Affective Computing Among Individuals in Deep Learning

Artificial Intelligence is a set of algorithm systems designed to think, sense and act like humans. This allows the definition of artificial intelligence as the concept of an agent that achieves what a person intended without human intervention. For example, robot vacuum cleaners that act differently depending on the type of room and the cleanliness of the room, or artificial intelligence washing machines that optimize the washing method according to the amount and type of laundry can also be seen as a kind of artificial intelligence as agents that achieve human goals on behalf of humans. Artificial intelligence, which was first created in 1956 at the Dartmouth Conference of 10 mathematicians, scientists and others, will reach the current stage of development through several evolution and decline. In the early days, attempts were made to translate human problem-solving logic into computer language. As artificial intelligence draws attention, machine learning and deep learning, which are core technologies, have also emerged as important keywords [7-9].

Deep learning is a kind of machine learning. The most important factor in the evolution of artificial intelligence is “learning.” Learning here is ‘a series of processes that create a system that extracts and classifies characteristics in any way’ and the choice of characteristics determines the performance of learning to recognize patterns and reduce error values. Machine learning refers to how to improve the performance of a particular task through experience. Due to the fact that there are so many different factors in the real world, and the exceptions that cannot be explained by the general rules, classical artificial intelligence required endless modifications and supplementation of the infinite cases in the application of the real problem in Fig. 1.

Fig. 1. Definition of Artificial Intelligence.

Download Original Figure

Artificial intelligence was first established by Ellen Tuning. Data-based artificial intelligence and self-learning artificial intelligence can be divided.

Also, the biggest difference between machine learning and machine learning is that deep learning can learn the data to be used for classification, while machine running must provide learning data manually, which is the biggest difference between deep learning and machine learning.

2.2.1. Machine Learning

The term machine learning has been around for a long time, but in the 2010s, it became a popular term for algorithms that use artificial intelligence. In terms similar to machine learning, deep learning, data mining and pattern recognition are also used, and it is not easy to understand what the difference is unless you are a related major, and it is easy to misunderstand that it is something different from machine learning. In conclusion, there is a slight difference, but I think intersecting is a very large and very similar meaning. There is no denying that there is a strategic purpose to make a new term appear trendier and sell better, even though there is no big difference in science or technology. Machine learning technology, as the term machine learning, is based on machine learning. But if you need data to learn [10].

The machine-learning algorithm can be classified by several criteria depending on how it works, with the biggest criteria being Machine Learning and Unsupervised Learning. First of all, leadership learning is the concept of giving and learning test genealogy with an answer sheet. The existence of the answer itself means that you look to guess the answer later on. Therefore, map learning is an algorithm used to find answers. An example of a map study is to give them multiple pictures and tell them to distinguish between a dog and a cat. At this time, they give a reply in advance to see which one is a picture of a cat or a dog. The data that gives you this learning is called training data, and the answer to training data is called labelling. Such labelled training data will allow the map-learning algorithm to learn in its own way and distinguish whether the photo coming in is a personal or a cat. This trained algorithm has become a model that can distinguish a dog from a cat. This model will then be able to properly distinguish cats and dogs even if unlabeled (unanswered) data is received in the future as shown in Fig. 2. Unlabeled data is called test data and how accurately the answer is correct will determine the performance of the algorithm. If you search “dog muffin AI” on Google, there are a lot of pictures of a mix of chihuahua and muffin to tell if you look at the computer as a joke, but these days, it is said to be extremely accurate.

Fig. 2. Process of Machine Learning.

Download Original Figure

On the other hand, giving and learning labeled (undescripted) data is called unguided learning. Because it does not give an answer, non-guided learning is not taught for the purpose of guessing the answer. However, it serves to provide useful information on problems that don’t give answers, such as grouping together what data are similar, or determining which nature defines the data well. This refers to the situation in which pictures are entered for the purpose of distinguishing pictures of dogs and cats as before. Non-guided learning does not tell which picture is a dog or a cat. That is, it is not labeled. However, the algorithm can identify which pictures are similar to each other by looking at the pictures. So, the cat-like photos are divided into one group and the dog-like photos into two groups.

This distinction can be made, but it cannot tell us who this group is. It is up to a person to determine whether the distinction has been made well and how the group is defined. In the above case, the algorithm does not answer that the cat picture is a cat and the dog picture is a dog, but since the cat picture and the dog picture are different groups, one can see the results and the other can be judged as a cat picture and the other as a dog picture. This non-directory learning cannot be used to inform the final answer, but it is widely used as it helps people make decisions by giving them useful information to identify the characteristics of the data.

2.2.2. Deep Learning

Deep Learning is a machine-learning model inspired by a biological brain circuit. Like humans, what do humans think, think, and store memories. The human brain consists of 100 billion neurons and 100 trillion synapses. When a synapse sends an electrical signal to a neuron, the neuron responds to this electrical signal and passes it to the next neuron. These neuron-synapses are a brain that allows humans to understand, think, and think in high dimensions. Deep running was designed with inspiration from these brain signals. The imitation of neurons in deep learning is called Perceptron, which was first proposed by Rosenblatt in 1958. Perceptron accepts input of x1 to xn and calculates the weighted sum of these inputs and predicts the final y by substituting the activation function. These basic perceptories alone are difficult to solve [11-12].

