Artificial neural network: Here's everything you need to know about black box of AI

• Artificial neural networks (ANNs) mimic biological neural networks in the human brain.

• ANN consists of an input layer, a hidden layer, and an output layer.

• Also called neural nets, ANNs are used daily in healthcare, social media when suggesting people you might know, and in marketing when recommending products to consumers.

Artificial neural networks (ANNs), also known as neural nets, are computing systems that are inspired by the way biological neural networks work in the human, or other animals, brain. In order to look further into artificial neural networks, it’s important to understand the basis of the origin and neural circuits of the brain itself.

Neural circuits are groups of neurons that are connected by synapses – a structure that allows neurons to pass chemical and electrical signals to other neurons. Neural networks, within the brain and in human intelligence, have been the inspiration behind creating artificial neural networks in artificial intelligence.

What is an artificial neural network?

An artificial neural network is a system of algorithms that works in a similar way. It is one of the subsets of machine learning under artificial intelligence. The models are based off of biological neurons in the brain forming a neural network.

The artificial neurons form the basis of artificial neural networks. Similarly to biological neural networks, which have neurons and synapses, ANNs have nodes and connections between nodes. As the ANN analyzes large amounts of data, it forms new connections and develops the capability to solve difficult problems or perform challenging tasks.

Past history of ANNs

In 1958, Frank Rosenblatt, an American psychologist, created one of the first prototype models of an artificial neural network, building on the work of researchers Warren McCullock and Walter Pitts, who published their concept of an artificial brain cell as a logic gate with binary outputs in 1943. The name of Rosenblatt's creation was the Perceptron. The machine was considered one of the first to "perceive an original idea." It was created to replicate how the human brain processes information and learned to identify various items.

“Stories about the creation of machines having human qualities have long been a fascinating province in the realm of science fiction,” Rosenblatt wrote in 1958. “Yet we are about to witness the birth of such a machine – a machine capable of perceiving, recognizing, and identifying its surroundings without any human training or control.”

Although the Perceptron was not able to recognize complex patterns, it was an innovative step forward in artificial neural network research.

How do artificial neural networks learn and work?

ANNs are inspired by neural networks in animals and are able to "learn" and improve in order to solve problems, such as those related to pattern recognition. The artificial neuron is a mathematical function that acts in some ways as a simulation of biological neurons.

The artificial neurons receive input and then use the information to create the output or data. Biological neurons similarly have input and output signals. However, ANN uses mathematical equations to connect all of the artificial neurons to create the artificial neural network.

The system of artificial neural networks is still an area within artificial intelligence that needs to be further studied and examined. The biological inspiration behind ANNs is extremely complex and distinctive.

Researchers are also taking a look at different ways to incorporate biological components of communication within the artificial neural network. Recently, scientists have developed a realistic artificial neuron that can communicate in diverse ways, both chemically and through electric pulses, more closely mimicking a biological neural network. The unprecedented study was published on November 7, 2022, in the journal Nature Electronics.

Layers within artificial neural networks

Artificial neural networks use a machine learning algorithm based on a model of neurons within the human brain. ANNs are comprised of layers of nodes which have three layers. These consist of an input layer, one or more hidden layers, and an output layer.

The input layer is the layer in which information is fed into the artificial neural network from datasets. The input nodes process the data, categorize it, and pass it on to the next layer.

The hidden layer or layers take input data from the input layer or from other hidden layers. Each hidden layer analyzes the output from the previous layer, processes it further, and passes it on to either the next hidden layer or to the output layer.

The output layer receives information from the hidden layer. The output layer often has more than one output node, for example, when working on multi-class classification problems.

In the input layer, weights and thresholds are assigned to the data. These determine the importance of any given variable. Each node performs a calculation on the data it receives to apply the assigned weight. If the result is above a specified threshold, that node is activated, and the data is sent to the next layer of the network, where weights are again applied. Otherwise, the data is not passed along to the next layer. In this way, the nodes become better at "learning" which data is more accurate and which can be discarded.

What is ANN used for?

Artificial neural networks are often used for pattern recognition and to predict outcomes. For example, ANNs have been utilized for growing healthier agriculture, predicting wildfires, and detecting diseases.

Artificial neural networks have been used to help improve some agricultural processes by improving the classification of gathered crops based on their usage, predicting the presence of infectious plant diseases, and improving the production efficiency for agriculture. ANNs can predict underlying trends within cultivation by using datasets to find solutions for agricultural systems and finding efficient replacements for conventional methods to cultivate crops with the output information. ANNs are used in agricultural production to confirm the presence of diseases on various crops, create intelligent weed control and optimize the storage of the crops.

This is all done with large amounts of information presented in datasets. This comes from the data collected during the crops’ growing season, along with using the appropriate software to enter the statistics. Researchers use the ANN tools to support agricultural production by making it more efficient to grow and harvest crops. An example of how an ANN was used in agriculture was published in the journal Agriculture.

The study discussed how an ANN was used in agriculture by examining the impact of infestation from fungal diseases. The research also assessed how weather conditions impacted the concentration of ferulic acid in the wheat crops, and the link to deoxynivalenol, and nivalenol – two mycotoxins produced by fungi known as Fusarium – in winter wheat grains.

