What is Deep Learning? - dummies

By John Paul Mueller, Luca Mueller

What is deep learning? Deep learning is a subcategory of machine learning. With both deep learning and machine learning, algorithms seem as though they are learning. This is accomplished when the algorithms analyze huge amounts of data and then take actions or perform a function based on the derived information.

An understanding of deep learning begins with a precise definition of terms. Otherwise, you have a hard time separating the media hype from the realities of what deep learning can actually provide. Deep learning is part of both AI and machine learning. To understand deep learning, you must begin at the outside — that is, you start with AI, and then work your way through machine learning, and then finally define deep learning. The following guide steps you through this process.

deep learning
Deep learning is a subset of machine learning which is a subset of AI.

Deep learning starts with artificial intelligence

Saying that AI is an artificial intelligence doesn’t really tell you anything meaningful, which is why so many discussions and disagreements arise over this term. Yes, you can argue that what occurs is artificial, not having come from a natural source. However, the intelligence part is, at best, ambiguous. People define intelligence in many different ways. However, you can say that intelligence involves certain mental exercises composed of the following activities:

  • Learning: Having the ability to obtain and process new information.
  • Reasoning: Being able to manipulate information in various ways.
  • Understanding: Considering the result of information manipulation.
  • Grasping truths: Determining the validity of the manipulated information.
  • Seeing relationships: Divining how validated data interacts with other data.
  • Considering meanings: Applying truths to particular situations in a manner consistent with their relationship.
  • Separating fact from belief: Determining whether the data is adequately supported by provable sources that can be demonstrated to be consistently valid.

The list could easily get quite long, but even this list is prone to interpretation by anyone who accepts it as viable. As you can see from the list, however, intelligence often follows a process that a computer system can mimic as part of a simulation:

  1. Set a goal based on needs or wants.
  2. Assess the value of any currently known information in support of the goal.
  3. Gather additional information that could support the goal.
  4. Manipulate the data such that it achieves a form consistent with existing information.
  5. Define the relationships and truth values between existing and new information.
  6. Determine whether the goal is achieved.
  7. Modify the goal in light of the new data and its effect on the probability of success.
  8. Repeat Steps 2 through 7 as needed until the goal is achieved (found true) or the possibilities for achieving it are exhausted (found false).

Even though you can create algorithms and provide access to data in support of this process within a computer, a computer’s capability to achieve intelligence is severely limited. For example, a computer is incapable of understanding anything because it relies on machine processes to manipulate data using pure math in a strictly mechanical fashion. Likewise, computers can’t easily separate truth from mistruth. In fact, no computer can fully implement any of the mental activities described on the intelligence list.

When thinking about AI, you must consider the goals of the people who developed it. The goal is to mimic human intelligence, not replicate it. A computer doesn’t truly think, but it gives the appearance of thinking. However, a computer only appears intelligent when it comes to logical/mathematical thinking. Unlike humans, however, a computer has no way to mimic intrapersonal or creative intelligence.

What is the role of AI in deep learning?

Remember, the first concept that’s important to understand is that AI (artificial intelligence) doesn’t really have anything to do with human intelligence. Yes, some AI is modeled to simulate human intelligence, but that’s what it is: a simulation. When thinking about AI, notice that an interplay exists between goal seeking, data processing used to achieve that goal, and data acquisition used to better understand the goal. AI relies on algorithms to achieve a result that may or may not have anything to do with human goals or methods of achieving those goals. With this in mind, you can categorize AI in four ways:

  • Acting humanly: When a computer acts like a human, it best reflects the Turing test, in which the computer succeeds when differentiation between the computer and a human isn’t possible. This category also reflects what the media would have you believe that AI is all about. You see it employed for technologies such as natural language processing, knowledge representation, automated reasoning, and machine learning (all four of which must be present to pass the test).

