Deep Learning: what it is, how it works, and use cases

Born as a specific subset of machine learning (ML), today deep learning (DL) tands out as one of the most fascinating techniques in the field of artificial intelligence, due to its proclaimed functional resemblance to the human brain. This resemblance also inspires its generalist approach to the problems it seeks to solve.

Thanks to the proven expertise of Moxoff, which has been engaged for years in developing DL systems within the enterprise sector, we will explore what deep learning consists of, the technologies that characterise it, the main differences compared to machine learning, and the areas of business where it is currently being successfully implemented.

What is deep learning

According to the definition provided by Gartner, deep learning is a “Variant of machine learning algorithms. It uses many layers to solve problems, gaining knowledge from raw data by transforming it at each layer. These layers progressively extract increasingly accurate features from the raw data, enabling the solution of complex problems with a high level of precision and limited human intervention.”

According to Techtarget: “Deep learning is a type of machine learning that mimics the way humans acquire certain kinds of knowledge. Deep learning models can be trained to perform classification tasks and to recognise patterns in images, text, audio, and various other data. It is also used to automate tasks that would normally require human intelligence, such as describing images or transcribing audio files.”
Thanks to this inherent capability, deep learning assists data scientists when they need to collect, analyse, and interpret large quantities and varieties of data, due to its strong analogy with the morphology of the human brain. In fact, DL features neural networks made up of multiple layers of nodes that work together in ways comparable to the millions of interconnected neurons in the human brain.

Deep Learning e Machine Learning

Deep learning is a sub-branch of machine learning, so it is reasonable to expect that DL and ML are based on similar principles, despite the wide variety of technologies that characterise them. However, their use can differ significantly.

Machine learning operates in more specific fields, offering solutions to problems that are generally known in advance and described through the data sets used to train the algorithms. Deep learning, on the other hand, uses more complex and layered technologies, as well as referring to much broader bodies of knowledge. The “deep” approach consists of trying to generalise problems without necessarily contextualising them in advance, with the aim of finding solutions according to the logic that the human brain would use when faced with unfamiliar situations.
With the sole intention of providing information, we could summarise the comparison between machine learning and deep learning as follows.

Machine Learning (compared to DL)

1. It uses algorithms to learn an approximate representation of the phenomenon to be modelled.
2. It is effective even with a smaller dataset.
3. Shorter training times.
4. Greater simplicity in interpreting the analytical result.
5. It requires more limited hardware resources.

Deep Learning (compared to ML)

1. It uses algorithms to learn an approximate representation of the phenomenon to be modelled.
2. It requires large volumes of data.
3. Training times can be very lengthy.
4. Effective for complex tasks.
5. Results are not always easy to interpret, with a high "black box" effect.
6. It requires significant computational resources.

Deep Learning: when and how it originated

The history of deep learning forms part of the fascinating chronology ofArtificial Intelligence, beginning in 1943, when Warren McCulloch and Walter Pitts created the first computational model for neural networks: “threshold logic”. In other words, the very first deep learning algorithm.
The first major milestone in DL was set in 1958, when Frank Rosenblatt introduced the famous perceptron, a pattern recognition algorithm based on a two-layer neural network, using the mathematical principles of addition and subtraction. This marked the beginning of a long period of research and development, which would only bear fruit in 1975.

In the following years, other increasingly sophisticated multilayer neural networks made their debut, such as the Neocognitron (1980), designed by Kunihiko Fukushima to tackle tasks like handwritten character recognition and other pattern recognition problems.
Despite the success of research in the field, which led to a flourishing scientific literature on the subject, deep learning—like artificial intelligence more broadly—entered a so-called "AI winter" that lasted around twenty years. The main reason for this abrupt slowdown was the complexity of the computations required by deep neural networks, which at the time took many days to run even on supercomputers. These circumstances, although they clearly indicated the potential of the technology, discouraged its adoption in current applications, both in the consumer and enterprise sectors.

In the 2010s, thanks to huge advances in computing technologies and the advent of cloud computing, deep learning resumed its rise to prominence—including in the media. This resurgence was demonstrated on various occasions by DeepMind, an AI lab acquired by Google in 2014, which developed applications such as AlphaFold. AlphaFold was able to solve the protein folding problem, quite literally ushering in a new era in the field of molecular biology.
Deep learning applications, fuelled by major incentives from big tech companies, have become part of everyday technology, accessible to everyone, often without users even realising it. One example is DeepFace, a deep learning technology implemented by Facebook for facial recognition and automatic tagging of user profiles. To give an idea of the technological complexity of DL, it is enough to note that the first version of DeepFace used deep neural networks with over 120 million parameters.
Today, deep learning is used in a wide range of business sectors. Thanks to functions such as object detection/recognition, image classification, and natural language processing, it underpins applications including automatic text and image generation, multilingual simultaneous translation, sentiment analysis and speech recognition.

How Deep Learning Works

To understand the basic principles of how deep learning works, it is useful to highlight the crucial stages in the development process.

