Scribd
Upload
47 views

Batch Normalisation

Batch normalization is a technique for improving neural network training by normalizing layer activations. It standardizes the layer's inputs to have zero mean and unit variance. This helps address the problem of internal covariate shift, where changes in previous layers disrupt learning. Batch normalization allows lower layers to perform nonlinear transformations while stabilizing learning. It can be applied to any hidden layer and includes learned scale and shift parameters to preserve expressiveness.

Uploaded by

4H5 VISHESH PETWAL
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
47 views

Batch Normalisation

Batch normalization is a technique for improving neural network training by normalizing layer activations. It standardizes the layer's inputs to have zero mean and unit variance. This helps address the problem of internal covariate shift, where changes in previous layers disrupt learning. Batch normalization allows lower layers to perform nonlinear transformations while stabilizing learning. It can be applied to any hidden layer and includes learned scale and shift parameters to preserve expressiveness.

Uploaded by

4H5 VISHESH PETWAL
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 17

BATCH

NORMALIZATION
INTRODUCTION

Not an optimization algorithm

A method of adaptive reparameterization

Motivated by the difficulty of training very


deep models
COMPOSITION OF SEVERAL FUNCTIONS OR LAYERS
DEEP MODELS
THE GRADIENT TELLS HOW TO UPDATE EACH PARAMETER,
UNDER THE ASSUMPTION THAT THE OTHER LAYERS DO NOT
CHANGE.

IN PRACTICE, WE UPDATE ALL OF THE LAYERS


SIMULTANEOUSLY.

WHEN WE MAKE THE UPDATE, UNEXPECTED RESULTS CAN


HAPPEN BECAUSE MANY FUNCTIONS COMPOSED TOGETHER
ARE CHANGED SIMULTANEOUSLY, USING UPDATES THAT
WERE COMPUTED UNDER THE ASSUMPTION THAT THE
OTHER FUNCTIONS REMAIN CONSTANT.
EXAMPLE
SUPPOSE WE HAVE A DEEP NEURAL NETWORK THAT HAS ONLY ONE UNIT PER
LAYER AND DOES NOT USE AN ACTIVATION FUNCTION AT EACH HIDDEN LAYER: ˆ
Y = XW1W2 W3 . . . WL.

Wi PROVIDES THE WEIGHT USED BY LAYER I.

THE OUTPUT OF LAYER I IS THE OUTPUT ˆ Y IS A LINEAR FUNCTION OF THE INPUT


X, BUT A NONLINEAR FUNCTION OF THE WEIGHTS Wi

SUPPOSE OUR COST FUNCTION HAS PUT A GRADIENT OF 1 ON Yˆ, SO WE WISH TO


DECREASE ˆ Y SLIGHTLY.

THE BACK-PROPAGATION ALGORITHM CAN THEN COMPUTE A GRADIENT

CONSIDER WHAT HAPPENS WHEN WE MAKE AN UPDATE


FIRST-ORDER TAYLOR SERIES
APPROXIMATION
ˆ Y PREDICTS THAT THE VALUE OF ˆ Y WILL DECREASE BY
EGTG.

IF WE WANTED TO DECREASE ˆ Y BY 0 .1, THIS FIRST-ORDER


INFORMATION AVAILABLE IN THE GRADIENT SUGGESTS WE
COULD SET THE LEARNING RATE

HOWEVER, THE ACTUAL UPDATE WILL INCLUDE


SECOND-ORDER AND THIRD-ORDER EFFECTS, ON UP TO
EFFECTS OF ORDER L.

THE NEW VALUE OF ˆY IS GIVEN BY


EXAMPLE OF ONE SECOND-ORDER
An example of one second-order term arising from this update is

This term might be negligible if is small, or might be exponentially large if the weights
on layers 3 through ’ l’ are greater than 1
This makes it very hard to choose an appropriate learning rate, because the effects of an update to the
parameters for one layer depends so strongly on all of the other layers.
Second-order optimization algorithms address this issue by computing an update that takes these second-order
interactions into account, but we can see that in very deep networks, even higher-order interactions can be
significant.
Even second-order optimization algorithms are expensive and usually require numerous approximations that
prevent them from truly accounting for all significant second-order interactions.

Building an n-th order optimization algorithm for n > 2 thus seems hopeless.
Batch normalization provides an elegant way of
reparametrizing almost any deep network.

WHAT CAN The reparametrization significantly reduces the problem


of coordinating updates across many layers.
WE DO Batch normalization can be applied to any input or hidden
INSTEAD? layer in a network.

Let H be a minibatch of activations of the layer to normalize,


arranged as a design matrix, with the activations for each example
appearing in a row of the matrix. To normalize H, we replace it with
WHAT CAN WE DO
INSTEAD?
where μ is a vector containing the mean of each unit and σ is a
vector containing the standard deviation of each unit.

