Machine learning differs from Traditional programming in a different way. When we talk about Traditional Programming, we usually have a set of rules/formulas to which we provide the data and get the answers from it.
Whereas in Machine learning, we train the computer to figure out the rules given a set of data and the answers. ML algorithms figure out the specific patterns in the dataset and try to build a set of rules around it.
Following are few great resources which explain ML and AL in detail
- Visual intro to machine learning
- Machine Learnings for Humans
- But what is a Neural Network? - By 3Blue1Brown
Let's look at a simple example to illustrate this with Neural Networks(NN). A neural network is a slightly more advanced implementation of machine learning which we call deep learning.
Understanding the problem
Like every first app, we start with something super simple that shows the overall scaffolding for how the code works. For neural networks, we can take an example where we ask the network to learn the relationship between two numbers.
const getYforX(x) {
const y = 3*x - 2;
return y;
}
// x -> input
// y -> answer
// rule -> 3x - 2
Let's program this NN to learn this rule of y=3x-2 by providing it some input and the output values.
We can divide this task into following steps:
- Defining and compiling the Neural Network
- Providing the data
- Training the network
- Prediction
Step 1: Defining and compiling the Neural Network
Since the input and output are scalars, we can define the network with a single layer having a single neuron.
const tf = require('@tensorflow/tfjs-node');
const model = tf.sequential({
layers: [tf.layers.dense({ units: 1, inputShape: [1] })]
});To compile this NN, we need to provide 2 functions. A loss function and an optimizer.
The NN can start with making a guess which could look like y=10x+4. The loss function measures the guessed answers against the known correct answers. It then tells us how well or how bad the assumption is.
It then uses the optimizer function to make another guess. Based on how the loss function went, it will try to minimize the loss.
It may endup with y=5x+2. This is still bad, but better than the previous answer.
We repeat this process for a number of EPOCHS which slowly brings us to the actual result minimizing the loss.
model.compile({ optimizer: 'sgd', loss: 'meanSquaredError' });The loss function is Mean Squared Error and the optimizer is Stochastic gradient descent
Step 2: Providing the data
Next, we feed in some data to the NN. As you can see below, the relationship between x and y is y=3x-2.
Tensorflow, as the name suggests, consumes tensors which are generalizations of matrices to N-dimensional space.
const x = tf.tensor([-1, 0, 1, 2, 3, 4, 5, 6, 7, 8]);
const y = tf.tensor([-5, -2, 1, 4, 7, 10, 13, 16, 19, 22]);Step 3: Training the network
Next up is the training process. The NN learns the relationship between input and output variables in the model.fit method.
We train the network for 600 epochs to give it sufficient time to learn the relationship, minimizing the loss.
model.fit(x, y, {epochs: 600}});Step 4: Prediction
Once the model completes the training task, the relationship between the input and output is encoded in the neurons.
We can use the model.predict function to give us the value of a new x which the network hasn't seen yet.
model.predict(tf.tensor([10])).print();
// outputs: 27.9856529
According to the formula, the correct answer should have been 28 (3*10 - 2). But, the answer is a little less.
This is because neural networks deal with probabilities. So, once we feed the NN with the data, it calculated that there is a very high
probability that the relationship between x and y is Y=3X-2, but with only 10 data points we can't know for sure.
Hence, the result for 10 is very close to 28, but not necessarily 28.
This is the comple code for a full glance.
const tf = require('@tensorflow/tfjs-node');
const model = tf.sequential({
layers: [tf.layers.dense({ units: 1, inputShape: [1] })],
});
model.compile({ optimizer: 'sgd', loss: 'meanSquaredError' });
const x = tf.tensor([-1, 0, 1, 2, 3, 4, 5, 6, 7, 8]);
const y = tf.tensor([-5, -2, 1, 4, 7, 10, 13, 16, 19, 22]); //y=3x-2
model
.fit(x, y, {epochs: 600}})
.then(history => {
model.predict(tf.tensor([10])).print();
});These are just ~10 lines but, the sheer magic that runs behind this is immense.
Interactive Demo
Follwing is a demo depicting what we have covered so far.
The graph below plots y=3x-2. The red dots are the known values and the blue dot is the value of y for x=10 which is currently being predicted by our NN.
Once you click on Train Model button, it will train the NN in your browser and update the prediction with every passing epoch. The code for this demo is available here.
Go ahead, give it a try!