Backpropagation - The Most Fundamental Training Systems Algorithm in Modern Generative AI

29 Apr 2024

An Introduction to Backpropagation

This article focuses on a fundamental concept called backward propagation (backpropagation).

After reading, you'll understand:

  1. What is backpropagation and how it's critically important in all of Artificial Intelligence, especially Generative AI.

  2. What Gradient Descent is, the mathematics behind it, its types, and how it enabled AI to solve every problem posed properly and systematically once enough data was provided.

  3. Why backpropagation is the most universal learning algorithm in all of Machine Learning, because of a theoretical scientific result called the general-purpose approximation theorem.

Let's delve into backpropagation in depth by starting with its history.

History of the Backpropagation Algorithm

The backpropagation algorithm is a fundamental technique used in training artificial neural networks (ANNs).

It is a supervised learning algorithm that allows the neural network to learn by adjusting the weights and biases of the connections between neurons based on the calculated error.

The algorithm propagates the error back through the network, allowing it to adjust the weights and biases in a way that minimizes the error. The origins of backpropagation can be traced back to the 1960s and 1970s, with various researchers contributing to its development.

Here's a brief history of the backpropagation algorithm, along with key references:

  1. In 1960, Henry J. Kelley proposed a method called the "Membrane Theory of Aging" that involved a process similar to backpropagation for adjusting weights in neural networks.

  1. Paul Werbos independently derived a procedure similar to backpropagation in his 1974 PhD thesis, which he called the "back-propagated derivative" method.

  1. In 1986, David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams published a paper titled "Learning Representations by Back-Propagating Errors".

This paper is widely credited with introducing and popularizing the backpropagation algorithm, making it a practical and widely used technique for training neural networks.

After its introduction, backpropagation became a fundamental algorithm in the field of neural networks and was widely adopted in various applications, such as pattern recognition, computer vision, natural language processing, and more. Numerous variations and improvements to the backpropagation algorithm have been proposed over the years, including techniques for improving convergence, handling vanishing and exploding gradients, and adapting learning rates.

What is Backpropagation, and Why Does it Matter in Neural Networks?

Backpropagation is a supervised learning algorithm used to train artificial neural networks.

It is a method for calculating the gradients of the error function with respect to the weights and biases in the network. This information is then used to update the weights and biases in order to minimize the error function.

The backpropagation algorithm works by first calculating the error at the output layer of the neural network. This error is then propagated backwards through the network, from the output layer to the input layer, hence the name "backpropagation."

At each layer, the error is used to compute the gradients of the error function with respect to the weights and biases of that layer. These gradients are then used to update the weights and biases of the layer using an optimization algorithm, such as gradient descent.

The process of forward propagation (computing the output of the network), backward propagation (computing the gradients), and updating the weights and biases is repeated iteratively until the error function is minimized.

Backpropagation is an efficient way to train neural networks because it allows for the computation of gradients of the error function with respect to all the weights and biases in the network using a single forward and backward pass.

This makes it possible for you to train large and complex neural networks with millions of parameters. One of the key advantages of backpropagation is that it enables neural networks to learn hierarchical representations of the input data.

As the network is trained, the lower layers learn to extract low-level features from the input, while higher layers learn to combine these features into more complex representations.

This hierarchical learning process is what enables neural networks to achieve remarkable performance on a wide range of tasks, including image recognition, natural language processing, and many others.

What is the Time Complexity of the Backpropagation Algorithm?

The time complexity of the backpropagation algorithm depends on the size of the neural network, specifically the number of layers, the number of neurons in each layer, and the number of training examples. In general, the time complexity of backpropagation can be expressed as O(n * m * p), where:

n is the number of training examples

● m is the number of weights and biases in the neural network

● p is the number of operations required to compute the activation function and its derivative for each neuron

To understand this complexity, let's break it down into the different phases of the backpropagation algorithm:

  1. Forward Propagation: During the forward propagation phase, the input data is passed through the neural network to compute the output. This step involves performing matrix multiplications and applying activation functions for each layer. The time complexity of this phase is O(m * p), where m is the number of weights and biases, and p is the number of operations required for the activation functions.

  2. Error Computation: After the forward propagation, the error between the predicted output and the true output is computed. This step typically has a time complexity of O(n), where n is the number of training examples.

  3. Backward Propagation: During the backward propagation phase, the gradients of the error function with respect to the weights and biases are computed. This step involves performing matrix multiplications and applying the derivatives of the activation functions for each layer. The time complexity of this phase is also O(m * p), similar to the forward propagation phase.

