed constants like n, it doesn't really matter which way you think of it. Either the model says that n is always 100, or n has an undeclared prior distribution not depending on any other parameter, and an observed value of 100. These two statements are equivalent.
Finally, your big question: Nothing in the code above says which parameter you want to look at. BUGS will compute the joint posterior distribution of every parameter. n and y will take on their fixed values, theta and logit.theta will both be simulated from the posterior. In another part of your code (or by using the WinBUGS menus) you can decide which of those to look at.

bnlearn::bn.fit difference and calculation of methods "mle" and "bayes"

I try to understand the differences between the two methods bayes and mle in the bn.fit function of the package bnlearn.
I know about the debate between the frequentist and the bayesian approach on understanding probabilities. On a theoretical level I suppose the maximum likelihood estimate mle is a simple frequentist approach setting the relative frequencies as the probability. But what calculations are done to get the bayes estimate? I already checked out the bnlearn documenation, the description of the bn.fit function and some application examples, but nowhere there's a real description of what's happening.
I also tried to understand the function in R by first checking out bnlearn::bn.fit, leading to bnlearn:::bn.fit.backend, leading to bnlearn:::smartSapply but then I got stuck.
Some help would be really appreciated as I use the package for academic work and therefore I should be able to explain what happens.
Bayesian parameter estimation in bnlearn::bn.fit applies to discrete variables. The key is the optional iss argument: "the imaginary sample size used by the bayes method to estimate the conditional probability tables (CPTs) associated with discrete nodes".
So, for a binary root node X in some network, the bayes option in bnlearn::bn.fit returns (Nx + iss / cptsize) / (N + iss) as the probability of X = x, where N is your number of samples, Nx the number of samples with X = x, and cptsize the size of the CPT of X; in this case cptsize = 2. The relevant code is in the bnlearn:::bn.fit.backend.discrete function, in particular the line: tab = tab + extra.args$iss/prod(dim(tab))
Thus, iss / cptsize is the number of imaginary observations for each entry in a CPT, as opposed to N, the number of 'real' observations. With iss = 0 you would be getting a maximum likelihood estimate, as you would have no prior imaginary observations.
The higher iss with respect to N, the stronger the effect of the prior on your posterior parameter estimates. With a fixed iss and a growing N, the Bayesian estimator and the maximum likelihood estimator converge to the same value.
A common rule of thumb is to use a small non-zero iss so that you avoid zero entries in the CPTs, corresponding to combinations that were not observed in the data. Such zero entries could then result in a network which generalizes poorly, such as some early versions of the Pathfinder system.
For more details on Bayesian parameter estimation you can have a look at the book by Koller and Friedman. I suppose many other Bayesian network books also cover the topic.

Preprocess data in R

Im using R to create logistic regression classifier model.
Here is the code sample:
library(ROCR)
DATA_SET <- read.csv('E:/1.csv')
classOneCount= 4000
classZeroCount = 4000
sample.churn <- sample(which(DATA_SET$Class==1),classOneCount)
sample.nochurn <- sample(which(DATA_SET$Class==0),classZeroCount )
train.set <- DATA_SET[c(sample.churn,sample.nochurn),]
test.set <- DATA_SET[c(-sample.churn,-sample.nochurn),]
full.logit <- glm(Class~., data = train.set, family = binomial)
And it works fine, but I would like to preprocess the data to see if it improves classification model.
What I would like to do would be to divide input vector variables which are continuoes into intervals. Lets say that one variable is height in centimeters in float.
Sample values of height:
183.23
173.43
163.53
153.63
193.27
and so on, and I would like to split it into lets say 3 different intervals: small, medium, large.
And do it with all variables from my set - there are 32 variables.
What's more I would like to see at the end correlation between value of the variables (this intervals) and classification result class.
Is this clear?
Thank you very much in advance
The classification model creates some decision boundary and existing algorithms are rather good at estimating it. Let's assume that you have one variable - height - and linear decision boundary. Your algorithm can then decide between what values put decision boundary by estimating error on training set. If you perform quantization and create few intervals your algorithm have fewer places to put boundary(data loss). It will likely perform worse on such cropped dataset than on original one. It could help if your learning algorithm is suffering from high variance (is overfitting data) but then you could also try getting more training examples, use smaller set (subset) of features or use algorithm with regularization and increase regularization parameter
There are also many questions about how to choose number of intervals and how to divide data into them like: should all intervals be equally frequent or of equal width or most similar to each other inside each interval?
If you want just to experiment use some software like f.e. free version of RapidMiner Studio (it can read CSV and Excel files and have some quick quantization options) to convert your data

