top of page

A Comprehensive Guide to Gradient Descent and Its Variants

Optimization algorithms are important tools utilized across various domains to either minimize or maximize objective functions. They find extensive applications in machine learning, deep learning, economics, engineering, and numerous other fields. These algorithms operate to discover the optimal solution for a given problem by iteratively adjusting parameters.

Gradient Descent Algorithm

Out of all the optimization algorithms, gradient descent is one of the most important. It's widely used in machine learning, especially for training neural networks. This algorithm is great at minimizing cost functions, which makes it useful for working with complex models.

Understanding how gradient descent works is crucial for professionals and researchers who rely on optimization techniques.

Gradient Descent

Gradient descent is a step-by-step process used to minimize a function by steadily moving toward the lowest point. Its main goal is to find the lowest value of a cost function, which measures the difference between predicted and actual values in a model.

Gradient descent works by figuring out the slope (or gradient) of the cost function concerning the model's parameters. Then, it adjusts these parameters in the opposite direction of the slope, aiming to reduce the cost function with each step.

The size of the adjustment is controlled by something called the learning rate, which determines how big the steps are toward the minimum.

This process repeats until a certain stopping point is reached, showing that the algorithm has effectively minimized the cost function.

Note: It's essential to choose an appropriate learning rate and convergence criterion to ensure that the algorithm converges efficiently without oscillating or getting stuck in a local minimum.

This process continues iteratively until the algorithm converges, meaning that the parameters reach values where the cost function is minimized or until the convergence criterion is met.

Types of Gradient Descent

There are several variants of gradient descent, each with its characteristics and applications. The three main types are

  • Batch Gradient Descent

  • Stochastic Gradient Descent

  • Mini-Batch Gradient Descent

Batch Gradient Descent

Batch Gradient Descent, also known as vanilla gradient descent, is the simplest form of gradient descent. In batch gradient descent, the model parameters are updated based on the gradients of the loss function computed over the entire training dataset. It involves calculating the gradients of the loss function for each parameter by considering all the training examples simultaneously.


  • Guarantees convergence to the global minimum for convex optimization problems.

  • Provides a smooth and stable convergence trajectory.


  • Computationally expensive for large datasets as it requires processing the entire dataset in each iteration.

  • Memory-intensive since it needs to store the entire dataset in memory.

Stochastic Gradient Descent

Stochastic Gradient Descent (SGD) is a variant of gradient descent where the model parameters are updated based on the gradients of the loss function computed for individual training examples. Unlike batch gradient descent, SGD processes one training example at a time.


  • Suitable for large datasets as it processes one example at a time, reducing memory requirements.

  • Faster convergence since it updates the parameters more frequently.


  • High variance in parameter updates due to the noisy estimation of gradients for individual examples, leads to erratic convergence.

  • Oscillatory convergence behavior due to the noisy updates.

Mini-Batch Gradient Descent

Mini-Batch gradient descent is a compromise between batch gradient descent and stochastic gradient descent. In mini-batch gradient descent, the training dataset is divided into small batches, and the gradients of the loss function are computed for each batch.


  • Combines the benefits of batch gradient descent and stochastic gradient descent.

  • Efficient use of computational resources by processing a subset of the data in each iteration.

  • Reduced variance in parameter updates compared to stochastic gradient descent.


  • Requires tuning the batch size hyperparameter, which can affect convergence behavior.

  • May introduce additional noise compared to batch gradient descent due to smaller batch sizes.

Comparison of the different types:


Batch Gradient Descent

Stochastic Gradient Descent

Mini-Batch Gradient Descent

Convergence Speed




Computational Efficiency

High (For large datasets)

Low (for large datasets)


Memory Requirements




Convergence Stability


Less Stable


Noise in Parameter Updates




Suitable for

Small Dataset, Convex problems

Large Dataset

Large Dataset

Tuning Parameters

Learning Rate

Learning Rate

Learning Rate, Batch Size

The Mathematics Behind Gradient Descent

Gradient Descent is a first-order iterative optimization algorithm for finding the minimum of a function. It’s widely used in machine learning and deep learning for training models.

Objective Function: The first step in Gradient Descent is to define the objective function, often denoted as f(θ).

This function represents the problem we want to solve. In the context of machine learning, this is often a loss function that we want to minimize.

Gradients: The gradient of a function at a certain point is a vector that points in the direction of the greatest rate of increase of the function at that point. The gradient is computed using partial derivatives.

