Deep learning: The mechanics of magic

What is deep learning? 

Often abbreviated to DL, deep learning is a subset of machine learning and artificial intelligence: a technique of training computers to mimic the way neurons in the human brain process and learn information – it is quite literally an artificial neural network. 

With deep learning, machines are taught to process data in a way that simulates the many layers of our brain, so they can now perform tasks that were once exclusive to human talents. We are already seeing numerous deep learning applications and uses for DL meaning things like image recognition, language translation, and even autonomous driving, are all within the realm of computer capabilities. In essence, deep learning is a powerful tool for unlocking complex insights hidden within vast amounts of data. 

How does deep learning work?

The natural human brain is made up of millions of interconnected neurons. Modelled on nature, deep learning algorithms use artificial neural networks made of multiple layers of interconnected nodes, called artificial neurons or units. Each layer processes information from the previous layer and passes it on to the next, gradually extracting higher-level features and patterns.

Deep learning models are trained on large amounts of labelled data, known as training sets, to learn how to recognize and classify patterns. This might be a simple challenge like the “Dogs vs Cats” dataset – a standard computer vision dataset in which photos must be classified as either containing a dog or cat. More sophisticated datasets might include wine appreciation or how to spot fake news amid authentic reporting.

During training, the parameters and weights of the neural network are adjusted to minimize the difference between predicted outputs and true outputs. Across multiple iterations, the model becomes increasingly finessed until it is capable of making predictions or performing tasks with a high degree of accuracy and precision.

Assorted programming languages have emerged for implementing deep learning algorithms. Many are looking to use deep learning with Python for its simplicity, readability and extensive “libraries” (pre-written bundles of code which make it easier for developers to solve common problems). Popular libraries like TensorFlow and PyTorch provide a rich set of tools and functionalities for building and training deep learning models.

Deep learning vs machine learning

Deep learning is an approach to creating rich hierarchical representations through the training of neural networks with many hidden layers. It is an evolution of machine learning and one that needs much less help from humans. While basic machine learning models do improve at performing their specified functions as they absorb new data, if they return an inaccurate prediction, an engineer must intervene and make adjustments. Deep learning, though, makes its own adjustments, course-correcting without the need for human assistance.

Deep learning applications

Deep learning is already positively impacting our lives as it is deployed across many different sectors. These applications can be grouped into four broad categories:

• Computer vision (interpretation by computers of images and videos)
• Speech recognition
• Natural language processing, or NLP
• Recommendation engines

For example, within the healthcare sector, computer vision is playing a crucial role in medical image analysis, speeding up detection and diagnosis and identifying the most effective treatments. Speech recognition software is available for real-time notes from the audio feeds of patient-doctor consultations, saving hours of transcribing time.

On the roads, deep learning is accelerating the advancement of autonomous vehicles by enabling them to perceive and understand their surroundings, to read road signs and to navigate safely.

In the finance sector, deep learning’s capacity to analyse complex patterns makes it invaluable in pinpointing fraudulent transactions and informing investment decisions through its market trends predictions. In business, analysis of long-form reports can be pulled together in seconds using NLP.

Meanwhile, robots that possess deep learning systems can navigate, manipulate and interact more and more like humans. The technology is already making major leaps in industrial automation and human-robot collaboration – both pivotal in the manufacturing and electronics sectors.

And on an everyday level, we’re getting ever more pertinent pointers from recommendation engines on platforms like Netflix, Amazon and social media thanks to deep learning’s analysis of our user preferences.

What are the limitations of deep learning?

Despite its remarkable capabilities and breathtaking potential, DL is not without its limitations and issues. It demands high memory and storage capacity and consumes data at such a rate that obtaining high-quality, diverse and well-labelled datasets can be challenging, especially in domains where data is scarce or expensive. It’s also hungry for training resources and power and can be susceptible to overfitting – performing well on training data but poorly on new, unseen data.

In common with most areas of AI, deep learning also presents ethical concerns around data security and the inherent human biases and inaccuracies which can pervade areas of its learning.

Framework for deep learning

Although the potential benefits of deep learning are enormous, good governance of the technology will be essential to help us realize them. The development of appropriate legal and ethical frameworks will be critical to build trust and mitigate risks.

International Standards will have an important role to play in ensuring the responsible adoption of deep learning within the field of AI. ISO, in collaboration with the International Electrotechnical Commission (IEC), has already published a number of standards through its dedicated group of experts on artificial intelligence, and is building an evolving framework of guidance for deep learning applications around the world.

Businesses and policymakers can already benefit from these standards as part of their assurance of ethical and safe deployment of deep learning models. Complying with International Standards builds trust between clients and contemporaries in all sectors.

The future of deep learning

Deep learning is still in its infancy, but it’s expected to grow exponentially and is likely to transform society. The integration of deep learning with other cutting-edge technologies – for example, combined with augmented reality or virtual reality – could revolutionize the way we experience and interact with the world around us. A simple walk down the street could be augmented by instant information about nearby buildings and landmarks. Virtual worlds will become increasingly immersive and realistic.

We can expect to see deep learning applied to fields such as agriculture, energy and manufacturing, where it has the potential to optimize processes, improve efficiency and drive innovation. And, more importantly, to accelerate solutions to planet-wide problems like climate change and food security.

Guided by International Standards and driven by human creativity and ingenuity, the future landscape of deep learning is fascinating, inspiring and almost without limit.