Emotional System for Military Target Identification

The thought of man-made systems and machines having emotions sounds like science fiction, however, few decades ago the idea of machines with intelligence seemed also like fiction, but today we are developing intelligent machines and using them successfully in different applications. Would we accept the idea of machines that could “feel”? What role would emotions play in machine learning and decision making? How can we model artificial emotions within intelligent systems? Can a machine’s decision capability be improved if it had emotions? These are questions one may ask when hearing that machines may have emotions, albeit artificial emotions. In this paper, we discuss these questions, and review briefly some of the latest works on modeling emotions within intelligent systems; including our own model that is based on an emotional neural network (EmNN). The EmNN has emotional neurons, weights, and two embedded emotional responses; anxiety and confidence. These emotional parameters are updated during task learning, and used during decision making. The paper will also present an application of the EmNN to military target identification, in addition to discussing the potential of using the emotional system to improve information exploitation. 1.0 INTRODUCTION In our daily lives, the amount of information we receive, perceive and then react to is tremendous. Much of our reaction to input information is formed into decisions that we make. What makes us act in a certain way, or decide for, or against an act has its roots back in our previous experiences. Whether we choose anchovies or pineapple on our pizza is a decision we make based on a previous experience with these flavors that leads us to decide what to have. Such a simple process of deciding on our favorite topping is as important and complicated as the more critical decisions we need to make throughout our lives. All decisions involve a learning process, resulting in association, classification and then decision making. During the learning process information which can take different forms, is exchanged between our natural sensors and the main processor; the brain. This sounds very technical and parallel to describing a computing system. However, computing systems lack one aspect of a human processing system; emotions. Over the past decades machines and systems have been developed and deployed to aid us in decision making and taking action; spanning application areas from simple electronic toys, industry and manufacturing, intelligence and security, to more complicated tasks in medicine, military applications, and space navigation. As the creators (designers) of these machines, we aim to assure that the action or the decision taken by the machine is correct and complies with our way of making a decision. In simplistic terms, we tend to create machines that would make the decisions on our behalf, and as information age progresses, and powerful high-tech systems grow even faster, our expectations of the machines are increasing. Most of the systems that we develop, and use to make decisions on our behalf, do not go through the learning process and the experiences that we possess. With the exception of some artificial intelligent systems that could interact with their input stimuli, and adapt their output or decision accordingly, systems Emotional System for Military Target Identification 18 2 RTO-MP-IST-087 UNCLASSIFIED/UNLIMITED UNCLASSIFIED/UNLIMITED and machines rely entirely on a set of commands that we provide to govern their actions, and this has been fine until our demands started requiring more complicated decisions by machines. Therefore, more and more intelligent systems are being developed, based in particular on neural networks which form the brain of a machine. These systems imitate our learning process and decision making by repeatedly exposing a neural network to examples of input information and its corresponding output, response, action or decision; this process models the previous experience process in humans. The neural network-based systems have been popularly used, and have shown success in various application areas, where association, classification and decision making can be obtained based on accumulated memory of past experiences. Despite the success of such intelligent systems, there has been a major and vital difference between a human decision-maker and a machine decision-maker, namely emotionwe have it, and machines do not. The idea of machines with affection or feelings is controversial, and some works expressed doubt about this idea [1,2], however, the concept of machines with emotions has lately attracted the attention of many researchers, and is currently gaining momentum with novel architectures emerging to artificially model emotions in one way or another. Recent definitions of emotion have either emphasized the external stimuli that trigger emotion, or the internal responses involved in the emotional state, when in fact emotion includes both of those things and much more [3]. The effective role of emotions on cognitive processing, learning and decision making in animals and humans has been emphasized by several researchers [4-8]. Emotions play an important role in human decision-making process, and thus they should be embedded within the reasoning process when we try to model human reactions [9]. Although computers and machines do not have physiologies like humans, information signals and regulatory signals travel within them; there will be functions in an intelligent complex adaptive system, that have to respond to unpredictable, complex information that play the role that emotions play in people [1]. Such computers will have the same emotional functionality, but not the same emotional mechanisms as human emotions. We may think of machine emotions as machine intelligence; we do not expect machines to “feel” the way we feel, but we could simulate machine emotions just as we do machine intelligence [9]. There have been examples of research works that attempted to incorporate emotions in machines in one way or another [9-20]. It was concluded from these works that if emotions such as anxiety, fear, and stress are included in systems that aim to simulate the human behaviour in certain circumstances, the system will be more user-friendly and its responses will be more similar to human behaviour. Other recent research works suggested the use of emotional components within neural models and control systems. For example, Abu Maria and Abu Zitar [21] proposed and implemented a regular and an emotional agent architecture which is supposed to resemble some of the human behaviours. They noticed that artificial emotions can be used in different ways to influence decision-making. Gobbini and Haxby [22] proposed a model for distributed neural systems that participate in the recognition of familiar faces, highlighting that this spatially distributed process involves not only visual areas but also areas that primarily have cognitive and social functions such