For a function f(θ) where θ = [θ1​,θ2​,...,θn​]

The gradient is defined as:

Gradient Descent

Update Rule: In each iteration of the Gradient Descent algorithm, we update the parameters θ using the following rule:

Gradient Descent 2


α is the learning rate, a hyperparameter that determines the step size at each iteration while moving toward a minimum of a loss function.

This process is repeated until the algorithm converges to an optimal solution, i.e., until the change in loss is smaller than a predefined threshold, or until a maximum number of iterations is reached.

Use of partial derivatives

They are used to compute the gradient, which tells us the direction in which we should change our parameters to minimize the loss. By taking steps proportional to the negative of the gradient, we move towards the minimum of the function.

The learning rate determines the size of these steps α

If the learning rate is too high, we may overshoot the minimum, and if it’s too low, the algorithm may take a long time to converge or get stuck in a local minimum. Therefore, choosing an appropriate learning rate is important for the efficient performance of Gradient Descent.

Step-by-Step Implementation of Gradient Descent

Initialize Parameters: Begin by initializing the parameters of the model or function you want to optimize. These parameters are typically denoted as θ and could represent coefficients in a linear regression model, weights in a neural network, or any other variables that affect the output of the function.

Define the Cost Function: Specify a cost function J(θ) that quantifies how well the model performs with the current set of parameters θ. The goal of gradient descent is to minimize this cost function.

Set Learning Rate and Convergence Criterion: Choose a learning rate (often denoted as α), which determines the size of the steps taken during optimization. Also, decide on a convergence criterion, such as a maximum number of iterations or a minimum change in the cost function between iterations.

Gradient Calculation: Compute the cost function gradient concerning each parameter θj​. This gradient indicates the direction and magnitude of the steepest ascent of the cost function and is given by the partial derivative: ∂J(θ) / ∂θ j​​.

Parameter Update: Update each parameter θj​ using the gradient and the learning rate. This step aims to move the parameters in the direction that decreases the cost function. The update rule for each parameter is: θj​ := θj ​− α ∂J(θ)​ / ∂θj​

This update is performed simultaneously for all parameters θj​.

Convergence Check: After updating the parameters, check if the convergence criterion is met. If the algorithm has converged (i.e., the change in the cost function is negligible or the maximum number of iterations is reached), stop the optimization process. Otherwise, return to step 4 and repeat the process.

Below is the implementation of Gradient Descent algorithms, which is used to find the minimum of a function.

function compute_gradient(X, y, theta):
    m = length(y)
    num_parameters = length(theta)
    gradient = zeros(num_parameters)
    # Compute the gradient for each parameter
    for j from 1 to num_parameters:
        gradient[j] = (1 / m)  sum((X  theta - y) * X[:, j])    

    return gradient


Consider the below example to find the minimum of a simple quadratic function of the form f(x)=x2

Here’s the Python code:

import numpy as np

# Define the function
def f(x):
    return x**2

# Define the derivative of the function
def df(x):
    return 2*x

# Gradient descent function
def gradient_descent(x_start, learning_rate, num_iterations):
    x = x_start
    for _ in range(num_iterations):
        grad = df(x)
        x = x - learning_rate * grad
    return x

# Initialize parameters
x_start = 5
learning_rate = 0.1
num_iterations = 100

# Run gradient descent
x_min = gradient_descent(x_start, learning_rate, num_iterations)

print(f"The minimum of the function occurs at x = {x_min}")

When you run this code, it will output something like:

The minimum of the function occurs at x = 7.888609052210118e-17

The exact value may vary slightly due to the precision of the floating-point calculations.

Variants and Extensions of Gradient Descent

Momentum-based Gradient Descent:

  • Momentum-based gradient descent enhances traditional gradient descent by incorporating momentum, which helps accelerate convergence and navigate through shallow local minima.

  • Momentum accumulates a velocity term that guides the parameter updates, allowing for smoother and faster convergence, especially in regions with high curvature.

  • This approach is particularly beneficial for optimizing deep neural networks and other complex models where traditional gradient descent may struggle.

Nesterov Accelerated Gradient Descent:

  • Nesterov accelerated gradient descent (NAG) is an improvement over momentum-based gradient descent, offering better convergence properties.

  • NAG first computes an intermediate update based on the current velocity and then evaluates the gradient at this updated position.

