Welcome. Today, we are going to cover how to build a basic single perceptron neural network.
I want to make this the first of a series of articles where we delve deep into everything – CNNs, transfer learning, etc. so be sure to bookmark the site and keep checking it. Also, this will include some math, so strap in.
What is Perceptron?
It is the most fundamental unit of a neural network (NN). In a complex NN, all data pass through several of these, often at the same time and performing different roles. But we’ll get to that later.
For now, what is a perceptron?
In supervised learning, a perceptron is a form of a linear classifier. So, any dataset that can be divided by a single straight line can be classified using a perceptron, i.e., any dataset that looks like this:
The last one cannot be considered linearly separable, because it is not a single line that can approximately separate the database.
Building A Single Perceptron Neural Network
Let’s move on to building our first single perceptron neural network today. For this, we’ll begin with creating the data.
1. Create our dataset
First, we need our data set, which in our case will a 2D array. Open up your code editors, Jupyter notebook, or Google Colab.
import pandas as pd import numpy as np import random
Let’s make our data. I consider a 20*20 plane in this example to keep our results small and concise.
#Dataset df = pd.DataFrame() df['x'] = [random.randint(1,20) for x in range(10)] df['y'] = [random.randint(1,20) for x in range(10)] df.head()
Now, we need to label these. So we’ll filter out based on a line (I considered y=x). So every point below the line is y<x and every point above the line is mean y>x.
label= for i in range(df.shape): if df.iloc[i,0] < df.iloc[i,1]: label.append(1) else: label.append(-1) df['label'] = label df
2. Initialize weights for the values
Now we can initialize the weights. We can’t use zero value, so we’ll just go with a random uniform distribution for weights:
#weights and bias weights = [np.round(random.uniform(-0.99,0.99),2) for i in range(2)]
Then we multiply the weights with the input data points and sum:
w = weights.copy() X = [[df.iloc[i,0],df.iloc[i,1]] for i in range(df.shape)] wx = [X[i]*w+X[i]*w for i in range(df.shape)]
Now, we have the sum of weights and inputs for each point.
So what the procedure is, is that we’ll plug in all these values one at a time into the activation function, and then based on output we’ll modify the weights.
3. Creating the Activation Function
Now we come to the activation function. The perceptron treats the sum and gives us a label, which we compare with the original label and determine if it is correct. If it is incorrect, the error is found and the weights are adjusted so that our solution *moves* in the direction of the original.
We’ll be using the signum function:
If wx <=0 , then output is 0. Else, the output is 1.
for i in range(df.shape): if wx[i]<=0: pred = 0 else: pred = 1
Remember, we only have two inputs: x and y. Not the whole dataframe. So, we will expand this activation function to take in one datapoint at a time, and then finding the error and then adjusting the error:
for i in range(df.shape): if wx[i]<=0: pred = -1 else: pred = 1 if pred != df['label'][i] : err = df['label'][i] - pred w = w + err w = w + err
This works perfectly now. To clearly see the outputs, we’ll put in a bunch of print statements:
for i in range(df.shape): print('wx : ',wx[i]) if wx[i]<=0: pred = -1 else: pred = 1 print('label=',df['label'][i]) print('pred = ',pred) if pred != df['label'][i] : err = df['label'][i] - pred print('err',err) print('before', w,w) w = w + err w = w + err print('after',w,w) else: print('w_i', w,w)
And now if we run this:
A simple print formatting statement gives us the final weights.
4. Testing our model on another database
Similar to how we did the train-test-split, we’ll use a different database for our testing here.
#Test Dataset new_df = pd.DataFrame() new_df['x'] = [random.randint(1,20) for x in range(100)] new_df['y'] = [random.randint(1,20) for x in range(100)] new_df.head()
Then we generate the labels based on y=x line, and the sum of weights*inputs:
label_text =  for i in range(new_df.shape): if new_df.iloc[i,0] < new_df.iloc[i,1]: label_text.append(1) else: label_text.append(-1) new_wX = w*new_df['x']+w*new_df['y']
Here is mine:
So, now the moment of truth, we apply our activation function, and then we can compare the given labels with the predicted labels:
new_df['given label'] = label_text pred_label_text =  for i in range(new_df.shape): if new_wX[i]>=0: pred_label_text.append(-1) else: pred_label_text.append(1) new_df['predicted labels'] = pred_label_text
As you can see, we did pretty well for ourselves 🙂
Congratulations on completing this tutorial. I hope this gave you a lot of insight into a “perceptron”. Stay in touch with us to read our future tutorials.