What is this session about?

Eyes of machines.

Eyes of Machine?

Yes! 

Today we'll learn how to teach machine to see.

So how do we see?

We see with our eyes right?

But how do our eyes work?

This is a million-dollar question!

Eye is a machine.

One of the most efficient machine in our body.

 

Somethings I'll not tell you.

• The Cambrian explosion
• The evolution of human eye
• The interesting debate of vision v/s language.

Let me form an analogy.

• Everything you've ever seen is training data.
• Everything that you've ever inferred is a target output
• Since eye is a machine which only shuts down when you sleep. 

Eye is machine that learns over time

Since eye are machine, vision is a learning process.

 

What I am saying is that with the same learning process, you can make your tongue see.

Yes, this is true.

 

Coming back to machines and learning.

 

Do we know the difference between machine learning and deep learning?

Let's start the machine learning fun.

For any machine learning algorithm to work, we need following:

• Data

• Hypothesis Function

• Loss Function

• Optimizer

Linear machine

Logistic regression.

Can someone relate how this is wrong for vision in machine?

So what do we lack in linear machines?

Since linear machines are only capable of interpreting one dimension of knowledge it hinders the visual sense of machine.

Non-Linear machine

 Neural Network in Keras

So now we are able to detect multiple dimension in our learning and use them efficiently.

Still, something is missing!

Convolution Neural Network

in Keras

The holy grail of computer vision

What changed?

• Parameter sharing - Spatial robustness / Faster training

 

• Multiple layers of kernels - multiple features and dimension learning

 

• Translation invariance due to kernel of different size.

Let's dive into code.

http://bit.ly/gdgatlan

Fin.