The original Turing Test didn’t include any physical contact. The newer, Total Turing Test does include physical contact in the form of perceptual ability interrogation, which means that the computer must also employ both computer vision and robotics to succeed. Modern techniques include the idea of achieving the goal rather than mimicking humans completely. For example, the Wright brothers didn’t succeed in creating an airplane by precisely copying the flight of birds; rather, the birds provided ideas that led to aerodynamics, which in turn eventually led to human flight. The goal is to fly. Both birds and humans achieve this goal, but they use different approaches.

  • Thinking humanly: When a computer thinks as a human, it performs tasks that require intelligence (as contrasted with rote procedures) from a human to succeed, such as driving a car. To determine whether a program thinks like a human, you must have some method of determining how humans think, which the cognitive modeling approach defines. This model relies on three techniques:
  • Introspection: Detecting and documenting the techniques used to achieve goals by monitoring one’s own thought processes.
  • Psychological testing: Observing a person’s behavior and adding it to a database of similar behaviors from other persons given a similar set of circumstances, goals, resources, and environmental conditions (among other things).
  • Brain imaging: Monitoring brain activity directly through various mechanical means, such as Computerized Axial Tomography (CAT), Positron Emission Tomography (PET), Magnetic Resonance Imaging (MRI), and Magnetoencephalography (MEG).

After creating a model, you can write a program that simulates the model. Given the amount of variability among human thought processes and the difficulty of accurately representing these thought processes as part of a program, the results are experimental at best. This category of thinking humanly is often used in psychology and other fields in which modeling the human thought process to create realistic simulations is essential.

  • Thinking rationally: Studying how humans think using some standard enables the creation of guidelines that describe typical human behaviors. A person is considered rational when following these behaviors within certain levels of deviation. A computer that thinks rationally relies on the recorded behaviors to create a guide as to how to interact with an environment based on the data at hand. The goal of this approach is to solve problems logically, when possible. In many cases, this approach would enable the creation of a baseline technique for solving a problem, which would then be modified to actually solve the problem. In other words, the solving of a problem in principle is often different from solving it in practice, but you still need a starting point.
  • Acting rationally: Studying how humans act in given situations under specific constraints enables you to determine which techniques are both efficient and effective. A computer that acts rationally relies on the recorded actions to interact with an environment based on conditions, environmental factors, and existing data. As with rational thought, rational acts depend on a solution in principle, which may not prove useful in practice. However, rational acts do provide a baseline upon which a computer can begin negotiating the successful completion of a goal.

How does machine learning work?

Machine learning is one of a number of subsets of AI. In machine learning, the goal is to create a simulation of human learning so that an application can adapt to uncertain or unexpected conditions. To perform this task, machine learning relies on algorithms to analyze huge datasets.

Currently, machine learning can’t provide the sort of AI that the movies present (a machine can’t intuitively learn as a human can); it can only simulate specific kinds of learning, and only in a narrow range at that. Even the best algorithms can’t think, feel, present any form of self-awareness, or exercise free will. Characteristics that are basic to humans are frustratingly difficult for machines to grasp because of these limits in perception. Machines aren’t self-aware.

What machine learning can do is perform predictive analytics far faster than any human can. As a result, machine learning can help humans work more efficiently. The current state of AI, then, is one of performing analysis, but humans must still consider the implications of that analysis: making the required moral and ethical decisions. The essence of the matter is that machine learning provides just the learning part of AI, and that part is nowhere near ready to create an AI of the sort you see in films.

The main point of confusion between learning and intelligence is that people assume that simply because a machine gets better at its job (it can learn), it’s also aware (has intelligence). Nothing supports this view of machine learning. The same phenomenon occurs when people assume that a computer is purposely causing problems for them. The computer can’t assign emotions and therefore acts only upon the input provided and the instruction contained within an application to process that input. A true AI will eventually occur when computers can finally emulate the clever combination used by nature:

  • Genetics: Slow learning from one generation to the next
  • Teaching: Fast learning from organized sources
  • Exploration: Spontaneous learning through media and interactions with others

To keep machine learning concepts in line with what the machine can actually do, you need to consider specific machine learning uses. It’s useful to view uses of machine learning outside the normal realm of what many consider the domain of AI. Here are a few uses for machine learning that you might not associate with an AI:

