Search anything:

Adam Optimizer

Binary Tree book by OpenGenus

Open-Source Internship opportunity by OpenGenus for programmers. Apply now.


This article at OpenGenus introduces the Adam optimizer, an adaptive algorithm widely used in machine learning and deep learning. It combines Adagrad and RMSprop, ensuring faster convergence and improved performance for various tasks like image classification, object detection, language translation, and speech recognition. The article emphasizes its significance in advancing deep learning research and applications while mentioning the need for memory management and hyperparameter tuning in large-scale scenarios.

Table of Contents

No. Topic
1. What is Adam?
2. Adagrad and RMSprop
3. Adam, an improvement
4. How Adam works
5. Mathematically
6. Advantages and disadvantages
7. Uses
8. Conclusion

What is Adam?

Adam optimizer is an adaptive optimization algorithm used in training machine learning and deep learning models. It was introduced by Diederik P. Kingma and Jimmy Ba in their 2014 paper, "Adam: A Method for Stochastic Optimization."

The aim of the Adam optimizer is to speed up the training process and improve convergence in the optimization of neural networks. It does this by adjusting the learning rate for each parameter during the training process, allowing it to adaptively scale the step size based on the historical gradient information.

The name "Adam" stands for "Adaptive Moment Estimation," which reflects the core idea of the algorithm: keeping track of the average of how much each parameter changes (the first-order moment or mean) and how much the parameter's changes vary (the second-order moment or variance).

To understand Adam better, we might need to look two other optimization algorithms, RMSprop and Adagrad.

Adagrad and RMSprop

Adagrad (Adaptive Gradient Algorithm) is also an adaptive learning rate optimization algorithm that adjusts the learning rate for each parameter based on the historical gradients of that parameter. It accumulates the squared gradients for each parameter during training and uses this information to scale the learning rate. The idea behind Adagrad is that parameters that receive large gradients will have their learning rates reduced, while parameters that receive small gradients will have their learning rates increased. This helps to handle sparse data effectively.
However, Adagrad has an issue of a decreasing learning rate over time, which can cause the learning process to slow down significantly after some time.

RMSprop (Root Mean Square Propagation) is designed to address the diminishing learning rate problem of Adagrad. Instead of accumulating all past squared gradients, RMSprop uses a moving average of squared gradients to compute the learning rate for each parameter. This allows the algorithm to adapt more efficiently to different scales of gradients and helps prevent the learning rate from becoming too small during training.

Adam, an improvement

Adam combines the ideas from both Adagrad and RMSprop, aiming to take the benefits of both algorithms and address their individual limitations. Adam utilizes moving averages of both the first-order (gradients) and second-order moments (squared gradients) of the parameters and also includes bias correction to counteract the initial bias of the moving averages. This helps Adam to maintain higher learning rates during training, especially in the early stages, which can result in faster convergence and better performance on various tasks.

How Adam works

Imagine you are trying to teach a computer how to play a racing game. The computer needs to adjust its actions in each round to get better at winning the race. The Adam optimizer helps the computer adjust these actions in a smart way.

We can think of the Adam optimizer like a virtual driving coach for the computer in the racing game. The coach adjusts the car's steering based on how it previously turned and how much it needs to improve. If the car turned too much left before, the coach will help turn it a bit less left next time, and vice versa.

Let’s take another example. let's say the computer is learning to recognize different animals. It looks at pictures and tries to identify them correctly. The Adam optimizer helps it adjust the knowledge of each animal as it sees more pictures. If the computer makes a mistake and thinks a cat is a dog, the optimizer will help it correct the knowledge about cats so that next time it's more likely to identify them correctly.

Here’s how it does it:

Initialization: Adam initializes two moving average variables for each parameter in the model. These variables are used to keep track of the historical gradients for each parameter. They are initialized to zero.

Compute gradients: During each iteration of the training process, the gradients of the model's parameters are calculated using the loss function and back propagation. These gradients represent the direction in which the parameters should be updated to minimize the loss.

Update moving averages: Adam updates the moving averages using an exponential decay. The moving averages, also known as the first moments, capture the recent history of the gradients for each parameter.

Update second moment estimates: Next, Adam calculates the second moments, which represent the variances of the gradients for each parameter. These are also updated using an exponential decay.

Bias correction: Since the moving averages and second moments are initialized to zero, they may have a bias towards zero, especially in the early iterations of training. To correct this bias, Adam applies a bias correction step.

