The impact and power of generative artificial intelligence was realized in large numbers in companies last year. Now it seems that every software company, supermarket brand and company, even tangentially related to technology, is building its own AI-based solution for 2024. But how much do we know about the foundations of these technologies and what they mean to do with our data behind the scenes?
How well is the black box of artificial intelligence that we all learn to believe really understood outside of specialized tech circles? What should you know about something that is opaque by design? First, let’s start from the beginning.
AI, in short, is a broad group of systems and data brought together to learn a single, niche task. In a large-scale language model (LLM), this can be used to generate text based on recognizing patterns observed in a large array of previously entered content. All current artificial intelligence, from OpenAI’s ChatGPT and DALL·E 2 to Adobe’s Sensei image generation tools, works on this very principle of learning to replicate a task by mimicking previous data inputs.
More general artificial intelligence, a technology capable of context-switching to perform any given task, is still thought by most experts to be more than a decade away. However, the average user could be forgiven for thinking that the more general technology is akin to science fiction based on marketing hype and social media commentary.
However, the more you know about today’s increasingly popular tools, the easier it is to envision how they will evolve and change in the coming months and years. Far from the magic of science fiction, today’s generative artificial intelligence is simply a well-executed evolution of machine learning (ML) technologies that existed decades before. So what has changed in the last five years?
Here, we’ll dive into the realm of deep learning — one of the foundational technologies behind today’s ML technologies and the key enabler of today’s AI revolution.
What is deep learning?
Deep learning is a specific application of neural networks, which are in themselves a branch of machine learning. A neural network-based ML algorithm trains an evolving piece of software by mimicking the way an organic brain part learns — using incremental refinement algorithms to get an approximate solution closer and closer to a proposed goal.
Within this discipline, the applications of deep learning algorithms are extremely diverse and complex. Almost every other week, new systems are discovered for recognizing images and speech and visualizing data patterns that would otherwise be missed. In areas such as health technology, these technologies are proving to be revolutionary and literally changing lives.
Deep learning models have been trained to diagnose a range of diseases using medical images with spectacular success over the past 10 years. What it will be able to offer patients in the next decade through increased access, growing capabilities and specialized technologies is staggering.
“The power of deep learning is what enables seamless speech recognition, image recognition, and automation and personalization in every conceivable industry today, so it’s safe to say you’re already experiencing the benefits of deep learning,” Sajid Sadi, vice president of research at Samsung, told Forbes.
To achieve these goals in the new project, engineers must first build a neural network capable of learning within its domain. This is achieved by creating a large number of small, simple and relatively isolated pieces of code to perform tiny transformations on pieces of input data. These pieces of code, called nodes, mimic neurons in the brain by making tiny transformations and passing their results on to the next through a large interconnected network.
Like neurons in the brain, a single software node is trivially simple and almost useless. Even in a small number of only hundreds of nodes, the transformation done to input the data would be almost untraceable. What happens when these nodes and the key connections between them work at scale, the system’s performance becomes truly remarkable.
How does a neural network work?
When individual nodes are connected together, their connections are weighted to prioritize some communication channels and restrict others. These connections act as synapses in the organic brain, determining how information spreads and acts on it as a whole.
Using artificial neurons and synapses, modeled as nodes and weighted connections, the neural network builds a vast number of layers to mimic the scale and complexity of an organic brain — enabling it to perform similarly complex tasks. These nodes are divided into three types with their unique role in the system:
- Input nodes: These nodes are responsible for simply receiving data from outside the system and forwarding it to the network.
- Hidden nodes: They work with each other in a series of layers to process the data and make decisions about its accuracy.
- Output nodes: At the outermost layer, output nodes return the finalized data product to the outside world.
Much like an organic brain, the connections between nodes have proven to be critically important to the success of the system. The strength of the brain’s organic synapses plays the biggest role in its overall abilities. By instructing the neural network to learn, programmers systematically change these weights after processing the input to measure the distance between the algorithm’s output and our goal.
Using the cost analysis function on a large training data set, engineers can evaluate the network’s performance between iterations and adjust its learning to achieve their goals. What sets deep learning apart from other neural networks is the number of layers of hidden nodes it applies to its targets.
Variants of deep learning algorithms
Convolutional Neural Network (CNN)
For applications involving computer vision, images, and video, convolutional neural networks are most commonly used. They are ideally suited for detecting subsets of features and patterns within a larger data set (for example, a raw image) — enabling object detection, face recognition, and a host of related intelligent features.
