Issues Limiting the Possibilities of Artificial Intelligence - Case Study Example

Neural Networking in Artificial Intelligence By and Introduction Artificial intelligence has gained much impetus in the recent past. It can be defined as an aspect of computer science that aims at improving the power of computers to make them think, act and perform like humans. It can also be perceived as a branch of computer science concerned with the creation of brainy machines that have the capabilities to mimic human intelligence, thereby reacting like man. Such aspects in contemporary computing as machine learning, voice recognition, use of computers to solve problems, planning and many more are the building blocks of artificial intelligence. The continued researches on the field all aim at coming up with intelligent systems and thus a very crucial aspect of technology. There are however some major issues limiting the possibilities of artificial intelligence research to develop a system that fully takes to human beings. Programming a computer to acquire such traits as knowledge, perception and reasoning could be very tedious. Making machines that can move and manipulate objects, learn and solve problems is close to impractical. To make such machines, a lot of information related to the day to day functioning of human beings need to be fed. To implement the aspects of intelligent systems, instilling reasoning and common sense in machines is required which is close to impractical. It can be a very tedious approach that can take you up to more than a century. Machine learning aims at producing system that can solve unrelated and diverse situational issues. Most artificial intelligent systems are built with a purpose in mind, for instance coming up with a gaming system that can outdo the best champions of say chess. This is however inevitable whenever creating a system intended to perform specific duties. To avoid the idea of developing task oriented artificial intelligent systems, computer professionals gained a lot of concerned on including artificial neural networks. The intelligent human brain is comprised of a network of neurons that make perceptions and connect depending on external stimuli. Neural Networks This is a major component of the recent developments in the artificial intelligence branch of computer science. Artificial neural network according to Gomes et al,.(2011) is a paradigm for information processing that seeks to mimic the manner in which such biological nervous systems as the human brains process information. The novel arrangement, as is in the works of Savaki (2006), of information processing systems forms the basic element of this model. These networks are composed of massive neurons which resemble processing elements which works in unanimity to solve definite issues as dispensed. The system are said to learn best through examples just as people do. Configurations for the Artificial Neural networks are made for individualized applications through the process of learning. About Neural networks Simulations of neural networks are a recent idea that has gained much escalation over the past decade with professional involvement in research and funding. They possess a notable ability to descend an implication from imprecise and complicated data, and can thus be resourceful in extraction of patterns for detection of complex trends that even humans cannot detect. Trained neural networks are therefore termed as “experts” in the context of the information being analyzed (Arinkinet al., 2013). The “experts” are therefore resourceful in making projections whenever exposed to new situations. These systems are said to possess the ability to learn adaptively through experiences and the data fed while training. They offer real time operations as different computations can be performed at per. Various models are available which seek to explain the concept of neural networks in artificial intelligence (Savaci, 2006). The models are categorized broadly in two; the biological and the mathematical model. The BiologicalModel In the mid-1940s, McCulloc and Pitts came up with models that simplified the structure of the neuron that gave birth to the artificial neural networks. The neurons were represented biologically as a networked abstract component made of circuits with an ability to execute computational tasks (Rodgers and Kabriski, 2007). The manner in which the biological neuron functions formed the basis model in which the artificial neuron was instituted. Neurons can be defined as the basic blocks of the human nervous system responsible for generation and transfer of signals within the body of a human being. Each neuron is taken as a discrete unit housing different processes which take place within it. A biological neuron comprises of four different major parts; the axon, cell body, the nucleolus and the axon (Rojas, 2006). The cell body acts as the home for the nucleolus providing it with replenishment and an environment for its functionality. The dendrites act as bio-receptors of signals sent from other neurons. The axons on the other hand acts as conductors of electrical signals that acts to enhance detection of actions. The electrical signals used by the neurons in conveyance of information from the brain are all similar. The brain detects the information received and interprets it on the basis of the path it was received. All the patterns of signals received by the brain are analyzed and the intended message interpreted. The axon gets insulation from myelin, a fatty tissue (Rojas, 2006). Some parts of the axon however are not insulated and are said to form what is known as the Nodes of Ranvier. They aid in regeneration of