The ANN was created with 14 inputs (data retrieved by the system) and with a multilayer perceptron (MLP) topology with two hidden layers. The MLP network is mainly used for classification data analysis, in which a category of data is analyzed and identified to allow for more accurate analysis using neural networks.

The output from the study was an effective preventive measure for limiting the formation of infestation of crops in the fields by pathogenic fungi. The results of the research using ANN show the potential for using neural networks to analyze agricultural data, and in this case, the concentration of ferulic acid, deoxynivalenol, and nivalenol in winter wheat grains. This allows for more effective preventive measures to be taken to grow healthier agriculture.

Artificial neural networks have also been used to help identify and detect human diseases by analyzing data from various medical examinations, such as CT scans, ultrasound images, and X-rays. The information collected from the medical test is used as input and analyzed by ANN software to provide predictions and medical insights, the output - such as identifying the presence of cancer. The ANN is often constructed using a feedforward network model, with many healthcare systems using the multilayer perceptron feedforward ANN design.

Artificial neural networks are often capable of diagnosing specific diseases efficiently and accurately. This can further assist with medical diagnosis and treatment options for patients. Eventually, advancements in ANN could potentially assist in catching ailments before any symptoms appear or preventing illnesses by identifying, for example, genetic mutations or patients at risk of particular diseases. The ANN could also be used to assess the preventative options for a patient with a predisposition to a disease, and send out accurate output information prior to getting sick, rather than treatment. ANN could possibly diagnose patients before diseases become chronic or irreversible.

An example of using an artificial neural network to predict disease was published this year in the journal Nature Medicine. In the study, ANN was used to predict and detect Parkinson’s disease (PD) in patients, along with tracking its progression, using data from nocturnal breathing. Researchers trained the ANN using data from 7,671 participants who had PD. The ANN model could output an estimated PD severity and progression as well, based on its datasets. The study showed that ANN could be useful in disease analysis, risk assessment, and evaluation before a clinical diagnosis.

It has also been used in speech recognition software, systems to provide recommendations for returning customers and consumers, and on social media to suggest people who may know each other on the sites. These everyday applications use ANN to analyze information thoroughly and make a prediction based on the weighting applied.

Artificial neural networks are called the black box of AI

ANNs have the ability to learn from analyzing datasets, using either supervised or unsupervised learning.

In supervised learning, the ANN is given labeled datasets that provide the right answer in advance. For example, training data for facial recognition may involve labeling each image with various terms related to emotion. After the network has been trained, it will start making "guesses" about the emotion of a new image of a human face that it has never processed before.

With unsupervised learning, the ANN is trained with unlabeled data so that offers it no hints about what it's seeing.

This type of ANN, in particular, has also been called black boxes since it is not clear how the ANN "learns" - how all the individual neurons in an ANN work together to arrive at the final output. While it is clear how an individual node applies the weighting it was given during training and how it computes the output, it is less easy to understand how multiple nodes, or "neurons", arrive at a final output.

In other words, studying and analyzing the structure of some types of ANN doesn’t provide full insight on its functionality and how it arrives at its predictions. Hence the term "black box."

Some models of artificial neural networks

There are numerous types of artificial neural network models used in supervised and unsupervised learning. These differ in terms of complexity, use cases, and structure. Some of the main ones are Perceptron, multilayer Perceptron neural networks, feedforward artificial neural networks, and radial basis function artificial neural networks. An additional model that can be used in both supervised and unsupervised learning is a recurrent neural network.

A feedforward artificial neural network is one in which data moves in only one direction between the input and output nodes. Although there may be many different layers with many different nodes, the data only moves in one direction. Feedforward models are mainly used for simplistic classification problems.

Radial basis function (RBF) networks use activation functions to make predictions. This is a mathematical function that calculates the absolute value between a center point and a given point. They are similar to an MLP, but use only one hidden layer instead of multiple layers - they usually have an input layer, a layer with radial basis function nodes with different parameters, and an output layer.

One common use for radial basis function neural networks is in system control, for example, systems that control power restoration after a power cut by prioritizing repairs to the greatest number of people.

A recurrent neural network processes data sequentially. This model of ANN will move the data forward and loop it back to prior steps within the artificial neuron network.

The hidden layers are recurrent, allowing the data to be looped and also retained. The output is then looped back into the input, improving the development for the next input. This allows it to make better predictions and achieve a task efficiently.

This type of model may be used, for example, in a predictive text system, where the memory of a previous word in a string helps to better predict the outcome of the next word.

Future of artificial neural networks

Although there have been numerous advances in improving artificial neural networks, some researchers believe that AI is a long way off before it can qualify itself as being synonymous with human thinking, and learning and even being comparable to the biology of the human brain.

A recent study at MIT showed that this is, in part, due to the large number of high-quality datasets that must be fed to the system by humans to create such a neural network. The study was published on November 3, 2022, in the journal Nature Reviews Neuroscience.

Deep learning models will give us insight about the brain, but only after you inject a lot of biological knowledge into the model,” said Mikail Khona, an MIT graduate student and author of the new study. “If you use the correct constraints, then the models can give you a brain-like solution,” Khnoa added, referencing a neuron within the brain known as a grid cell, and using ANNs to mimic it.

In the future, an artificial neural network may be able to learn distinctively from itself, along with prior datasets, allowing it to learn in a manner more closely resembling human intelligence.