Trying to do a simulation in R

I'm pretty new to R, so I hope you can help me!
I'm trying to do a simulation for my Bachelor's thesis, where I want to simulate how a stock evolves.
I've done the simulation in Excel, but the problem is that I can't make that large of a simulation, as the program crashes! Therefore I'm trying in R.
The stock evolves as follows (everything except $\epsilon$ consists of constants which are known):
$$W_{t+\Delta t} = W_t exp^{r \Delta t}(1+\pi(exp((\sigma \lambda -0.5\sigma^2) \Delta t+\sigma \epsilon_{t+\Delta t} \sqrt{\Delta t}-1))$$
The only thing here which is stochastic is $\epsilon$, which is represented by a Brownian motion with N(0,1).
What I've done in Excel:
Made 100 samples with a size of 40. All these samples are standard normal distributed: N(0,1).
Then these outcomes are used to calculate how the stock is affected from these (the normal distribution represent the shocks from the economy).
My problem in R:
I've used the sample function:
x <- sample(norm(0,1), 1000, T)
So I have 1000 samples, which are normally distributed. Now I don't know how to put these results into the formula I have for the evolution of my stock. Can anyone help?
Using R for (discrete) simulation
There are two aspects to your question: conceptual and coding.
Let's deal with the conceptual first, starting with the meaning of your equation:
1. Conceptual issues
The first thing to note is that your evolution equation is continuous in time, so running your simulation as described above means accepting a discretisation of the problem. Whether or not that is appropriate depends on your model and how you have obtained the evolution equation.
If you do run a discrete simulation, then the key decision you have to make is what stepsize $\Delta t$ you will use. You can explore different step-sizes to observe the effect of step-size, or you can proceed analytically and attempt to derive an appropriate step-size.
Once you have your step-size, your simulation consists of pulling new shocks (samples of your standard normal distribution), and evolving the equation iteratively until the desired time has elapsed. The final state $W_t$ is then available for you to analyse however you wish. (If you retain all of the $W_t$, you have a distribution of the trajectory of the system as well, which you can analyse.)
So:
your $x$ are a sampled distribution of your shocks, i.e. they are $\epsilon_t=0$.
To simulate the evolution of the $W_t$, you will need some initial condition $W_0$. What this is depends on what you're modelling. If you're modelling the likely values of a single stock starting at an initial price $W_0$, then your initial state is a 1000 element vector with constant value.
Now evaluate your equation, plugging in all your constants, $W_0$, and your initial shocks $\epsilon_0 = x$ to get the distribution of prices $W_1$.
Repeat: sample $x$ again -- this is now $\epsilon_1$. Plugging this in, gives you $W_2$ etc.
2. Coding the simulation (simple example)
One of the useful features of R is that most operators work element-wise over vectors.
So you can pretty much type in your equation more or less as it is.
I've made a few assumptions about the parameters in your equation, and I've ignored the $\pi$ function -- you can add that in later.
So you end up with code that looks something like this:
dt <- 0.5 # step-size
r <- 1 # parameters
lambda <- 1
sigma <- 1 # std deviation
w0 <- rep(1,1000) # presumed initial condition -- prices start at 1
# Show an example iteration -- incorporate into one line for production code...
x <- rnorm(1000,mean=0,sd=1) # random shock
w1 <- w0*exp(r*dt)*(1+exp((sigma*lambda-0.5*sigma^2)*dt +
sigma*x*sqrt(dt) -1)) # evolution
When you're ready to let the simulation run, then merge the last two lines, i.e. include the sampling statement in the evolution statement. You then get one line of code which you can run manually or embed into a loop, along with any other analysis you want to run.
# General simulation step
w <- w*exp(r*dt)*(1+exp((sigma*lambda-0.5*sigma^2)*dt +
sigma*rnorm(1000,mean=0,sd=1)*sqrt(dt) -1))
You can also easily visualise the changes and obtain summary statistics (5-number summary):
hist(w)
summary(w)
Of course, you'll still need to work through the details of what you actually want to model and how you want to go about analysing it --- and you've got the $\pi$ function to deal with --- but this should get you started toward using R for discrete simulation.

Cons of setting MaxNWts in R nnet to a very large number

I am using the nnet function package from the nnet package in R. I am trying to set the MaxNWts parameter and was wondering if there is any disadvantage to setting this number to a large value like 10^8 etc. The documentation says
"The maximum allowable number of weights. There is no intrinsic limit
in the code, but increasing MaxNWts will probably allow fits that are
very slow and time-consuming."
I also calculate the size parameter by the following calculation
size = Math.Sqrt(%No of Input Nodes% * %No of Output Nodes%)
The problem is that if I set "MaxNWts" to a value like 10000 , it fails sometimes because the number of coefficients is > 10000 when working with huge data sets.
EDIT
Is there a way to calculate the number of wts( to the get the same number calculated by R nnet) somehow so that I can explicitly set it every time without worrying about the failures?
Suggestions?
This is what I have seen in the Applied Predictive Modeling:
Kuhn M, Johnson K. Applied predictive modeling. 1st edition. New York: Springer. 2013.
for the MaxNWts= argument you can pass either one of:
5 * (ncol(predictors) + 1) + 5 + 1)
or
10 * (ncol(predictors) + 1) + 10 + 1
or
13 * (ncol(predictors) + 1) + 13 + 1
predictors is the matrix of your predictors
I think it is empirical based on your data, it is a regularization term like the idea behind shrinkage methods, ridge regression (L2) term for instance. It's main goal is to prevent the model from over fitting as is the case often with neural networks because of the over-parameterization of its computation.

Resources