# 6.2 Breakdown of a neural network

 Page 1 / 1
This module breaks down a standard neural network, describing the different parameters, hyper-parameters, and functions that are necessary for building a neural network.

## Cost functions

The different cost functions we explored using were the sigmoid function, rectified linear units (ReLU), and the softmax normalization function.

The sigmoid activation function is the most general nonlinear activation function used in neural networks. Intuition would naively suggest that the activation of a neuron would be well modeled by the step function, but the issue is its non-differentiability. The stochastic gradient descent algorithm requires that activation functions be differentiable. The solution would be to approximate the step function using a smooth function like the sigmoid or the hyperbolic tangent. The issue with the sigmoid function is that its derivative far from the origin is near zero, so if any individual weight on a neuron is very wrong, it is unable to use the gradient to adjust its value. As a result, outlier weights can significantly impact the performance of the network.

The advantage of using rectified linear units is threefold. First, its derivative is a constant (either 0 or 1) making the computation of the gradient much faster. Second, it is a better approximation of how biological neurons fire, in the sense that there is no activation in the absence of stimulation. Third, rectified linear units speed up learning by not being able to fire with zero net excitation. This means that if an excitation fails to overcome a neuron’s bias, the neuron will not fire at all. And when it does fire, the activation is linearly proportional to the excitation. The sigmoid function in comparison allows for some activation to occur with zero and even negative net excitation. However, a lower learning rate needs to be used with ReLU because its zero derivative for a net excitation less than zero means that the neuron effectively stops learning once its net excitation hits zero.

Softmax activation is particularly useful on the output layer, as it normalizes the output. Exponentiating each of the net excitations gives a more dramatic representation of the differences between them. Weak excitations become weaker activations and strong excitations become stronger activations. Everything is then normalized, giving the layer the effect of becoming a decision-maker.

## Activation functions

The different cost functions we explored using for the gradient descent learning algorithm were mean-squared error, cross-entropy, and log-likelihood.

Mean-squared error is the simplest measurement of difference that can be used to good effect in a neural network. It can be used with any activation function and is the more versatile option, though not always the most effective one. One of its shortcomings is that neurons with a sigmoid activation function become saturated quickly and are unable to learn more as a result of the relatively small magnitude of the sigmoid’s derivative far from the origin.

Cross-entropy treats the desired output as some probability distribution and the network’s output as another probability distribution, and measures the distance between the distributions. The main attraction to using cross-entropy is that when used in conjunction with the sigmoid activation function, its gradient is linearly dependent on the error, solving the issue with neurons becoming saturated quickly.

Log likelihood maximizes only the output neuron corresponding to which neuron should be firing. Used in conjunction with a softmax layer, all other output neurons would be minimized as a result of maximizing the desired output neuron. In this sense, a softmax layer has to be used, or the activations of the final layer will be too close together to draw meaningful conclusions.

Stochastic gradient descent is the algorithm used in our network to adjust weights and biases according to the evaluation of the gradient of a given cost function. The gradient determines whether a parameter should increase or decrease and by how much. The learning rate of a network is a constant associated with how much a parameter should travel down its gradient at each reevaluation. In the original algorithm, parameters are updated after each given input. A common practice with neural nets is to only reevaluate the gradient after a so-called minibatch of multiple inputs is passed. This way, the cost function has multiple samples and can better construct a curve, yet the gradient is somewhat different every time it’s evaluated. This introduces some noise into the gradient to make it harder for parameters to get stuck in a local minimum of the gradient.

## Dropout

Overfitting is an issue experienced in networks when neurons are trained to identify specific images in a training set rather than the more general concept that an image represents. Instead of recognizing a 7, the network may only recognize the particular 7s that were in the training data set. To prevent this, we implemented dropout in our network. Random neurons in our interconnected layers were turned off between mini-batches, meaning that certain weights were not able to be used in determining an output. This essentially means that we were training a slightly different network each mini-batch, encouraging more neurons to learn meaningfully, as weights will typically be more fairly distributed. In evaluating the network, all neurons are turned back on and their weights are scaled down by the dropout rate. As a result, neurons are less strongly associated with particular images, and more applicable to a more expansive set of images.

what does nano mean?
nano basically means 10^(-9). nanometer is a unit to measure length.
Bharti
do you think it's worthwhile in the long term to study the effects and possibilities of nanotechnology on viral treatment?
absolutely yes
Daniel
how to know photocatalytic properties of tio2 nanoparticles...what to do now
it is a goid question and i want to know the answer as well
Maciej
Abigail
for teaching engĺish at school how nano technology help us
Anassong
Do somebody tell me a best nano engineering book for beginners?
what is fullerene does it is used to make bukky balls
are you nano engineer ?
s.
fullerene is a bucky ball aka Carbon 60 molecule. It was name by the architect Fuller. He design the geodesic dome. it resembles a soccer ball.
Tarell
what is the actual application of fullerenes nowadays?
Damian
That is a great question Damian. best way to answer that question is to Google it. there are hundreds of applications for buck minister fullerenes, from medical to aerospace. you can also find plenty of research papers that will give you great detail on the potential applications of fullerenes.
Tarell
what is the Synthesis, properties,and applications of carbon nano chemistry
Mostly, they use nano carbon for electronics and for materials to be strengthened.
Virgil
is Bucky paper clear?
CYNTHIA
so some one know about replacing silicon atom with phosphorous in semiconductors device?
Yeah, it is a pain to say the least. You basically have to heat the substarte up to around 1000 degrees celcius then pass phosphene gas over top of it, which is explosive and toxic by the way, under very low pressure.
Harper
Do you know which machine is used to that process?
s.
how to fabricate graphene ink ?
for screen printed electrodes ?
SUYASH
What is lattice structure?
of graphene you mean?
Ebrahim
or in general
Ebrahim
in general
s.
Graphene has a hexagonal structure
tahir
On having this app for quite a bit time, Haven't realised there's a chat room in it.
Cied
what is biological synthesis of nanoparticles
what's the easiest and fastest way to the synthesize AgNP?
China
Cied
types of nano material
I start with an easy one. carbon nanotubes woven into a long filament like a string
Porter
many many of nanotubes
Porter
what is the k.e before it land
Yasmin
what is the function of carbon nanotubes?
Cesar
I'm interested in nanotube
Uday
what is nanomaterials​ and their applications of sensors.
what is nano technology
what is system testing?
preparation of nanomaterial
how did you get the value of 2000N.What calculations are needed to arrive at it
Privacy Information Security Software Version 1.1a
Good
Got questions? Join the online conversation and get instant answers!