signals as they travel down through the axon. Different neurons get into contact with each other at points known as the synapse. Due to presence of a cleft, the neurons do not get into contact and signals are therefore exchanged through chemical processes. The membrane has potential that results from imbalances in the concentration of ions, which leads to generation of the electrical signals. The neurons could be polar, bipolar or multipolar based on the number of processes they can carry out. The Mathematical Model The artificial model achieved from the biological neuron should be keenly evaluated during its evaluation. This is due to the fact that the synapses of the neuron are modeled as weights. This synapse interconnects the neural networks, adding strength on to the network. In the case of artificial neurons, the neuron is represented by a number, which later represents a synapse. Where negative weight is present, this acts as an inhibitory connection (Gomes, Ludermir and Lima, 2011). Conversely, plus values elect excitatory links. These components act as a representation of the actual activities taking place in a neural cell. All the values input are added up and computed using the weight value in a process referred to as linear combination. At last, the amplitude of output is controlled by an activation function. The figure below acts as a mathematicaldescription of the process. From the model the interval activity of the neuron can be derived using the formula: This could therefore result to an output of the neuron yk, which is a form of an activation function. Activation functions In the neural networks, there exist functions that specify the output from a neuron into an input and are known as activation functions. The functions are said to possess squashing effects and the neurons therefore act as switches producing the binary ‘1’ in their full activation form and the binary ‘0’ when their activation status is null. A wide range of these functions are supported by neural networks, with only a few being applied by default. The rest of the functions are however used for customization functions. Some of the activation functions available in neural networks according to Gomes and Luderma (2012) include identity functions, hyperbolic functions, logistic functions, softmax function, exponential functions, sine, square root and the unit sum functions among others. Function Definition Range Identity X (-inf,+inf) Logistic (0,+1) Hyperbolic (-1,+1) -Exponential (0, +inf) Softmax (0,+1) Unit sum (0,+1) Square root (0, +inf) Sine sin(x) [0,+1] Ramp [-1,+1] Step [0,+1] Table 1.1: Mathematical definition of activation Functions (adopted from ELECTRONIC TEXTBOOK, 2002). The identity function is common in a number of networked systems and is usually characterized by its activation levels getting directly passed on as output. The logistic function on the other hand is represented by use of a sigmoid shaped curve with a range of (0,1) as the output. Where the range of the output in the sigmoid curve lies between (-1 and +_1), the sigmoid curve is said to be a hyperbolic function. In such cases, performance of the function is highly rated due to its symmetry. In the case of multilayered perceptron, customization using this function is more preferred, more so in the event of hidden layers. The exponential functions are mostly preferred for outspread units (Gomes, Ludermir and Lima, 2011). Combining negative exponential functions of activation with radial functions amounts to production of the Gaussian model. This is a bell shaped model whose standard deviation is achieved through the prescription below. The letter “d’ denotes the deviation of the unit stored in a threshold of units. The unit sum function acts as a normalization tool for converting outputs in a manner that they can interpreted in probabilistic forms. The softmax function works with the normalized data to produce a sum of 1.0 across the activation layer. Most common in the output layers to enable interpretations of outputs as probabilities. The square root is employed in transforming the distance activation functions that have been squared in a network to present the actual values in the output. The sine function is only applicable where data that has been distributed radially requires recognition. The Ramp has been presented as a unit linear sigmoid function whose training performance is poor, but rather with faster execution rates. Distributed representation framework A typical artificial neural network is composed of a pool of simple units of processing that communicate through signaling each other over massive numbers of weighted connections. A number of major aspects of a distributed system model that distinguishes it includes; possession of processing units of cells and neurons, an activation state for each unit (yk) which equates the output, rules for propagation, units connections, a function Fk for activation (this determines new levels of activation on the basis of effective input), an external input for each single unit, methods to be employed in gathering information and finally the environment under which the system ought to operate. Neural Network topologies Neural network topologies define the manner in which nodes and interconnections are arranged in a system. The choice of a topology is dependent on the problem at hand and the ease with which such a problem can be solved (Gomes, Ludermir and Lima, 2011). The pattern in which nodes are interconnected form networks that are broadly classified as feed-forward and recurrent neural networks. The Feed-forward topology contains nodes that are arranged hierarchically in a manner that input flows in