Feedforward Neural Networks
Edit: I came across this resource, and it does an excellent job of explaining machine learning: http://www.r2d3.us/visual-intro-to-machine-learning-part-1/
Artificial learning is all the rage these days, so I decided to jump on the bandwagon.
I’ve started with learning about the most basic type of neural network – a feedforward. There’s probably some gaps in my knowledge, so feel free to chime in:
What is a feedforward neural network?
A feedforward neural network was the first and simplest type of neural network. It does not form any type of cycle and only moves in one direction.
How it works
An input is passed to a layer of “weights” which takes the input and performs calculations on it. If the output of the calculations is >0, then it passes it along to the next layer until you reach a final output.
That’s a little confusing, so let’s use an example. Let’s say you have one data point called X. That data point will be passed onto another layer of 3 neurons. Those neurons will take the input, perform a simple learning algorithm, and pass it to the next layer until a final output is reached. If the final output is >0, then your algorithm is working well against your dataset.
How does it get smarter?
You might be asking, how is this a neural network? It just seems like a bunch of equations.
These networks are trained by using a learning algorithm, which creates adjustments to the weights based on final output. If the sum of the final product is <0, the neural network will adjust until the sum of the calculations are >1.
The most common learning technique
There are multiple leaning techniques you can use, but the most common is back-propagation. Here’s how it works:
– The final output of your neural network is compared with a correct answer to calculate the value of an error-function
– The error is then fed back through the system and the algorithm adjusts the weights of each connection to reduce the value of the error function
– After repeating this process across a large number of training cycles, the network will typically converge to a state where the error of calculations is small
This type of learning has its shortcomings. If you only have a few samples for you to train your neural network against, your system might overgeneralize based on limited data.