Deep learning has become something of a catchphrase-du-jour. It is a trendy term that is being employed to address the latest wave of artificial intelligence (AI) technologies. Let’s take a look at what it is, how it compares to artificial intelligence, and how it is being applied.
Also see: Top AI Software
What is Artificial Intelligence?
Artificial intelligence can be thought of as the opposite of human intelligence. If living creatures are born with natural intelligence, man-made machines can be said to possess artificial intelligence. So any “thinking machine” has artificial intelligence.
In practice, however, computer scientists use the term artificial intelligence to refer to machines doing the level of thinking that humans have taken to a very high level.
Computers are very good at making calculations — at taking inputs, manipulating them, and generating outputs using algorithms. But in the past they have not been capable of other types of work that humans excel at, such as understanding and generating language, identifying objects by sight, creating art, or learning from past experience.
The development of artificial intelligence is an effort to change that, and give computers exceptionally advanced, human-like intelligence.
Also see: What is Artificial Intelligence
What is Deep Learning?
Deep learning could be defined as a form of artificial intelligence based on neural networks. This is modeled after what researchers believe are some of the native capabilities of the human brain.
As such, deep learning harnesses multiple processing layers to extract progressively better and more high-level insights from data. In essence, it is simply a more sophisticated application of AI platforms and machine learning.
How Does Deep Learning Tie Into Artificial Intelligence?
Many see deep learning as a subset of both AI and machine learning. They view AI as the overarching subject, and separate technological factors as follows:
- AI attempts to mimic human intelligence and behavior.
- Within the broad subject of AI, machine learning is a technique used to help computers learn using training that is modeled on results gleaned from large data sets.
- Within both AI and machine learning sits deep learning. This technique is essentially a more complex version of machine learning using neural networks.
Deep learning, then, should be considered as a subset of machine learning which, in turn, is a subset of AI.
“AI was traditionally limited to low-complexity tasks and simple decision-making until deep learning came,” said Petr Baudis, CTO and chief AI architect at Rossum. “Deep learning is a type of machine learning that is based on artificial neural networks.”
He explained that these mathematical structures – loosely inspired by the structure and function of the brain – can learn by example in a way that is similar to the way humans learn. The big shift is that deep learning finally allows machines to tackle problems of similar complexity to those humans can solve. It has been responsible for some of the most impressive AI achievements in recent years, added Baudis.
Also see: The Future of Artificial Intelligence
Deep Learning Means Larger, More Accurate AI Models
A lot of deep learning excitement concerns how to scale up large general models, technically termed foundation models. Deep learning-based models can generate images from text and even video from text. Thus, users of AI see deep learning as the key to being able to scale the current limits of AI modeling.
Algorithms are being used that dig deeper into data, preferences, and potential actions. They may even be capable of providing multi-faceted answers to complex situations and problems.
“Current models are limited in several ways, and some of the community is rushing to point those out,” said Peter Stone, Ph.D., executive director at Sony AI America. “It will be interesting to see what capabilities can be achieved with neural networks alone and what novel methods will be uncovered for combining neural networks with other AI paradigms.”
But, this isn’t some instant route to breakthrough insight. Deep learning platforms take time to support data analytics. They have to sift through a lot of data to spot patterns.
The model seeks to learn how things work in the real world and, thus, more accurately predict and provide accurate analyses. The good news is that advances in neural networking are speeding up the process. Some say the technology will soon reach the same level of maturity as data analytics.
Also see: The History of Artificial Intelligence
Deep Learning Needs Hardware and Software
It isn’t just software harnessing algorithms, AI, and modeling that is needed. The hardware is just as important. If you have huge datasets, you need the underlying compute infrastructure to deal with it in near real time. That means a whole lot of processing power. Graphics processing units (GPUs), for example, are needed, as are the latest memory fabrics that bring big memory to computing and AI platforms.
“Organizations can enhance their AI platforms by combining open source projects and commercial technologies,” said Bin Fan, vice president of open source and founding engineer at Alluxio. “It is essential to consider skills, speed of deployment, the variety of algorithms supported, and the flexibility of the system while making decisions.”
Software advances are vital, too. The latest trend is for deep learning workloads to be placed in containers to provide isolation, portability, more scalability, and address dynamic behavior.
“Organizations will find their AI workloads running on more flexible cloud environments in conjunction with Kubernetes,” said Fan.