one order from layer to layer without acceptance of any form of a feedback. In this model, data processing extends to multiple units in the following layers but cannot be sent back to previous units. It can be well understood from the illustration below. The recurrent topology on the other hand is designed in a manner that feedback links between the units is possible and in any direction without restrictions. Information can be stored by nodes in their output with each state being dependent on previous ones in the topology. The Adaline and perceptron neural networks forms precise examples of feed forward topology while the Hopfield and Kohonen models fall under the recurrent category. Training of artificial neural networks Afterestablishinga structure for a network of particular application, it becomes ready for training eventually. The training process starts with the random selection of the initial weights, after which learning commences. Three main approaches are used to train a neural network which includes the reinforced learning, supervised and unsupervised learning. Supervised learning entails provision of input data manually with an aim of producing desired results or output using the input data. Figure 1.1: an illustration of supervised learning (adopted from techonthenet-blog.com) In the case of unsupervised training, the network is left with the task of making sense of the inputs to produce desired output without any external control. The biggest number of networks relies on supervised training. The latter is applied mostly in the performance of tasks where an expectation of desired output prevails. The system is left with the mandate of exploring the underlying data structures and relations to organize them into categories for easier execution. Figure 1.2: an illustration of unsupervised learning (adopted from techonthenet-blog.com) The third form of neural machine learning isreinforcement learning. This form of training entails mimicking of the manner in which human beings learn as they interact with the physical environment (Li and Zhan, 2014, pp. 1683-1695). Inputs are presented to the network without any expectations for desired output. Where the desired output is generated, all the connections associated with the output are highlighted and strengthened. In the case of the rear, such connections are weakened thereby making this approach a perfect choice for instilling control. Modifying patterns of connectivity of Neural Networks In the different paradigms of learning, whether supervised or supervised, there is resultant regulation of the weights of interconnection between the different units of a neural network. These result from some sort of rules for modification. Where there are two different units that are simultaneously active, need for strengthening the interconnection prevails. This is derived from the suggestions made in the works of Hebb (1949), the Organization of Behavior. If for instance an element j is input to another unit k, the most basic form of Hebbian learning principles recommends modification of the weight under the operation in which the symbol ϒ represents a positive constant for proportionality to be applied in the rate of learning. A different approach can be adopted in which the real activation of unit k is not applied. Rather, to adjust the weight, the difference between the desired and the actual values are computed as is presented in the scenario below. This value dk used in this model denotes the activation desired by an instructor. Often, this is referred to as the Widrow-Hoff rule, and sometimes as the delta rule. Conclusion In spite of the advancements in technology in the modern days, there are certain tasks and functionalities that the different programs cannot perform. Even though, an implementation of different programs in a neural network can help to solve such issues. Through training, the computers are made to perform duties that they could earlier not perform. Better still, the neural networks learn, replacing the need for reprogramming and thus easy achievement of desired goals. References Arinkin, V., Digel, I., Porst, D., Artmann, A., and Artmann, G., 2014. Phenotyping date palm varieties via leaflet cross-sectional imaging and artificial neural network application. BMC Bioinformatics, 15(1), pp. 1-16. Gomes, G., Ludermir, T., and Lima, L., 2011. Comparison of new activation functions in neural network for forecasting financial time series. Neural Computing & Applications, 20(30), pp. 417-439. Gomes, G., and Ludermir, T., 2013. Optimization of the weights and asymmetric activation function family of neural network for time series forecasting. Expert Systems With Applications, 40(16), pp. 6438-6446. Li, G., Niu, P., Duan, X., and Zhang, X., 2014. Fast learning network: a novel artificial neural network with a fast learning speed. Neural Computing and Applications, 24(7/8), pp. 1683-1695. Rogers, S. K., and Kabrisky, M., 2007. An introduction to biological and artificial neural networks for pattern recognition. 5th Ed. Bellingham, Wash. USA, SPIE Optical Engineering Press. Rojas, R., 2006. Neural networks: a systematic introduction. 7thEdition,.Berlin [u.a.], Springer. Salsani, A., Daneshian, J., Shariati, S., Yazdani-Chamzini, A., and Taheri, M., 2014. Predicting road header performance by using artificial neural network. Neural Computing and Applications, 24(7/8), pp. 1823-1831. Savaci, F. A., 2006. Artificial intelligence and neural networks: 14th Turkish symposium.3rd ed. Berlin [u.a.], Springer. ELECTRONIC TEXTBOOK, 2002. [online] Available at: < http://www.fmi.uni-sofia.bg/fmi/statist/education/textbook/ENG/glosa.html > [accessed 25 June 2014) Read More