At this time, the characteristics of a cat such as pointed ears and brown fur are referred to as re-preservation. Adding a Hidden layer to a deep learning model means adding one more layer to extract this presentation. The deeper the layer is stacked, the more capable it is to combine the basic features to extract high-dimensional features one step above. The difference between machine learning and deep learning is this. The classic machine-learning algorithm tries to map directly to input features (e.g., the molecular unit of a cat) to determine whether they are cats. As the term Deep Learning is found, the algorithm itself is embedded in the process of re-preservation learning, which extracts useful information from a very abstract molecular cat and converts it into a high-dimensional feature that can judge a cat. Deep learning is challenging the human-specific realm, an area that is easy to do but difficult to formalize and explain, an intuitive and high-dimensional one can do-it-your Because deep learning is implemented only through data and through numerous parameters that locate the rules of data, it requires a lot of data.

For example, let’s create a deep learning system to find cats. It takes a lot of cat pictures to figure out if the machine is a cat. It will not be easy for the machine to distinguish the difference in the small details from the cat to distinguish the fox and the cat that look similar to the cat. Therefore, no matter how good a deep-learning model is implemented if numerous datasets are not supported. On the other hand, it is true that deep learning is inspired by brain science, but it does not necessarily apply brain science knowledge as it is [10]. Deep learning has a rather engineering aspect. In fact, even without a biological level, it is useful if it helps improve performance (Fig. 3).

Fig. 3. Deep Learning Process.

Download Original Figure

Affective computing can be divided into six application areas. They include learning, robotics, customer-centered management, image recognition and language recognition (Fig. 4).

Fig. 4. Model of Emotional Computing.

Download Original Figure

The OCC model was established in the late 1980s at the University of Illinois to study sensibility through psychological methods. The OCC model is the emotional evaluation model of the three researchers, Ortony, Collins and Clore, which is a research method that types into emotional clusters resulting from similar causes rather than analyzing all the emotions one can express. The principle of this is that these emotional types are evaluated and separated by three components (event, object, agent) and are classified into the final 28 emotional types. Robots have been developed to generate characteristics that can affect the generation of emotions using information from internal and external sensors, and based on this, to express emotions in various forms. Various emotional models have been used to generate robot sensitivities to internal and external stimuli. Emotional models can be divided into two main types [16-17].

First, there is a continuous multi-dimensional emotional model that defines emotional space in two or three-dimensional space and when mapping. A representative model for this is the 3D emotional model used by WE-4RII and Kismet. By placing a characteristic axis in the 3D emotional vector space and calculating the value, it is possible to maintain a continuous change of emotion and to provide a variety of emotional expressions. The calculations are high and somewhat complicated.

Second, it can be divided into discrete emotional models with tree-shaped structures predetermined to satisfy the conditions. A representative model for this is the OCC model bought in OZ project. While the OCC model is easy to implement, it is not easy to express a continuous change in sensitivity. Emotions generated through these emotional models should be expressed in the form of faces, gestures, emoticons, sounds and text. In this chapter, we will explore the research trends for typical emotional related robots and the creation of emotion and the expression of row movements in each robot. Emotional inference and expression is the recognition of emotion and the prediction and expression of emotional state through inference about the situation in which it occurs. To this end, the University of Illinois has established an OCC model since the late 1980s, conducting a study on sensibility. The OCC model relates to psychology and is an emotional evaluation model of three people, Ortony, Collins and Clore. Instead of trying to describe all of the emotions that a person can express, this defines the emotional clusters that are generated and distinguished by similar causes as the emotional type. For example, agony involves emotions such as sadness (sad), feeling (distraught), and love’s pain (lovesick) depending on the difference between the cause and so on. The three factors that evaluate emotional types are events, objects, and agents. Events refer to actions related to the objectives of agents, objects represent different objects that exist in equal capacities, and agents have emotional types depending on the event and the objects. Depending on these three factors, emotional types are divided into three categories and finally into 28 emotional types. The emotional evaluation process consists of an evaluation of the satisfaction of the events associated with the agent’s goals, the degree of approval for the agent’s actions or actions, and whether the agent’s attitude likes the objects involved. This process allows the OCC model to calculate the generation and intensity of sensibility, and to generate specific sensitivities according to the interpretation of a given situation. Typical sensibilities defined in the OCC model include joy, anguish, hope, fear, pride, shame, admiration, shame, anger, gratification, and regret.

