After seeing a number of “I Forced a Bot” tweets go viral, I wanted to see whether a computer program could actually execute something like this, and if so, what would it look like?
Inevitably, this led me to lots of readings on machine learning and neural networks. Drawing inspiration from human/animal brains, a neural network is a system that’s used to build a computer program that learns from data. Just as a brain’s neuron’s “fire” or activate other connected neurons in order to send messages together, a neural network is a collection of nodes that transmit information from one to another.
brain.js is a library of neural networks, written in JavaScript. After reading the documentation on GitHub, I found creating your own (very basic) neural network can be relatively easy, thanks to brain.js.
Objective: Teach a program to learn something and distinguish elements or create assumptions, based on data.
With brain.js, depending on how complex and abstract you want to go, you can teach a program to learn anything from identifying color contrast (the READ.me uses this as an example), write a children’s book, recognize a face, or in this case, distinguish the difference between a Donald Trump and Kim Kardashian tweet.
3 Main Components to a Basic Neural Network:
1. Brain.js file
Luckily, the library is well-documented and you can import the file directly into your program. The brain.js file will hold all of the logic that will instantiate a neural network.
2. Index.js file
With a few functions added(I only included four), the code in this file will execute the neural network and take in any new information a user feeds it. A few key functions:
encode( )
function encode(arg) {
return arg.split('').map(x => (x.charCodeAt(0) / 255));
}In order for brain.js to work with the data we give it (and train it on), the values must be between 0 and 1. Since I worked with strings, and ultimately would input a string into the program, I had to find a way to convert the strings into values between 0 and 1. After converting every single character of each string into ASCII code and dividing by 255(number of characters in extended ASCII), we have an array for each string that looks like the below.
train(data)
After encoding each string of our training data (more on that soon), the train function creates a new neural network and trains the neural network on the encoded data provided.
3. Training-data.js
Lastly, and most importantly, our training data. Since we’ve created a function that encodes our data into numeric values, we can store our training data as regular text (which I copy pasted into my file, instead of retrieving). I only used 8 examples each for Kim Kardashian and Donald Trump in my file, however, the more training data you include, the more accurate your neural network.
With that a (super) basic neural network is set up!
To execute and see it run in console, I just create an execute function, and call it with a new tweet the program has never seen before(Kim Kardashian’s) as an argument.
function execute(input) {
let results = trainedNet(encode(input));
let output;
console.log(results)
results.trump > results.kardashian ? output = 'Trump' : output = 'Kardashian';
return output;
console.log(output)
}
train(trainingData);
console.log(execute("Beyond vibing #KUWTK"));- results is what the neural net returns once it processes the encoded input and determines a number between 0 and 1 for the two choices (Kim K and Trump), which can be interpreted as a percentage of certainty for each.
- output is the key for the result which was calculated at the highest certainty
via console, you can see that not only did the program correctly guess that the tweet string was Kim Kardashian’s, but it also was able to do so with 57% certainty.
Using a Trump tweet below,
“Whether we are Republican or Democrat, we must now focus on strengthening Background Checks!”
the program was able to guess with 87% certainty that the tweet was authored by Trump.
Another example:
While I didn’t have time to fully write a bot that can write a script for an episode of Jerry Springer(yet), this exercise and brain.js showed me the possibilities of training programs and the exciting things you can do with celebrity tweets!