Update parameters: Finally, the parameters of the model are updated using the bias-corrected moving averages and a learning rate. The learning rate scales the size of the updates, and the moving averages help in adjusting the learning rate for each parameter based on their historical gradients.


The above can be mathematically expressed as follows:


m = 0 (initial first moment estimate)
v = 0 (initial second moment estimate)
t = 0 (initial iteration count)

Calculation of gradients ∇θ of the model's parameters with respect to the loss function.

Updating moving averages:

m = β1 * m + (1 - β1) * ∇θ

Updating second moment estimates:

v = β2 * v + (1 - β2) * (∇θ)2

Bias correction:

m̂ = m / (1 - β1t)
v̂ = v / (1 - β2t)

Updating parameters:

θ = θ - α * m̂ / (sqrt(v̂) + ε)


  • m is the first moment estimate (moving average) of the gradients.
  • v is the second moment estimate (uncentered variance) of the gradients.
  • β1 and β2 are the exponential decay rates for the first and second moments, respectively (typically set to 0.9 and 0.999).
  • α is the learning rate (step size) for updating the parameters.
  • t is the iteration count, starting from 0 and incremented at each step.
  • m̂ and v̂ are bias-corrected moving averages to account for the initialization bias.
  • ∇θ represents the gradients of the parameters with respect to the loss function.
  • ε is a small constant (usually around 1e-8) to prevent division by zero and add numerical stability.

Advantages and disadvantages


Efficient learning: Adam is an adaptive learning rate optimization algorithm, which means it automatically adjusts the learning rate for each parameter based on historical gradients. This adaptivity allows it to efficiently learn from data and converge faster, especially in the early stages of training.

Suitable for large datasets: Adam performs well on large datasets and tasks with a large number of parameters. It can handle sparse gradients effectively, making it suitable for tasks involving natural language processing and computer vision, where datasets can be huge and complex.

Robust to hyperparameters: Adam is relatively less sensitive to hyperparameter choices compared to some other optimization algorithms like traditional gradient descent. This characteristic makes it easier to use and saves time in hyperparameter tuning.


Memory intensive: Adam needs to store moving averages of past gradients for each parameter during training and hence it requires more memory than some other optimization algorithms, particularly when dealing with very large neural networks or extensive datasets.

Slower convergence in some cases: While Adam usually converges quickly, it might converge to flawed solutions in some cases or tasks. In such scenarios, other optimization algorithms like SGD (stochastic gradient descent) with momentum or Nesterov accelerated gradient (NAG) may perform better.

Hyperparameter sensitivity: Although Adam is less sensitive to hyperparameter choices than some other algorithms, it still has hyperparameters like the learning rate, beta1, and beta2. Choosing inappropriate values for these hyperparameters could impact the performance of the algorithm.


Image Classification: In tasks where the goal is to classify images into different categories (for e.g., recognizing whether an image contains a cat or a dog), Adam optimizer is often used to train convolutional neural networks (CNNs).

Object Detection: Object detection frameworks, such as YOLO (You Only Look Once) and SSD (Single Shot Multibox Detector), often use Adam optimizer during the training process.

Language Translation: For neural machine translation tasks, where the aim is to translate text from one language to another, recurrent neural networks (RNNs) and transformer-based models trained with Adam optimizer are used.

Speech Recognition: Automatic Speech Recognition (ASR) models, like those used in virtual assistants or speech-to-text applications, use Adam optimizer for training recurrent neural networks (RNNs) and transformer-based models.


Adam optimizer has become one of the most popular and widely used optimization algorithms in the field of deep learning. Its ability to adaptively adjust learning rates based on historical gradients makes it efficient and effective in training complex neural networks, particularly those involving large datasets. By combining the benefits of Adagrad and RMSprop while addressing their limitations, Adam has provided significant improvements in convergence speed and overall performance. Despite its advantages, programmers should be mindful of its memory requirements and carefully tune hyperparameters for optimal results. As the field of deep learning continues to advance, Adam remains a crucial tool for researchers and programmers.

Agniva Maiti

Agniva Maiti

Hi! I'm Agniva, an aspiring data scientist with a love for languages and coffee. I'm currently in my 2nd year of B.Tech (CSE) in KIIT, Odisha, India. I'm also interested in Web and App Development.

Read More

Improved & Reviewed by:

Benjamin QoChuk, PhD Benjamin QoChuk, PhD
Adam Optimizer
Share this