  4. Weight and Bias Updates: Finally, the weights and biases are updated using an optimization algorithm, such as gradient descent. This step typically has a time complexity of O(m), where m is the number of weights and biases.

Since the forward propagation, backward propagation, and weight/bias updates need to be performed for each training example, the overall time complexity of the backpropagation algorithm is O(n * m * p).

It's important to note that the actual runtime of the backpropagation algorithm can be significantly influenced by various factors, such as the hardware used, the implementation details, and the specific characteristics of the neural network and the training data. Additionally, modern deep learning frameworks and libraries often employ various optimization techniques and parallelization strategies to improve the computational efficiency of the backpropagation algorithm.

In practice, the backpropagation algorithm can be computationally expensive, especially for large neural networks and large datasets. This has motivated the development of more efficient training algorithms, such as mini-batch gradient descent, and the use of hardware accelerators like graphics processing units (GPUs), tensor processing units (TPUs), and now, language processing Units (LPUs from the company Groq), to parallelize the computations and accelerate the process on a massive scale.

Gradient Descent and its Variants

Gradient descent is an optimization algorithm widely used in machine learning and deep learning to find the minimum of a function. It is particularly useful for training neural networks, where the goal is to minimize the error or loss function by adjusting the weights and biases of the network.

The basic idea behind gradient descent is to iteratively adjust the parameters of a function in the direction of the negative gradient of the function with respect to those parameters. The gradient represents the direction of the steepest increase of the function, and by moving in the opposite direction (negative gradient), the algorithm can approach the minimum of the function.

There are several types of gradient descent algorithms, each with its own advantages and trade-offs:

  1. Batch Gradient Descent: In batch gradient descent, the entire training dataset is used to compute the gradients and update the parameters in each iteration. This method ensures that the parameters are updated in the direction that minimizes the error across all training examples. However, batch gradient descent can be computationally expensive for large datasets, and it may converge slowly or get stuck in local minima.

  2. Stochastic Gradient Descent (SGD): Stochastic gradient descent is a variation where the gradients are computed and the parameters are updated based on a single training example at a time. This method is more efficient than batch gradient descent for large datasets because it does not require computing the gradients for the entire dataset in each iteration. However, SGD can lead to noisy updates and may require more iterations to converge.

  3. Mini-batch Gradient Descent: Mini-batch gradient descent is a compromise between batch gradient descent and stochastic gradient descent. In this method, the training dataset is divided into small batches, and the gradients are computed and the parameters are updated based on the average gradient of each batch. This approach strikes a balance between the stability of batch gradient descent and the efficiency of stochastic gradient descent.

  4. Momentum-based Gradient Descent: Momentum-based gradient descent introduces a momentum term to the parameter updates. This term accumulates the gradients of past iterations, allowing the algorithm to accelerate in the direction of the minimum and potentially escape local minima. Momentum can help the optimization process converge faster and more reliably.

  5. Adaptive Learning Rate Algorithms: Algorithms like AdaGrad, RMSProp, and Adam adapt the learning rate (step size) for each parameter based on the historical gradients. This can help the optimization process converge more quickly and avoid issues like the vanishing or exploding gradient problems common in deep neural networks.

The choice of gradient descent algorithm depends on various factors, such as the size of the dataset, the complexity of the problem, and the desired trade-off between computational efficiency and convergence speed. However, in practice, adaptive learning rate algorithms like Adam or RMSProp are often preferred for training deep neural networks due to their ability to handle sparse and noisy gradients effectively.

The Universality of the Backpropagation Algorithm

The backpropagation algorithm is a fundamental component in the training of neural networks, and it has played a crucial role in the success of generative AI models across various domains, including transformers, generative adversarial networks (GANs), autoencoders, deep learning, and deep reinforcement learning. Here's why backpropagation is used extensively in these areas:

  1. Transformers:

    Transformers, such as the popular models like BERT, GPT, and their variants, rely heavily on the backpropagation algorithm for training. These models consist of multiple layers of self-attention and feed-forward neural networks, and backpropagation allows for efficient computation of gradients and parameter updates during the training process. Without backpropagation, it would be extremely difficult to train these large and complex models effectively.

  2. Generative Adversarial Networks (GANs):

    GANs are a type of generative model that involves training two neural networks simultaneously: a generator and a discriminator. The backpropagation algorithm is used to train both the generator and the discriminator networks. The generator learns to generate realistic data samples by backpropagating the gradients from the discriminator's output, while the discriminator learns to distinguish real data from generated data by backpropagating the gradients from its own output.

