NumberVision – Machine Learning Application

-
The next way we used our Neuron class was to classify digits. Using a UI, our program was assigned to take an input digit (a 5x7 grid, each box has a shade of grey between black and white) and classify the digit as a number between 0 and 9.

A Smarter Neuron

Graphical depiction of a Neuron receiving more than 2 inputs.
Previously, I had only been using a Neuron with two inputs: an X and Y coordinate. Previously, this allowed me to plot my points on a plane like I did with the Half Moon project. However, due to the 5x7 grid of inputs I was getting, I had to alter my Neuron class to be able to take in 35 inputs. Mostly, this involved changing a few numbers. 


Applying the Neuron to NumberVision

The digit '1', with progressively more noise added from left to right.
Good algorithms should be able to generate a result even if the input is a little distorted. I applied the same goal to NumberVision. The flow chart above shows various levels of added distortion, or noise. At no extra noise, the digit is easily recognizable. At the fifth iteration of noise, even a human would have a tough time identifying the number.

In order for my Neuron to be able to recognize a number, even with noise in the input, I trained it with a noise-y set of inputs. Each digit had 30 variations, totalling 300 inputs.

However, it's impossible to train a single neuron to recognize any of ten digits, as each Neuron can only have one of two outputs: yes or no. So, since there are 10 digits, I created 10 Neurons, each trained on a different digit. Then, to recognize a digit, the program runs the input through each of the Neurons. If trained correctly, only one Neuron will trigger, and there you have your digit. 

Comments

Post a Comment

Popular updates

Half Moon Project Part 2 - More Accuracy

-
Image
This time around, we used the Neuroph library to create a neural network  that can divide the double half-moons more accurately. Neural Networks A Neural Network is a combination of Neurons that feed into each other. Using some really complex math, they can weight each input (from a previous Neuron) differently. Then, after training, they can use backpropagation to calculate which weights should be edited. What this looks like can be seen below. More Accuracy for Half-Moons Since, previously, we were only using one neuron to calculate the line to separate the two moons, we could only draw a linear line. This time, with a more complex neural network, it can separate the two with a function that's super complicated. Theoretically, a human could also derive the equation, but it's so complex that at this point, it's easier to get the Neural Network to do it for us.
Read more

Machine Learning & Half Moon Project

-
Image
This was our first Machine Learning assignment. Machine Learning is something that I was (and still am) super excited about. Basically, the algorithm 'learns' something, and can then apply it in other situations. Learning & Neurons This was the toughest concept to learn (ha). We started with a Neuron, similar to the neurons in our brains. The neuron can trigger (yes/no) based on a series of inputs, in our case two inputs. The equation for whether or not the neuron triggers is SUMMATION(Weight(n) * Input(n)) > threshold . Each input is weighted differently, and the weights are what we can change to 'train' the neuron. We use this equation:  ΔWeight = learningSpeed * (target - output) * input . For example, if we start with a speed of 0.2, an input of 1, a weight of 0.5, and a target of 0, the weight would change by -0.2 (0.2 * (1 - 0) * 1). Now, with a weight of 0.3, repeat the process. Once each input returns the desired output, the best weights have been fo
Read more