DL Series: Artificial Neural Networks (ANNs)

Photo by Alina Grubnyak on Unsplash

*Before reading this article, make sure to check out the first part of this four-part series on deep learning.


What’s this about: Artificial Neural Networks, or ANNs, are one of the most common types of deep learning networks. These biologically inspired computational networks consist of three layers and are used to solve a wide range of problems involving speech recognition, text translation, fraud detention, and much more.


Go Deeper to Learn More


Artificial neural networks are based on the studies of the brain and nervous system, and to an extent, they mimic biological neural networks. The most equal comparison between the two comes from ANN models’ ability to simulate the electrical activity of the brain and nervous system. In simpler terms, ANNs are created when regular computers are programmed to behave like they have interconnected brain cells.


How ANNs Work

Artificial neural networks rely on three types of layers of mathematical processing to analyze information:

  • Input Layer: The input layer receives different types of data from the outside world with the aim of processing it, analyzing it, and extracting insights from it.


  • Hidden Layer/s: After the input layer, the data moves onto the hidden layer, which can consist of more than one unit. Here the data is transformed into something the output layer can process. These layers are connected to one another, with each connection being weighted. The higher the weight, the greater influence one unit has on another. As the data moves through each one of these units, the network learns more about it. It’s important to note that the architecture of the hidden layer can vary greatly, much more than the input and output layers.


  • Output Layer: The output layer is where the network responds to the data it was fed by performing calculations via its neurons to produce the final result.

ANNs require massive amounts of information in the form of a training set in order to learn and extract insights. For example, if you are training an ANN how to spot differences between cats and dogs, it would require hundreds of thousands of images tagged as a dog. This would enable the network to identify important characteristics of a dog and classify future data based on it’s previous knowledge.

When the ANN is in the training phase, its output is compared to human descriptions of what it should be. The machine is validated if these align, but if it is incorrect, it goes through a back propagation process to adjust its learning. Back propagation involves the network working backwards through the layers to adjust the mathematical equations. This is the fundamental key to deep learning and is what makes networks intelligent.

The different layers of an ANN

ANNs can undergo various learning methods to acquire or enhance knowledge. The three main learning paradigms can be classified as supervised, unsupervised, and reinforcement learning. Supervised learning requires a human supervisor to classify training examples into classes, while unsupervised learning models can identify patterns independently. Reinforcement learning relies on a trial-and-error process where it is either awarded or penalized depending on its actions. The vast majority of practical ANN applications currently use supervised learning.

Advantages of ANNs

Here are some of the main advantages of ANNs:

  • Storing information on entire network: One of the greatest advantages of ANNs is that information is stored on the entire network, rather than on a database. This means if a few pieces of information are lost in one place, the network will not stop functioning.

  • Can work with incomplete knowledge: In some cases following the training of an ANN, the data can produce output even if there is incomplete information.

  • Fault tolerance: ANNs can still generate output even when one or more cells are corrupted, making the networks fault tolerant.

  • Parallel processing: ANNs can perform multiple tasks simultaneously.

Disadvantages of ANNs

ANNs also have a few disadvantages, such as:

  • Hardware dependence: ANNs require processors with parallel processing power due to their structure. In some cases, they require highly specialized hardware like graphic cards and AI chips.

  • Unexplained behavior: ANNs don’t tell why or how they produce a probing solution, which reduces network trust.

Network structure: Because there is no specific rule for determining the structure of ANNs, it must be achieved through trial and error.

Applications of ANNs

Artificial neural networks have many great properties that make them useful for a wide range of real-world applications:

  • Image and character recognition: ANNs are able to take in a lot of inputs before processing them to identify complex relationships, which makes them useful for image and character recognition.

  • Content recommendation: Major companies like Spotify use ANNs to recommend songs or other media to users.

  • Ad targeting: ANNs are used to create ads and target audiences, analyze the real-time performance of ads, and analyze the impact of ad content.

  • SEO: ANNs enable search engines to analyze content and ensure the best possible answers to queries. They can also be used to analyze content’s relevancy so that it can be tweaked.

  • Food science: A more unique application, ANNs are often used in food science to ensure food safety by carrying out tasks like the modeling of microbial growth, and the prediction of physical, chemical, functional, and sensory properties of food products during processing and distribution.

  • Other applications: Some of the other top applications for ANNs include named entity recognition (NER), part-of-speech tagging, question answering, paraphrase detection, machine translation, spell checking, stock market prediction, and more.

Solving Real-World Problems With ANNs

Artificial neural networks have advanced dramatically over the past years. First a theoretical undertaking, ANNs are now deployed throughout society to solve real-world problems in nearly every field. Commercial applications of ANNs are steadily rising, solving complex problems that were nearly impossible before.

With the rise of big data, where companies have been increasingly collecting massive amounts of data and labeled datasets, and parallel computing, data scientists are able to develop the most complex ANNs. With interest in ANNs increasing in both the public and private sectors, the technology has a long road ahead of it. This is just the beginning!


>>> Follow us on Twitter, Instagram, and LinkedIn for AI-related insights and news.

>>> Sign up for our newsletter here.

Giancarlo Mori