  • By anticipating the future position of the parameters, NAG reduces the oscillations and overshooting commonly encountered with standard momentum-based methods, leading to faster convergence.

Adaptive Learning Rate Methods (e.g., AdaGrad, RMSprop, Adam):

  • Adaptive learning rate methods dynamically adjust the learning rate during training, alleviating the need for manual tuning and improving convergence performance.

  • AdaGrad adjusts the learning rate for each parameter based on the historical gradients, scaling down frequently updated parameters and amplifying the updates for infrequently updated parameters.

  • RMSprop addresses the diminishing learning rate issue of AdaGrad by using a moving average of squared gradients, which prevents the learning rate from decreasing too rapidly.

  • Adam combines momentum and adaptive learning rate techniques, offering the benefits of both approaches. It maintains separate adaptive learning rates for each parameter and includes bias correction mechanisms to mitigate initialization biases.

Second-order Optimization Methods (e.g., Newton's Method, Quasi-Newton Methods):

  • Second-order optimization methods utilize information beyond the first derivative (gradient) to approximate the curvature of the cost function.

  • Newton's method directly incorporates the second derivative (Hessian matrix) of the cost function, allowing for more precise parameter updates and faster convergence in well-conditioned problems.

  • Quasi-Newton methods, such as BFGS and L-BFGS, approximate the Hessian matrix iteratively, offering computational advantages over Newton's method for large-scale optimization tasks.

  • These methods can converge more rapidly than first-order methods but require additional computational resources for storing or computing second-order information.

Applications of Gradient Descent

Gradient descent finds applications in various real-world scenarios where optimization is required. Some common applications include:

  1. Machine Learning Models Training: Gradient descent is extensively used in training machine learning models, including linear regression, logistic regression, support vector machines, and neural networks. It helps adjust the parameters of these models to minimize the error between predicted and actual values.

  2. Deep Learning: In deep learning, particularly in training neural networks with multiple layers, gradient descent (along with its variants) is crucial for adjusting the weights and biases of neurons during the backpropagation process. This optimization enables neural networks to learn complex patterns and make accurate predictions in tasks such as image recognition, natural language processing, and speech recognition.

  3. Optimization Problems: Gradient descent is applied in solving optimization problems across various domains, such as finance, engineering, and physics. It helps in finding the optimal solution for problems like portfolio optimization, parameter tuning in engineering designs, and fitting mathematical models to experimental data.

  4. Natural Language Processing (NLP): In NLP tasks like language translation, sentiment analysis, and text generation, gradient descent is used to optimize the parameters of deep learning models such as recurrent neural networks (RNNs) and transformers. These models learn from large amounts of text data to perform tasks like language understanding and generation.

  5. Computer Vision: Gradient descent is employed in computer vision tasks like object detection, image segmentation, and image classification. Convolutional neural networks (CNNs) are trained using gradient descent to learn features from images and predict their contents.

Benefits of Gradient Descent:

  1. Efficiency: Gradient descent is computationally efficient, especially for large-scale optimization problems, due to its iterative nature and ability to work with large datasets.

  2. Versatility: It can be applied to a wide range of optimization tasks and is not limited to specific types of models or functions.

  3. Scalability: Gradient descent scales well with the size of the dataset and the number of parameters, making it suitable for training complex models.

  4. Global Optimization: While gradient descent may converge to local minima, various extensions, and techniques, such as momentum and adaptive learning rates, can help mitigate this issue.

Limitations of Gradient Descent:

  1. Sensitivity to Learning Rate: The performance of gradient descent is sensitive to the choice of learning rate. A too-small or too-large learning rate can lead to slow convergence or divergence, respectively.

  2. Local Optima: Gradient descent may converge to local minima, especially in non-convex optimization problems, resulting in suboptimal solutions.

  3. Saddle Points: In high-dimensional spaces, gradient descent can get stuck at saddle points, where the gradient is zero but the point is not an optimum.

  4. Non-Differentiable Cost Functions: Gradient descent requires the cost function to be differentiable, limiting its applicability in cases where the cost function is non-differentiable or discontinuous.


The Gradient Descent algorithm is a powerful optimization tool that has found extensive use in machine learning and deep learning. Its ability to iteratively minimize a cost function makes it ideal for training models, and its simplicity allows for easy implementation and understanding. We’ve explored the mathematics behind Gradient Descent, its different types, and the step-by-step process of its implementation. We’ve also discussed its applications, advantages, and disadvantages.


bottom of page