Classification Methods Based on Neural Networks

Loading...

From the course by The State University of New York

Big Data, Genes, and Medicine

90 ratings

The State University of New York

Big Data, Genes, and Medicine

90 ratings

This course distills for you expert knowledge and skills mastered by professionals in Health Big Data Science and Bioinformatics. You will learn exciting facts about the human body biology and chemistry, genetics, and medicine that will be intertwined with the science of Big Data and skills to harness the avalanche of data openly available at your fingertips and which we are just starting to make sense of. We’ll investigate the different steps required to master Big Data analytics on real datasets, including Next Generation Sequencing data, in a healthcare and biological context, from preparing data for analysis to completing the analysis, interpreting the results, visualizing them, and sharing the results. Needless to say, when you master these high-demand skills, you will be well positioned to apply for or move to positions in biomedical data analytics and bioinformatics. No matter what your skill levels are in biomedical or technical areas, you will gain highly valuable new or sharpened skills that will make you stand-out as a professional and want to dive even deeper in biomedical Big Data. It is my hope that this course will spark your interest in the vast possibilities offered by publicly available Big Data to better understand, prevent, and treat diseases.

From the lesson

Predicting Diseases from Genes

After this module, you will be able to 1. Build classification and prediction models. 2. Evaluate the performance of classification and prediction methods. 3. Write R scripts to classify and predict diseases from gene expressions.

Overview of Classification and Prediction Methods8:36

Classification Methods Based on Analogy12:11

Classification Methods Based on Rules13:03

Classification Methods Based on Neural Networks7:26

Classification Methods Based on Statistics3:54

Classification Methods Based on Probabilities7:59

Prediction Methods4:26

Meet the Instructors

Isabelle Bichindaritz
Associate Professor
Computer Science

Very important other set of specification methods is based on your neural network.

Neural networks are based on the brain metaphor for information processing.

It's also referred to as Neural computing, or

Artificial neural networks, and more recently, deep learning.

There are many uses for artificial neural network.

Main one is in pattern with cognition, for example on images and

videos and so forth, they give very good results.

But they can also be used for any kind of prediction and classification task,.

So that's here you see at the bottom of this slide here you see some real neurons.

As you know the neurons are the cells that are present in the brain and

nervous system.

So these are specialized cells and they have some

communication between them as you see here some channels like Axons and Dendrites.

And they communicate through what's called synapses that

release signals that go from neuron to neuron.

And so based on this network model

an artificial neuron is represented here.

This is at the basis of the architecture pretty much of a neuron network.

So a single neuron has inputs and outputs as we can see here, so

the inputs are a set of features.

Attribute features, variables, however we call them, columns, that

we have in the data for a particular instance.

And then to each input is associated a weight.

And then the neuron unit, what it does is calculate a weighted sum,

so it adds all the x's multiplied by the weights.

And based on this total, pretty much,

input uses what's called a transfer function to calculate the input.

Those inputs will be Y1, Y2 or Yn.

So as you see this is a model based

on the actual real natural neural model.

Now there are many different types of architectures for

neural network and the architecture will be based on the task to address.

Because it can be applied to classification.

Prediction tasks but also clustering and optimization and so forth.

The most popular architecture is called the Feedforward, multilayered perceptron

using what's called the backpropagation learning algorithm.

It can be used for both classification and prediction by the problems.

But there are other ones for example, self-organizing feature maps will

be used for clustering tasks like the one's we're going to see in module six.

So that an example of architecture of a feature word

Multi layer per set electron that has one hidden layer.

So, here you said only connecting the input with the output.

For example, here we want to predict with a yes or no, is a prediction task.

So, it's a typical classification task.

We see here a hidden layer, so yes no, could be for

example, particular diagnosis.

Yes, this is breast cancer.

No this is not breast cancer, for example.

So, here you have one hidden layer and the hidden layers are very important.

At some point, research on your networks was tough,

because, the hidden layer was not existing.

And the type of problems that could be solved with this type of

model was too simple.

It could only classify, well,

problems where the data were what we said linearly separable.

And so this was very limiting and once hidden layers were introduced now

the range of applications and complexity

is just incomparable with what it was before.

And you may have several hidden layers.

So here, again, it's going to be a choice in the architecture and

fine tuning the perfect architecture for your particular neural network.

So the neural network, we learned the prediction model and

here what is not actually, if I go back to this weights, it's a set of weights.

That's what actually is a model is going to learn.

And once a weight has been learned and weight satisfaction,

satisfying performance then this is when we are ready for deployment of the model.

So that's really what the learning is about, it's about adjusting the weights.

So each time, based on the input data that we have in our training set,

we'll calculate the output based on the current weights.

And is designed output is a cheap and we would

stop learning it means it's particular weights have been learned satisfactorily.

Also why is going to adjust the weight and

each weight until it finds that the weight are learned.

Once trained, the neural network stops learning and then can be deployed.

Deep learning is based on advanced neural networks,

it's going to use more advanced, particularly, mapping functions.

So they are excellent for, in particular, pattern recognition and

they are one of the major methods used in machine learning.

However, they require important computing power and

have a set of parameters to adjust.

And also one limitation is that say often are not match understandable.

And we've seen comparison with some other models

that some considers this as a drawback.

So it really depends on what kind of task you're trying to carry out.

Thank you.

Explore our Catalog

Join for free and get personalized recommendations, updates and offers.

Coursera provides universal access to the world’s best education, partnering with top universities and organizations to offer courses online.

© 2018 Coursera Inc. All rights reserved.

Download on the App Store

Get it on Google Play