The arithmetic here is based on broadcasting the vector μ and the


vector σ to be applied to every row of the matrix H .

Within each row, the arithmetic is element-wise, so Hi,j is


normalized by subtracting μj and dividing by σj . where δ is a small positive value such
as 10−8 imposed to avoid encountering
The rest of the network then operates on H0 in exactly the same the
way that the original network operated on H.
undefined gradient of z at z = 0.
Crucially, we back-propagate through these operations for computing the mean and the standard deviation, and
for applying them to normalize H.

This means that the gradient will never propose an operation that acts simply to increase the standard deviation or mean of hi;
the normalization operations remove the effect of such an action and zero out its comthe gradient.ponent in

This was a major innovation of the batch normalization approach. Previous approaches had involved adding penalties to the
cost function to encourage units to have normalized activation statistics or involved intervening to renormalize unit statistics
after each gradient descent step.

The former approach usually resulted in imperfect normalization and the latter usually resulted in significant wasted time as the
learning algorithm repeatedly proposed changing the mean and variance and the normalization step repeatedly undid this
change.

Batch normalization reparametrizes the model to make some units always be standardized by definition, deftly sidestepping
both problems.
At test time, μ and σ may be replaced
by running averages that were
collected

during training time. This allows the


model to be evaluated on a single
example,

without needing to use definitions of


μ and σ that depend on an entire
minibatch
Revisiting the ˆ y = xw1 w2 . . . wl example, we see that we can
mostly resolve the difficulties in learning this model by
normalizing hl−1. Suppose that x is drawn from a unit Gaussian.

Then hl−1 will also come from a Gaussian, because the


transformation from x to hl is linear.

However, hl−1 will no longer have zero mean and unit variance.

After applying batch normalization, we obtain the normalizedˆ hl−1


that restores the zero mean and unit variance properties.

For almost any update to the lower layers, ˆhl−1 will remain a unit
Gaussian.
The output ˆ y may then be learned as a simple linear function ˆ y = wlˆh l−1. Learning in
this model is now very simple because the parameters at the lower layers simply do not
have an effect in most cases; their output is always renormalized to a unit Gaussian.

In some corner cases, the lower layers can have an effect.

Changing one of the lower layer weights to 0 can make the output become degenerate,
and changing the sign of one of the lower weights can flip the relationship between ˆ hl−1
and y.

These situations are very rare. Without normalization, nearly every update would have an
extreme effect on the statistics of hl−1.

Batch normalization has thus made this model significantly easier to learn.

In this example, the ease of learning of course came at the cost of making the lower layers
useless.
the lower layers no longer have any harmful effect, but they also
no longer have any beneficial effect.

This is because we have normalized out the first and second order
statistics, which is all that a linear network can influence.
LINEAR
EXAMPLE In a deep neural network with nonlinear activation functions, the
lower layers can perform nonlinear transformations of the data, so
they remain useful.

Batch normalization acts to standardize only the mean and


variance of each unit in order to stabilize learning, but allows the
relationships between units and the nonlinear statistics of a
single unit to change.
Because the final layer of the network is able
to learn a linear transformation, we may
actually wish to remove all linear
relationships between units within a layer.

Unfortunately, eliminating all linear


interactions is much more expensive than
standardizing the mean and standard
deviation of each individual unit, and so far
batch normalization remains the most
practical approach.
Normalizing the mean and standard deviation of
a unit can reduce the expressive power of the
neural network containing that unit.

In order to maintain the expressive power of the


network, it is common to replace the batch of
hidden unit activations H with γH0 + β rather
than simply the normalized H0.

The variables γ and β are learned parameters


that allow the new variable to have any mean
and standard deviation.
At first glance, this may seem useless—why did we set the mean to , and then
introduce a parameter that allows it to be set back to any arbitrary value β?

The answer is that the new parametrization can represent the same family of
functions of the input as the old parametrization, but the new parametrization has
different learning dynamics. In the old parametrization, the mean of H was
determined by a complicated interaction between the parameters in the layers
below H.

In the new parametrization, the mean of γH0 + β is determined solely by β .

The new parametrization is much easier to learn with gradient descent.


Most neural network layers
It is natural to wonder whether
take the form of φ(XW + b)
we should apply batch More specifically, XW + b
where φ is some fixed
normalization to the input X, or should be replaced by a
nonlinear activation function
to the transformed value XW normalized version of XW .
such as the rectified linear
+b .
transformation.

The bias term should be


The input to a layer is usually The statistics of the input are
omitted because it becomes
the output of a nonlinear thus more non-Gaussian and
redundant with the β
activation function such as less amenable to
parameter applied by the
the rectified linear function in standardization by linear
batch normalization
a previous layer. operations.
reparametrization.

You might also like