as person knowledge and emotional responses. Coutinho and Cangelosi [7] suggested the use of modelling techniques to tack into the emotion/cognition paradigm, and presented two possible frameworks that could account for their investigation, one of which explored the emergence of emotion mechanisms. Most of these previous attempts on incorporating emotions in to machine learning have shown successful results, and provided a positive trend to developing machines with emotions, albeit simulated. Lately, we proposed an emotional neural network (EmNN) which was based on the novel emotional back propagation (EmBP) learning algorithm [23], and used it to solve a facial recognition problem. In other works [24,25], we explored the potential of using emotional neural networks in different tasks, such as Emotional System for Military Target Identification RTO-MP-IST-087 18 3 UNCLASSIFIED/UNLIMITED UNCLASSIFIED/UNLIMITED more complicated face recognition tasks and in blood cell identification. The difference between an emotional system and the more traditional approaches; including intelligent systems, is the simulated artificial emotions of the system. These additions have several advantages over traditional approaches. The embedded artificial emotions narrow the gap between humans and systems; thus instead of “human and machine interaction” we have “human and human-like machine interaction”. The closer and more coherent communication of information between humans and emotional systems has the advantage of faster communication, since both systems (human and machine) have emotions. This is not the case with traditional systems, where often a human operator perceives information and makes decisions which could differ due to his/her emotions. In this paper, we present the emotional neural network (EmNN) and describe its emotional parameters. The EmNN will also be applied to identify potential military targets, such as navy ships, helicopters, jetfighters, tanks and other assorted military vehicles. One of the aims of this work is to mimic the way a human would recognize these targets, by: firstly, using different images of targets with random orientations, angles, and backgrounds, secondly, avoiding complicated image pre-processing phases, and using only global image pattern averaging to simulate a human’s “glance” or quick look at a target image, and finally, using the EmNN to perform the target identification, by repeatedly exposing random target images to the network during its training phase; this process simulates the human “getting familiar” with the objects, without the need to look into edges, local features, angles or colors of a potential target. The paper is organized as follows: Section 2 presents the EmNN and describes the differences between conventional neural networks and emotional neural networks. Section 3 presents the application of the EmNN to military target identification, describing the image database and the EmNN topology. Section 4 describes the implementation results. Finally, Section 5 concludes the work that is presented within this paper and suggests further work. 2.0 EMOTIONAL NEURAL NETWORKS Neural networks in intelligent systems model the structure and the function of a biological brain, albeit at a much smaller scale. The function of a neural network is to learn to associate certain inputs with already known (to the designers) outputs; this is called training the neural network. Training is completed when the network reaches an acceptable minimum error value, or the number of times (called epochs or iterations) that the network is exposed to the pairs of input/output examples. During training, the memory of a neural network is represented by a set of values, called synaptic weights, which are updated in each training epoch. Once the network learns (completes training) the final values of the synaptic weight are considered the “experience” memory, and are thus used for testing or implementing the trained system. Applications vary from association, to classification, identification and pattern recognition. When compared to conventional neural networks, the emotional neural network (EmNN) has additional emotional neurons, two emotional parameters (anxiety and confidence), and emotional weights. The emotional parameters are updated during learning, and the final emotional weights are used together with the network’s conventional weights to make decisions. The incorporation of the simulated emotionality within a neural network structure aims at further improving the network’s learning and decision making capabilities. The EmNN model is based on the emotional back propagation (EmBP) learning algorithm [23]. The additional emotional coefficients (anxiety and confidence) are updated each iteration or epoch during the learning process, and their values are used to update the emotional weights associated with the emotional neurons. Our hypothesis when suggesting such a model for an emotional system is that when we learn a new task, our anxiety level is high, while our confidence is low at the beginning of learning. Emotional System for Military Target Identification 18 4 RTO-MP-IST-087 UNCLASSIFIED/UNLIMITED UNCLASSIFIED/UNLIMITED Over time, training and with positive feedback, our anxiety level decreases, while our confidence level increases; thus making better decisions in less time. Figure 1 shows an emotional neural network structure that can be used for pattern recognition and object identification based on presenting the network with input images. There are two emotional neurons feeding the hidden and output layers in the EmNN. These emotional neurons differ from the normal neurons in that they are non-processing neurons which receive global average values of input images, rather than segments or pixel values from that image, also their associated emotional weights are updated using the two emotional coefficients, rather than the conventional learning and momentum rates. In practice, what the two emotional parameters mean is that when the emotional neural network is trained, and as the epochs progress, one term (anxiety level) tells the system to pay less and less attention to the derivative of the error of the training pattern using all nodes as the input average value of the samples of the training pattern, while the other term (confidence level) tells the system to pay more and more attention to the previous change it made to the weights, which is some sort of an increasing inertia term to modify the level of change from one pattern to the next as the training epochs progress. Full description of the learning algorithm of the emotional neural network can be found in our recent work [23]. However, we describe in this paper the definitions of the two emotional responses which are updated during the machine learning process. The anxiety level of the EmNN is represented by the anxiety coefficient (μi) value, which is defined as: i AvPAT i E Y + = μ (1) where YAvPAT is the average value of all presented patterns to the EmNN in each iteration; and is defined as: