What is Adam Optimization?

Adam, short for Adaptive Moment Estimation, is an optimization algorithm that is used in the field of machine learning to update network weights iteratively based on training data. Adam is regarded as an extension to stochastic gradient descent and is known for its effectiveness in handling sparse gradients and its robustness to the choice of hyperparameters.

How Does Adam Work?

Adam combines the advantages of two other extensions of stochastic gradient descent, namely Adaptive Gradient Algorithm (AdaGrad) and Root Mean Square Propagation (RMSProp). AdaGrad adapts the learning rate to the parameters, performing smaller updates for parameters associated with frequently occurring features, and larger updates for parameters associated with infrequent features. RMSProp, on the other hand, utilizes a moving average of squared gradients to normalize the gradient itself, which has an effect similar to momentum.

Adam maintains two moving averages for each weight in the neural network: the first moment (the mean of the gradients) and the second moment (the uncentered variance of the gradients). Essentially, Adam adapts the learning rate by considering how quickly the first moment decays compared to the second moment.

The algorithm calculates an exponential moving average of the gradient and the squared gradient, and the parameters beta1 and beta2 control the decay rates of these moving averages. The moving averages themselves are estimates of the first moment (the mean) and the second moment (the uncentered variance) of the gradients. However, these moving averages are biased towards zero, so Adam performs bias-correction to account for this.

The general steps of the Adam optimization algorithm are as follows:

  1. Initialize the weights of the neural network.
  2. Calculate the gradients of the loss function with respect to the weights.

  3. Compute the moving averages of the gradients and their squared values.
  4. Correct the bias in the moving averages.
  5. Update the weights using the bias-corrected moving averages.
  6. Repeat steps 2-5 until the stopping criterion is met (e.g., a certain number of epochs or a threshold loss value).

Advantages of Adam

Adam has several advantages that make it suitable for a wide range of optimization problems in machine learning:

  • Efficiency: Adam is computationally efficient, requiring minimal memory.
  • Adaptive Learning Rate: It calculates individual adaptive learning rates for different parameters, which helps in faster convergence.
  • Handling of Sparse Gradients:

    Adam works well with sparse gradients, which are common in various problems such as natural language processing and computer vision.

  • Robustness: It is less sensitive to the choice of hyperparameters, especially the learning rate.

Hyperparameters in Adam

Adam has a few key hyperparameters that need to be set before the optimization process begins:

  • Learning Rate (alpha): The proportion that weights are updated (typically between 0.001 and 0.1).
  • Beta1: The exponential decay rate for the first moment estimates (commonly set to 0.9).
  • Beta2: The exponential decay rate for the second moment estimates (commonly set to 0.999).
  • Epsilon: A small scalar used to prevent division by zero (commonly set to 1e-8).

Limitations of Adam

Despite its popularity, Adam is not without limitations and may not always be the best optimizer for every machine learning problem:

  • Convergence: There is theoretical evidence that suggests Adam can fail to converge to the optimal solution under certain conditions.
  • Hyperparameter Tuning: While robust, Adam still requires careful tuning of its hyperparameters for best performance on specific problems.
  • Generalization: Some studies suggest that models trained with Adam may generalize worse on unseen data compared to models trained with other optimizers like stochastic gradient descent with momentum.


Adam is a powerful optimization algorithm that has become a default choice for training deep neural networks. Its ability to handle sparse gradients and adapt its learning rate for each parameter makes it a versatile and effective tool. However, practitioners should be aware of its limitations and consider alternative optimizers depending on the specific nature and requirements of their machine learning problem.


Kingma, D. P., & Ba, J. (2014). Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980.

Please sign up or login with your details

Forgot password? Click here to reset