GO Ahead. Compose: Understanding AI machine learning in 3 minutes #MGAC2018 #YourFutureYourCall

Recall that a quadratic function is:

f(x)= ax^2+bx+c

The architecture of a deep learning neural network is actually quite similar. As an example, we would like to train our system to and differentiate between an image of a cat and a dog.

We wish our system to be able to output

f(image of a cat) = cat

f(image of a dog) = dog

as we feed an image to the system.

But how? How would our system learn? How do we teach the system?

In essence, we need a specific method of quantifying the performance of our system, i.e. give the system a score for its performance.

For example, when being fed an image of a dog, the system might output a 40% chance of it being a dog and a 60% chance of it being a cat.

Thus the score shall be 60, as we know the ideal output is 100% dog. The idea is to train our system such that score gradually decreases until it reaches 0, for both dog and cat images being fed to the system.

Concerning the diagram, each circle represents a neuron, which has a value of 0 to 1. The strength of the interlayer connection between each neuron represents the weight of the system. 

As the performance is quantified, we nudge the value of the weights (a,b and c) by a minuscule amount to improve the performance (this involves finding the derivative and global minimum in a multi-variable calculus sense).

This process is repeated with new data. Thus this is, in essence, the machine "learning", i.e. gradually adjusting the weights towards the desired result.

#MGAC2018 #YourFutureYourCall

Citations

Cs231n.github.io. (2018). CS231n Convolutional Neural Networks for Visual Recognition. [online] Available at: https://cs231n.github.io/neural-networks-1/ [Accessed 20 Jul. 2018].