Neural network structure

The first step consists of choosing the type of neural network to use. Deep neural networks (with three or more layers) make use of intermediate layers (hidden layers) to build various levels of abstraction, aiming to use the minimum resources necessary to solve a mathematical problem.
The expertise of the data scientist and data engineer in optimising resources according to the required level of complexity is fundamental in this context. A deep neural network with many hidden layers can achieve highly accurate results, but this comes at the cost of increased complexity in terms of dataset definition, mathematical modelling, and the computational resources needed to carry out the necessary processing.

Network training

Deep learning systems are trained using labelled data with specific meta tags. A common example is image recognition. A dataset containing images of cows will have a series of files tagged with the metatag “cow”, but no further instructions are given to the system.

It is the neural network itself, through the interaction of its various hierarchical layers, that learns to identify the distinctive features of a cow and to recognise it as such. DL systems therefore use a trial and error learning procedure, which requires large datasets in order to generalise the problem.
Today, there are various frameworks available for training deep neural networks. Notable examples include TensorFlow, an open-source library developed by Google, and PyTorch, an open-source project supported by Meta, built on the foundations of the Caffe2 and ONNX frameworks.

Backpropagation and optimisation

The morphological complexity of a neural network is reflected in the objective difficulty of adjusting and weighting its various nodes. For training deep neural networks, error backpropagation algorithms are commonly used; these algorithms automatically adjust the weights of neural connections when the outcome does not meet expectations (in the case of an error).
In other words, the backpropagation algorithm iteratively optimises the result, constantly seeking the most appropriate weights to use based on the estimated error.

Testing and validation

In machine learning, datasets are usually divided into three types: training, validation and test setThis approach allows the model to be trained on one set, tested and refined with a second set, and validated with a further set. Such a procedure offers obvious advantages when working with large datasets, as is common in deep learning.

This is a process that requires skills and experience that cannot be improvised. Determining the appropriate sizes for the training, validation, and test sets depends on various factors, such as the size of the dataset, its complexity, the number of model parameters, and the nature of the problem being addressed.
Years of proven field experience now enable Moxoff to make the most effective decisions regarding the definition of datasets used for training deep learning systems.

Implementation

Once developed, a deep learning application can be implemented to deliver rapid results, even for highly complex tasks. However, as expressly recommended by the recent (AI Act)it is always advisable to establish an appropriate workflow for monitoring and evaluating results, aimed at assessing the validity of outcomes over time, ensuring compliance with regulatory requirements, and updating documentation to explain the functioning of the AI system itself.

Key components of deep learning

The terminology of deep learning refers to at least the following key components: artificial neurons, neural networks, and activation functions.

Artificial neurons

Artificial neurons are the fundamental components of mathematical models known as neural networks, whose operational scheme is explicitly inspired by the neural connections in the biological brain. Neurons are linked in successive layers by weighted connections, numerical values multiplied by each neuron's value. Artificial neurons are present at all levels of the neural network: input, hidden (intermediate), and output, and they effectively govern how the network operates.

Neural networks

A neural network is traditionally composed of three types of layers: input, hidden (intermediate), and output layers.

• • 1. Input layer: responsible for the input of data into the artificial neural network
    2. Hidden layers: once the data has been received from the input layer, the hidden layers continuously adapt their behaviour and, thanks to neuronal interconnections, enable the analysis of a phenomenon from numerous perspectives.
    3. Output layer: consists of a set of nodes that generate output data. In a yes/no system, there are two output nodes, but this number can rise to hundreds depending on the complexity of responses for which the deep learning system is programmed.

Among the most widespread neural networks are convolutional neural networks (CNNs), used for computer vision and image classification, and recurrent neural networks (RNNs), used for processing and speech recognition applications..

Activation functions

Artificial neurons receive inputs and weights, which are converted into outputs. This is made possible by specific activation functions, whose general role is to help the neural network learn complex relationships and patterns in the data through its various layers.
There are various types of activation functions (Sigmoid, Softmax, ReLU, Swish, ELU, etc.) to meet a wide range of application requirements. The choice of which to use depends on the expertise and practical experience of the data specialist.

The Advantages of Deep Learning

A correct and well-informed implementation of deep learning systems within business processes can yield significant benefits, chief among them the ability to process informationin great depth and with remarkable versatility, using a continuous improvement approach typical of machine learning.

Advanced information processing

DL applications deliver much more insightful results than traditional machine learning techniques, thanks to their ability to process unstructured data, which is increasingly prevalent in modern multimedia contexts. Deep neural networks enable analyses that can fully unlock the informational potential of organisational data.

Continuous improvement

Like machine learning, deep learning is an artificial intelligence technique based onautomatic learning, allowing the system to be trained and to progressively improve its effectiveness over time. Once implemented in a business setting, a deep learning system represents a long-term investment that can generate ongoing value, reinforcing a data-driven culture within organisations.

