Linear Regression Using a Neural Network

Share on facebook
Share on twitter
Share on linkedin
Share on telegram
Share on whatsapp


This article was published as part of the Data Science Blogathon


In this article, we will answer these basic questions and build a basic neural network to perform linear regression.

What is a neural network?

The basic unit of the brain is known as a neuron, there are approximately 86 billion neurons in our nervous system that are connected to 10 ^ 14-10 ^ 15 synapse. meEach neuron receives a signal from the synapses and outputs it after processing the signal.. This idea is extracted from the brain to build a neural network.

Each neuron performs a scalar product between the inputs and the weights, add biases, applies a trigger function and outputs the outputs. When a large number of neurons are present together to give a large number of outputs, a neural layer is formed. Finally, multiple layers combine to form a neural network.

Arquitectura de red neuronal

Neural networks are formed when multiple neural layers combine with each other to give a network, or we can say that there are some layers whose outputs are inputs for other layers.

The most common type of layer to build a basic neural network is the fully connected layer, in which adjacent layers are completely paired and single-layer neurons are not connected to each other.


In the figure above, neural networks are used to classify data points into three categories.

Naming conventions. When the N-layer neural network, we do not count the input layer. Therefore, a single-layer neural network describes a network with no hidden layers (input is mapped directly to output). In the case of our code, we are going to use a single layer neural network, namely, we don't have a hidden layer.

Output layer. Unlike all layers in a neural network, neurons in the output layer commonly do not have a firing function (or you can think they have a linear identity activation function). This is because the last layer of output is usually taken to represent the class scores (for instance, in the classification), which are arbitrary real value numbers or some kind of real value target (for instance, in regression). Since we are doing the regression using a single layer, we don't have any activation function.

Neural network sizing. The two metrics that people commonly use to measure the size of neural networks are the number of neurons or, more commonly, the number of parameters.


We will use three basic libraries for this model, numpy, matplotlib and TensorFlow.

  • Numpy: this adds support for large, multidimensional arrays and arrays, along with a large collection of high-level math functions. In our case, we are going to generate data with the help of Numpy.
  • Matplotlib: this is a plotting library for python, we will visualize the final results using graphs in Matplotlib.
  • Tensorflow: this library has a particular focus on deep neural network training and inference. We can directly import the layers and train, test functions without having to write the whole program.
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf

Generating data

We can generate our own numerical data for this process using the np.unifrom function () that generates uniform data. Here, we are using two input variables xs and zs, adding some noise to randomize the data points and, Finally, the target variable is defined as y = 2 * xs-3 * zs + 5 + noise. The size of the data set is 1000.


After generating the data, save them to an .npz file, so they can be used for training.

np. know('TF_intro',input=generated_inputs,targets=generated_target)

Our goal is to get the final weights as close as possible to the actual weights, namely [2,-3].

Defining the model

Here, we will use the dense layer of TensorFlow to make the model and import the stochastic gradient descent from the Keras optimizer.

A gradient is the slope of a function. Measures the degree to which one variable changes with changes in another variable. Mathematically, gradient descent is a convex function whose output is the partial derivation of a set of parameters from its inputs. The higher the slope, the steeper the slope.

Starting from an initial value, Gradient Descent runs iteratively to find the optimal values ​​of the parameters to find the minimum possible value for the given cost function. The word “stochastic” refers to a random probability system or process. Therefore, en Stochastic Gradient Descent, some samples are randomly selected, instead of the dataset for each iteration.

Given the, the entrance has 2 variables, inlet size = 2 and output size = 1.

We set the learning rate at 0.02, which is neither too high nor too low, and the epoch value = 100.

models = tf.keras.Sequential([

Get weights and biases

We can print the predicted values ​​of weights and biases and also store them.

[array([[ 1.3665189],
        [-3.1609795]], dtype=float32), array([4.9344487], dtype=float32)]

Here, the first matrix represents the weights and the second matrix represents the biases. We can clearly see that the predicted values ​​of the weights are very close to the actual value of the weights..


Prediction and precision

After prediction using the given weights and biases, a final RMSE score of 0.02866, which is quite low.

RMSE is defined as the root mean square error. The root mean square error takes the difference for each observed and predicted value. The formula for the RMSE error is given as:



variable i



number of data points not missing

x_ {i}


real time series observations

hat {x} _ {i}


estimated time series

from sklearn.metrics import mean_squared_error
mean_squared_error(generated_target, out, squared=False)

If we plot the predicted data on a scatter plot, we get a graph like this:

plt.ylabel('Predicted Output')

Hurray! Our model is trained correctly with very few errors. This is the end of your first neural network. Note that each time we train the model we can obtain a different precision value, but they won't differ much.

Thank you for reading! You can contact me at [email protected]

The media shown in this article is not the property of DataPeaker and is used at the author's discretion.

Subscribe to our Newsletter

We will not send you SPAM mail. We hate it as much as you.