  3. Autoencoders:

    Autoencoders are neural networks designed for unsupervised learning tasks, such as dimensionality reduction and data denoising. They consist of an encoder network that compresses the input data into a lower-dimensional representation, and a decoder network that reconstructs the original input from the compressed representation. Backpropagation is used to train both the encoder and decoder networks by minimizing the reconstruction error between the original input and the reconstructed output.

  4. Deep Learning:

    Deep learning models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), are widely used for various tasks, including image recognition, natural language processing, and speech recognition. These models often have multiple layers of neurons, and backpropagation is the primary algorithm used to train them. By backpropagating the errors from the output layer to the input layer, the weights and biases of the neural network can be adjusted to minimize the overall loss function.

  5. Deep Reinforcement Learning:

    In deep reinforcement learning, neural networks are used to approximate value functions or policy functions for decision-making in complex environments. The backpropagation algorithm is used to train these neural networks by backpropagating the temporal difference errors or policy gradients. This allows the agent to learn optimal behaviors by adjusting the weights of the neural network based on the rewards received from the environment.

The reason backpropagation is so ubiquitous in generative AI is its ability to efficiently compute gradients and update the parameters of complex neural network models. This capability is essential for training large-scale models with millions or billions of parameters, which are common in many generative AI applications.

Furthermore, the backpropagation algorithm is highly versatile and can be applied to different types of neural network architectures, loss functions, and optimization objectives, making it a powerful tool for a wide range of generative AI tasks.

Mathematical Intuition and General-Purpose Approximators

The neural networks can be formulated as dynamical systems, moving through what is known as an energy landscape. They move towards the global minimum of all the energy configurations that are approximated by the system of parameters.

The neural network traces a trajectory through the energy landscape.

You can think of it in the following manner:

The entire number of input, hidden, and output weights form an n * m * p matrix with dimensions of (you guessed it) n x m x p. So the hypersurface of dimensions - a 10,000 or 50,000 dimension space. Now human beings can’t imagine even 4-space, let alone 1000-space dimensions or 10,000-space dimensions.

The prevailing theory was that backpropagation, which follows the direction of the steepest descent (the gradient) would get stuck in local minima. But in practice, there are so many dimensions that absolute local minima, or stopping points for the operation of the backpropagation algorithm, simply cannot occur.

This means that given enough computational power, deep learning systems of dimensionality greater than even 1,000 parameters rarely get stuck in local minima. This means that deep learning systems can approximate the answer to any problem, once there is enough data! And, with the new advances in computation and hardware, there is both computational power and storage capacity to solve any given problem. This remarkable statement is summed up in the universal approximation theorem in neural networks system theory.

The Universal Approximation Theorem

The general purpose approximation theorem, also known as the universal approximation theorem, is a fundamental result in the field of neural networks and machine learning. It states that a feedforward neural network with a single hidden layer and a sufficient number of neurons can approximate any continuous function on a compact subset of real numbers to an arbitrary degree of accuracy.

In simpler terms, this theorem suggests that neural networks, which are mathematical models inspired by the human brain, have the remarkable ability to learn and mimic virtually any complex relationship or pattern present in data. Given enough neurons (the computational units within the network) and appropriate training, these networks can essentially approximate any function, mapping inputs to desired outputs with a high level of precision.

The significance of this theorem lies in its implications for the versatility and power of neural networks. It means that, in theory, a neural network with a suitable architecture and training process can be used to solve a wide range of problems, from image recognition and natural language processing to forecasting and decision-making tasks. The theorem provides a theoretical foundation for the successful application of neural networks in various domains, as they can effectively learn and generalize from complex data patterns.

Simply put, a neural network that is deep enough and has enough data can solve any problem presented to it. Humanity has been given the golden key to all scientific knowledge. All we need - is data! And computational power! And now, with GPUs, TPUs, and LPUs, we have more computational power than we ever need.

The deepest secrets of the cosmos lie within our grasp. With the right knowledge and data, humanity can solve any problem! That is the essence of the universal approximation theorem - the foundation of all artificial intelligence systems theory. And the foundation of the modern AI revolution. This is a tool whose true complexity and power eludes all but the most brilliant minds. And this is what has brought the human race to this junction in history. The answer to the darkest, deepest questions of the universe are now solvable.

It’s no surprise that every researcher secretly dreams of working in AI. This is its true potential! We can solve any problem that exists! And the limits are only what we can imagine. In essence, we can do anything. Solve anything. Achieve anything. It just takes enough understanding and insight. And of course, hands-on experience with neural networks.


There is no limit to what the human race can achieve. There is no stopping this revolution of cosmic potential. And Artificial General intelligence is just around the corner.

What will our human race create?

I can’t wait to find out!

May the glorious joy of the incredible potential of the future be with you.


All Images generated by DALL-E-3.