  • Access control: In many cases, access control is a yes-or-no proposition. An employee smartcard grants access to a resource in much the same way as people have used keys for centuries. Some locks do offer the capability to set times and dates that access is allowed, but such coarse-grained control doesn’t really answer every need. By using machine learning, you can determine whether an employee should gain access to a resource based on role and need. For example, an employee can gain access to a training room when the training reflects an employee role.
  • Animal protection: The ocean might seem large enough to allow animals and ships to cohabitate without problem. Unfortunately, many animals get hit by ships each year. A machine learning algorithm could allow ships to avoid animals by learning the sounds and characteristics of both the animal and the ship. (The ship would rely on underwater listening gear to track the animals through their sounds, which you can actually hear a long distance from the ship.)
  • Predicting wait times: Most people don’t like waiting when they have no idea of how long the wait will be. Machine learning allows an application to determine waiting times based on staffing levels, staffing load, complexity of the problems the staff is trying to solve, availability of resources, and so on.

Moving from machine learning to deep learning

Deep learning is a subset of machine learning, as previously mentioned. In both cases, algorithms appear to learn by analyzing extremely large amounts of data (however, learning can occur even with tiny datasets in some cases). However, deep learning varies in the depth of its analysis and the kind of automation it provides. You can summarize the differences between machine learning and deep learning like this:

  • A completely different paradigm: Machine learning is a set of many different techniques that enable a computer to learn from data and to use what it learns to provide an answer, often in the form of a prediction. Machine learning relies on different paradigms such as using statistical analysis, finding analogies in data, using logic, and working with symbols. Contrast the myriad techniques used by machine learning with the single technique used by deep learning, which mimics human brain functionality. It processes data using computing units, called neurons, arranged into ordered sections, called The technique at the foundation of deep learning is the neural network.
  • Flexible architectures: Machine learning solutions offer many knobs (adjustments) called hyperparameters that you tune to optimize algorithm learning from data. Deep learning solutions use hyperparameters, too, but they also use multiple user-configured layers (the user specifies number and type). In fact, depending on the resulting neural network, the number of layers can be quite large and form unique neural networks capable of specialized learning: Some can learn to recognize images, while others can detect and parse voice commands. The point is that the term deep is appropriate; it refers to the large number of layers potentially used for analysis. The architecture consists of the ensemble of different neurons and their arrangement in layers in a deep learning solution.
  • Autonomous feature definition: Machine learning solutions require human intervention to succeed. To process data correctly, analysts and scientists use a lot of their own knowledge to develop working algorithms. For instance, in a machine learning solution that determines the value of a house by relying on data containing the wall measures of different rooms, the machine learning algorithm won’t be able to calculate the surface of the house unless the analyst specifies how to calculate it beforehand. Creating the right information for a machine learning algorithm is called feature creation, which is a time-consuming activity. Deep learning doesn’t require humans to perform any feature-creation activity because, thanks to its many layers, it defines its own best features. That’s also why deep learning outperforms machine learning in otherwise very difficult tasks such as recognizing voice and images, understanding text, or beating a human champion at the Go game (the digital form of the board game in which you capture your opponent’s territory).

You need to understand a number of issues with regard to deep learning solutions, the most important of which is that the computer still doesn’t understand anything and isn’t aware of the solution it has provided. It simply provides a form of feedback loop and automation conjoined to produce desirable outputs in less time than a human could manually produce precisely the same result by manipulating a machine learning solution.

The second issue is that some benighted people have insisted that the deep learning layers are hidden and not accessible to analysis. This isn’t the case. Anything a computer can build is ultimately traceable by a human. In fact, the General Data Protection Regulation (GDPR) requires that humans perform such analysis. The requirement to perform this analysis is controversial, but current law says that someone must do it.

The third issue is that self-adjustment goes only so far. Deep learning doesn’t always ensure a reliable or correct result. In fact, deep learning solutions can go horribly wrong. Even when the application code doesn’t go wrong, the devices used to support the deep learning can be problematic. Even so, with these problems in mind, you can see deep learning used for a number of extremely popular applications.