Emotional inference means emotional inference according to events that occur in multi-agent systems as inter-agent interactions. The agent’s emotional reasoning for this depends on the interpretation of the events that have occurred, and this interpretation affects everyone’s goals, expectations, learning, etc. Each has different internal models depending on its own perspective, and it is now known that there is a total of 24 emotional forms. The above-mentioned emotional/psychological cognitive skills are recently included in the facial expression recognition field, which belongs to bioinformatics. This technique is used to recognize the emotional state of a person by reading facial the technique that accounts for the largest portion of the rain is the technology that recognizes facial expressions. The changes in facial expressions represent different emotions 2 due to changes in the facial muscles of the human face (eyes, eyebrows, mouth, etc.) [18-19].

Sensitivity using facial expressions is recognized through the transformation of characteristic elements derived from basic emotional expressions such as anger, sadness, anaerobic error, joy, surprise, etc., and studies are under way to take into account various changes in face shape and state environment that are frequently changed by glasses, hairstyle, facial expressions, etc. Active Appearance Model (AAM) recognition, which automatically extracts the position of eyes, nose, mouth, etc. from the input by modeling the shape and appearance of the face as the main component analysis (PCA), and the input phase through the characteristics of the face with GW (Gabor Wavelet) are developing as major technologies. The Facial Action Coding System (FACS) categorizes the measurement of emotional state by AC (Action Units) to database the emotional state of changes in facial expressions. Facial image-based facial expression recognition technology is a technology that detects subtle facial changes and recognizes the emotions of the current employees and has drawn attention in the human-computer interaction field [21], [22]. In addition, a database was implemented to enable the recognition of facial expressions only for Koreans who excluded racial effects based on the standard image of the Korean public using facial expression training technology. When big data and artificial intelligence become an issue recently, Seo et al. [20] based emotional or emotional technology is also being developed to build AI, create cognitive models, and recognize emotions through facial learning of each emotional expression. Performance improvement and utilization of facial expression training technology are expected.

With pet robots and portable robots in the spotlight, interest in the interaction between humans and robots through contact is growing. In robots, contact sensors have been used for such purposes as measuring location for autonomous driving, avoiding obstacles, balancing the body when walking on two-legged robots, or controlling movement when a robot manipulates objects with its own hands. The Communication robot project aimed at measuring human sensitivity through physical interaction and developing robot that can interact naturally with people through contact, suggesting that as a result, a way to recognize hitting behavior, light flick behavior and stroking behavior using I-SCAN contact sensors developed by NITTA Corp., which can measure the pressure of contact. The Huggable Project announced a Sensitive Skin so that the robot could feel contact with all of its body, with the aim of developing a portable robot that can be used for emotional therapy. The Sensitive Skin consists of a QTC force sensor and a heat sensor. And these contact sensors have produced an initial study that recognizes nine different contact behaviors that a robot can feel in its users.Shigeki Sugano defined PIFACT items, saying that robots should determine the contact conditions felt by robots in order to coexist with humans in their daily lives. He developed the human-robot PIFACT definition system and evaluated the system developed by applying it to WENDY. To detect contact, an FSR sensor, a force sensor, is attached to the WENDY shoulder and arm area.

Humans like and dislike someone and can create and think new things. In comparison, artificial intelligence learns, acts and determines based on numerous data. With artificial intelligence now entering deep into human life as a key technology in our society, the issue of social and ethical value judgment on artificial intelligence has become an important issue. Recently, the issue of choosing between large and small is being considered in autonomous driving cars, nursing robots, and former two-robot or drones. U.S. government-affiliated research instruments have supported various research projects related to artificial intelligence ethics. Ronald Arkin of Georgia Institute of Technology in the U.S. has been studying the ethics of war robots. The purpose of this study is to reduce the cost of non-combatants in war, and to make robots have the ability to make ethical judgments to assess opponents’ sins. However, robots lack emotions, so it is difficult to properly judge human dignity and see it as a matter of comparative judgment of values in which mercy can be involved depending on circumstances (in Fig. 5). Above is talking about emotional computing. Starts with Emotion Engine and can be sensitively predicted for emotion, emotion, etc. Describe this architecture. Such UNC Berkeley’s “Center for Human-Competitive AI” is conducting research to enable AI systems to provide essential services to society in a manner consistent with human values.

Fig. 5. Tactile Sensitizer Architecture.

Download Original Figure

In other words, using the reverse reinforcement learning approach, the development of artificial intelligence services to observe human row movements and learn human values. They are conducting research on whether self-driving cars have enough capacity to detect the conditions of surrounding roads, whether they can make the right decision when they are faced with unexpected events and unfamiliar objects or situations, and whether they are safer than human-driving vehicles. So the field of study in this paper should produce an immediate and automatic emotional response to certain patterns of external stimuli. Emotional generation is like synthesizing emotions in a machine. To synthesize emotions is to make it seem like a machine has real feelings, or to have internal mechanisms similar to those of humans or animals. Emotional generation requires, for example, the process of synthesizing various tones and speeds depending on the emotion or situation or synthesizing necessary actions to represent the state of emotion. For example, in an emotional agent, synthesis determines what state the machine should be in. These studies have long been carried out. In particular, emotional synthesis of speech or facial expressions has been under way for almost 30 years and is now being studied.