Baudis noted the impact of big data on deep learning. He said that the deep learning revolution has been made possible by the availability of big data. Deep learning algorithms require large amounts of data to learn effectively. The recent proliferation of data sources such as social media, sensors, and online transactions has made it possible to train deep learning models on a scale that was previously unimaginable.
Some of the most impressive examples of deep learning in action include “large language models” that can generate legible text on various topics or generate realistic images from text prompts. Breakthroughs are being announced regularly.
These innovations are quickly trickling down into business and IT. For example, deep learning is now being used to generate high-accuracy text transcripts from audio recordings of business meetings and phone calls. This can be a huge time-saver for businesses, as it frees up employees from having to manually transcribe these recordings.
“Deep learning is also being used to automatically capture data from business documents with high accuracy,” said Baudis. “This can save businesses a lot of time and money, as it eliminates the need for manual data entry.”
Also see: Best Machine Learning Platforms
How is Deep Learning Evolving to Support AI?
Deep learning is evolving explosively, as it has done over the last few years. Baudis calls it by far the fastest evolving area of AI. He goes as far as to call non-deep learning areas of AI “niche” due to the progress being made on the deep learning front.
But, challenges lie ahead. The goals include:
- Learning how to achieve better-than-human performance in highly complex tasks. Progress has been made, but there is still a long way to go.
- Deep learning needs to acquire enough robustness so AI appliances can become trusted partners for humans; humans need to know they can rely on deep learning deployments.
“In some applications, this may mean the system is able to explain the reasoning behind its decisions,” said Baudis. “In others, it’s about the AI quickly learning from its mistakes and immediately adapting to do better next time.”
A practical example in the enterprise could be the document data capture system:
- First, low error rates simply mean a lot of manual work saved for human operators.
- Second, whenever a human corrects an AI mistake, the AI should not repeat the same mistake again.
This kind of behavior moves software adoption to a higher, far more functional level – suddenly it’s much less of an IT robot and more like onboarding a new coworker.
AI and Deep Learning Codependency
Deep learning and AI are deeply intertwined, at the foundational level. Rick Wagner, senior director of product management at SailPoint, explains this dependency.
“To provide artificial intelligence, processes and procedures must be learned, forming patterns derived and even nuance/variance patterns detected,” Wagner said. “Deep learning can be leveraged to analyze the exceptions when a human intervenes with AI decisions. Those exceptions can be analyzed for patterns which ultimately will improve the effectiveness of AI.”
Deep learning leads AI on to new innovations. On its own, artificial intelligence is the most beneficial when processes and procedures can be analyzed for patterns and exceptions to patterns. To achieve autonomous governance and autonomous identity lifecycle management, machine learning will play a pivotal role with deep learning providing a level of predictability of human exception decision processes, said Wagner.
“Learned processes and procedures must take place in order for artificial intelligence to be applied,” he said.
Shallow AI and Advanced Deep Learning
So, can standard AI be thought of as shallow AI, and is deep learning advanced, state-of-the-art AI? Some think so.
“Unlike shallower categories of AI, a deep learning approach involves the solution continuously ingesting large quantities of data to train a multi-layered deep neural network (DNN),” said Elaine Lee, principal data scientist at Mimecast. “DNN is designed to mimic human thought and understanding.”
Once DNN observes enough labeled data, it can successfully identify or categorize new, unlabeled data, or even a threatening anomaly, which users could be alerted to by an integrated AI instance. On the cyber defense side, for example, deep learning is being used to detect intrusions or malicious activity and classify malware and cyberattacks. It can also help organizations fortify their AI models, which are themselves vulnerable to attacks that wield misleading data.
“The integration of AI, deep learning, and algorithms is more important than ever in the world of cybersecurity as highly sophisticated threat actors increasingly weaponize their own AI-enabled malware,” said Lee. “Cyber criminals are increasingly relying on the use of social engineering attacks to propagate ransomware, employing AI to manipulate bots, poison data, create deepfakes, and assist with the scale and effectiveness of their attack campaigns.”
On the opposite side of the fight, organizations are shoring up their AI defenses to combat the evolving sophistication and complexity of attacks. According to Mimecast’s 2022 State of Email Security Report, 46% of organizations are already leveraging these technologies, with another 46% planning to follow suit. By mimicking human intuition, AI is helping to detect and prevent threats more effectively while reducing human error.
Also see: Top Digital Transformation Companies