Recurrent Neural Network (RNN)
In comparison, for areas such as natural language processing and speech recognition, recurrent neural networks are often used because of their ability to work well with sequential data.
One feature that all deep learning algorithms have in common is that they are significantly more complex than other ML techniques. It is this inherent complexity that allows deep learning algorithms to build a thorough working model of the tasks they seek to accomplish. However, as we’ll see, this is a feature that has significant engineering trade-offs to counter its revolutionary benefits.
Deep learning in practice
Deep learning is capable of doing something relatively unique in the field of machine learning. Unsupervised learning, taking unstructured raw data to categorize and distinguish fully autonomously, provides a new way of working on vast data sets without the need for human classifications.
With the processing power and storage available in today’s cloud systems, the algorithms used to work on our data are now capable of learning in an increasingly real sense of the word.
Already used to enable speech recognition in noisy environments, enabling accurate handwriting recognition and enabling facial recognition features used in mobile devices and smart homes, these systems are already something we take for granted every day.
Streaming giants rely on technology to recommend TV shows, games and music based on how large numbers of users have previously navigated their service. In the real world, users often find their predictions and recommendations to be eerily accurate.
Deep learning is a technology success story. Unsupervised learning is one of his latest ‘killer apps’. Used to build some of the largest and most capable platforms, it’s easy to imagine it as a ready-to-plug-in solution to any problem you can accurately describe. However, there are some downsides to this extremely powerful tool.
Disadvantages of deep learning approaches
The biggest pitfall of deep learning comes from the very nature of neural networks as a black box solution to artificial intelligence. Even when a system gives the correct answer 99 times out of 100, it’s hard to fully trust any software if you can’t see its inner workings.
The high degree of opacity of neural networks makes them difficult to recommend for applications that require extensive monitoring or regulation. Medical and aerospace applications are two key areas where this can raise significant issues.
The biggest problem in this area is a phenomenon that engineers define as dataset shift — where the training data is poorly matched to real-world examples or suffers from fundamental bias. The result of data movement problems often shows up in the unexpected behavior of algorithms deployed in real-world use cases.
Worryingly, this issue may not become apparent in early use or testing. Machine learning systems tend to fail silently. Problems resulting from data drift can lie dormant for months or years in use and can be fully diagnosed because the system makes decisions with poorly understood reasoning.
Another issue plaguing deep learning comes from under-specification — a problem Google recently discovered. Like data drift, this problem causes hiccups in real-world performance that may go unnoticed. Unlike dataset change, the main driver of this problem is not a lack of well-fitted training data, but a poorly optimized cost function that does not capture enough detail about the intended target.
Changes in algorithm output large enough to produce significant differences, but small enough to ensure that the cost analysis still provides a passing score that can be transferred to the resulting system and produce unintended consequences further down the line.
Solving real-world deep learning challenges
A problem common to both of these problems is the unknown and unknowable nature of current deep learning algorithms. When used in practice, it’s almost impossible to tell a good system from a bad one because we don’t have enough information about what’s going on underneath. But engineers are developing ways to overcome this obstacle.
The field of explainable artificial intelligence is growing at a rapid pace to develop algorithms that humans can better understand. This would circumvent many of the biggest problems inherent in the field, but we still have a long way to go before fully explainable deep learning algorithms begin to solve major medical problems. The advice today for engineers building deep learning systems is to test, test, test and then test some more.
“The biggest, immediate takeaway is that we need to do a lot more testing,” Alex D’Amour, who led the industry-leading underspec study, told MIT Technology Review. “We’re asking more from machine learning models than we can guarantee with our current approach.”
Extensive and thorough testing, continuous integration, and thorough validation are necessary when implementing a deep learning system to ensure success. Related fields such as synthetic data are helping engineers overcome these challenges by making larger amounts of higher data available through AI-enabled data synthesis.
The future of deep learning
Deep learning technologies have proven to be extremely powerful tools that have impacted all fields. From self-driving vehicles on the road today to modern advances in health diagnostics and accessibility applications — this is not a technology that can be dismissed easily.
Despite the potential pitfalls and drawbacks outlined here, this area is worthy of significant investment and research. If 2023 was the year of GPT, then 2024 surely marks the beginning of a decade of powerful, explainable and highly robust AI systems powering the software solutions of tomorrow.