Versatility

Among the results achievable with DL is the ability to automate many routine operations traditionally carried out by humans, especially those with low added value. This leads to significant improvements in performance and reductions in errors. Furthermore, the growing availability of cloud services is democratising DL applications, making them accessible to a wider range of businesses, including SMEs.

The challenges of deep learning

Deep learning is an emerging technology that is maturing at an impressive rate. However, alongside its significant advantages in terms of accuracy and scalability, there are also critical challenges that need to be carefully considered before adoption. Ideally, this should be done with the support of qualified consultancy, capable of analysing business needs on a case-by-case basis in order to find the most effective solution for the specific context.

High Demand for Computational Resources

Deep neural networks require highly complex computations and, as a result, considerable computing resources. Modern AI PCs, equipped with neural processing units (NPUs), allow much of the traditional workload handled by CPUs and GPUs to be streamlined, particularly for short and frequent tasks. However, for training DL systems such as large language models (LLMs), high performance computing (HPC) systems are required, the energy consumption and running costs of which are far from negligible.

Need for large datasets

Deep learning models require large amounts of labelled data for training, which can be costly to acquire and manage, and also require appropriate data governance—not always feasible for every organisation. Today, synthetic data and other techniques can help facilitate the creation of datasets, but the effort involved remains significant, especially for large-scale models.

Lack of transparency and interpretability

The complexity of deep neural networks effectively creates a "black box," where results are produced without it being clear how they were generated. In addition to the challenge of interpreting the outputs, there are also possible issues with overfitting, which can make the system too tailored to specific cases, limiting its ability to deal with more general problems.
Moxoff’s expertise enables companies to overcome these potential adoption barriers, or to recommend alternative artificial intelligence techniques, when they prove to be more suited to the specific needs to be addressed.

Practical applications of deep learning

Deep learning is now used in a wide range of fields, including machine vision (computer vision), natural languageprocessingand patternrecognition).

Machine Vision

Traditional machine vision applications have achieved significant improvements in accuracy and performance thanks to deep learning techniques, which are widely used in areas ranging from medical diagnostics to quality control in dustrial production processes..
The main DL functions in the field of machine vision are image classification, semantic segmentation, object recognition, and anomaly detection. Convolutional neural networks (CNNs) are primarily used for these purposes.

Natural language processing

Deep learning is increasingly used in the context of natural language processing (NLP) for applications such as sentiment analysis, simultaneous translation, document management, chatbots, information retrieval, content summarisation, anti-spam systems, grammar correction, and the generation of textual content.

Pattern Recognition

Pattern recognition is a technique based on deep learning that uses data analysis to identify patterns and any anomalies that may arise compared to standard situations. In fact, the data can be anything: text, images, video, audio, etc. One of the main strengths of this technique lies in its ability to recognise hidden patterns even in unfamiliar objects, thanks to the very nature of deep learning.

Deep Learning use cases

Deep learning-based applications have now become standard practice in sectors such as finance, healthcare, and manufacturing.

Settore finanziario

The financial sector uses deep learning techniques in a variety of applications, ranging from predictive analysis of market trends and investment risk assessment to anti-fraud systems. In particular, these anti-fraud systems employ anomaly detection techniques to identify potentially suspicious behaviours indicative of fraud against banking systems (such as POS, online banking transactions, credit cards, etc.).

Healthcare Sector

Image recognition techniques based on DL are used in radiology to identify anomalies in diagnostic images. DL and ML applications can detect features that the human eye would not be able to spot, enabling the discovery of diseases at an early stage and significantly reducing the number of false positives.

Manufacturing sector

Deep learning systems are widely used in the manufacturing industry for applications such as predictive maintenance, intelligent process automation (IPA), quality control, forecasting market and component demand for production, contract management, optimisation of production processes, digital twins, IT/OT cybersecurity, and many more.

Technologies complementary to deep learning

Deep learning-based applications are often used in conjunction with other artificial intelligence techniques, such as machine learning itself—from which it is derived and more recent applications, such as: RAG (retrieval-augmented generation).

Integration with Machine Learning

The different purposes of machine learning and deep learning make the use of these two artificial intelligence techniques perfectly complementary: ML can be used to solve known problems, while DL offers a generalist approach, even to situations that are unknown in advance.

Integration with RAG

Today, RAG (retrieval-augmented generation) is widely used to make commercial large language models (LLMs), which are trained on vast datasets—more specific to the needs of implementing organisations. These models do not include proprietary organisational data by default. RAG works by querying the model and adding specific information into the prompt, thereby avoiding the need to retrain the model.

Moxoff and Deep Learning

Consultancy and the development of solutions based on machine learning simplify the implementation of systems designed to improve and make business decisions more informed , as well as to increase the efficiency of business processes. Moxoff offers its proven expertise in ML and DL to help companies get the most out of these technologies, successfully overcoming all the challenges and critical issues associated with their adoption.

Moxoff enables companies to implement a bespoke solution to meet their specific needs, thanks to an end-to-end approach that includes business and data analysis, system design, integration, and the maintenance of deep learning systems to enhance their efficiency over time.