Research Paper On Artificial Neural Networks
Research Paper On Artificial Neural Networks
Artificial Neural Networks (ANNs) and Neurochips are computer(s) that operate on the principles of neural patterns, as opposed to modern computing which operates in a linear fashion. Neural patterns are when signals are sent electronically utilizing the same patterns as a human brain uses. In this paper, I will analyze the applications of ANNs including both the advantages and disadvantages. I plan to analyze not only the benefits, but also the dangers, or rather, problems with this technology. With a technology this cutting-edge, analyzing the applications can be just as important as researching the features and drawbacks of the technology itself. I will look at specific applications where this technology is effective and ineffective. I have noticed that the literature in this field tends
to express either support or skepticism around full implementation of the technology, and yet none seem to mention the idea of partial implementation, which I will discuss here. Neural Networks were originally theorized in 1890 by William James in his work Psychology: The Briefer Course (Anderson J. A., 1988). However, primitive Artificial Neural Network research began around 1943 (Basheer & Hajmeer, 2000). The first versions of artificial neural computers were built in the 1950s (Basheer & Hajmeer, 2000). In the 1980s and 1990s, neurocomputing began to have the computing power necessary for more complex systems and the field flourished (Basheer & Hajmeer, 2000). Additionally in 2007, two researchers at a lab in Tel-Aviv discovered how to implant additional memories of neural firing patterns into a pre-existing neural network (Ben-Jacob & Baruchi, 2007). In addition, in 2010, three researchers confirmed a theory that many smaller ANNs were more efficient than one large one. (Chen, Xue, & Tokuda, 2010). Ironically, the field of ANN computing has died roughly three times as of this writing, mostly due to lack of
resources, and the limitations of the technology at the time. However, Artificial Neural Networks have been adopted into some fields specifically because of their advantages over bit computers. In order to help visualize Neural Networks and Artificial Neural Networks work, I included figure 1.1 which shows that ANNs can have multiple inputs and multiple outputs. Hidden units are nodes that receive the inputs from the input nodes, read and combine the inputs, and pass the inputs further down the network (Reingold, 1999).
Figure 1.11 ANN computing is an alternative format for computing, where signals are electronically sent in the same way that a human brain would. ANNs, unlike bit computers which operate in a linear mode, are able to work like a human brain due to the use of multiple nodes for many changing inputs and outputs as opposed to a single linear process,
http://www.psych.utoronto.ca/users/reingold/courses/ai/nn.html
as depicted in Figure 1.1 (Reingold, 1999). Generally, current ANN hardware (called a neurochip) utilizes actual biomaterial for parts of the processor. One of the more useful features of these neurochips are that they lend themselves to self-learning and adaptation more readily than bit computers which, unless re-written by a programmer, have their instructions set in stone. ANN computers, though a rapidly advancing technology, cannot be employed in every field due to high costs and infeasibility of use. Certain fields, however, still employ ANN computers in effective methods. ANN computers are utilized mainly in areas of probability mapping, robotics (Pretorius, du Plessis, & B., 2009), statistics, and simulations. Additionally, there are theories that ANNs may be useful for odds projection in areas such as gambling (Davidson, Billings, Schaeffer, & Szafron, 2002) and stock market forecasting (Wu & Ruey-Pyng, 1993). After training an artificial neural network, it becomes very accurate (Pretorius, du Plessis, & B., 2009), however, before training, the accuracy of the ANN is approximately nil. For instance, assuming a neural network has input values of 0 and 1, before training it has a 50% chance of correctly identifying the number. This situation gets much worse when multiple inputs are added. Four inputs, before training, would reduce accuracy dramatically. Learning for an ANN is basically a trial-and-error process (as exemplified by ANNs that focus on statistics). After training, accuracy can approach 90% or greater (Pretorius, du Plessis, & B., 2009). Examples of where accuracy would be useful include path-following and object avoidance in semi-autonomous/fully autonomous robots. One of the drawbacks of ANNs in relation to their accuracy is that they are not at all accurate
out of the box. Thus; they cannot be employed effectively without specifically tailoring the ANN to the task at hand. One example where ANNs should not be used is in personal computing, due to their inaccuracy and actual inferiority to bit computers out of the box. With a large enough dataset, ANN computers become able to generalize data, much like a human brain, however, if the data given is too small, the resulting ANN displays bad or even no generalization ability. Moreover, if the amount of hidden nodes is too large, the ANN cannot generalize as effectively or even at all (Basheer & Hajmeer, 2000). Additionally, if the ANN is over-trained then the ANN is amazing at recall, but it also becomes horrendous at generalizing unfamiliar or new data (Basheer & Hajmeer, 2000).
Figure 1.22 As one can see from this graph, which tracks the angle of the ANN-equipped robot, the deviation between the predicted and actual value of angle in the robot was minimal,
2
depicting the accuracy and prediction ability of a properly tuned Artificial Neural Network in a real-life context. Though ANNs can be highly accurate and can generalize well if given enough setup time and a large enough dataset size, they take such a large amount of time to set up that they become impractical for application in every field, which is why I advocate for use of ANNs alongside bitcomputers for the fields in which they are useful. As the number of hidden nodes in an ANN increases, so does the training time (Basheer & Hajmeer, 2000). A possible solution, noticed by Chen, Xue, and Tokuda, noted that hierarchical subnetworks within a larger assembler network were more efficient than a single, more complex network (Chen, Xue, & Tokuda, 2010). Another alternative is to build onto the network as needed, however that approach takes far longer and is more inefficient than the possible idea of many sub-networks within assembler networks (Basheer & Hajmeer, 2000). Asim Roy argues that training cannot be called autonomous, as human correction and verification influences the results, and thus, autonomous networks are truly only semiautonomous at best (Roy, 2000). I would argue, however, that this argument, though valid, has no real bearing on ANN research, due to the fact that as long as the system works properly, debating the difference between semi-autonomous and fully-autonomous is a moot point. The overlap between ANNs and actual human brains may eventually get to the point where humans begin to question the idea of what is human. Current-generation ANN hardware utilizes actual brain cells for the CPU, which may challenge our idea of being human due to the fact that ANNs are able to emulate a human brain quite
effectively. The question is whether a computer system that utilizes biological components and operates exactly like a human brain can be considered human itself. At what point do we consider a system sentient? While this may seem to be many years off at this point, technology is rapidly blurring the line between biological and artificial. When we think about sentience, do we talk about a quality within an organism, or rather an emotional response from humans themselves? If it is the latter, then one could technically consider laptops sentient, because of the fact that people develop feelings towards their personal computers. Machines are unemotional and rational as of now, while humans are emotional and sometimes rational or irrational. ANNs are unemotional; however ANNs can be both rational and irrational. Assuming machines (including ANNs) learn emotions, will we then consider them to be sentient? If anything, human emotions are both a flaw and an asset at the same time. Machines (especially ones running ANNs and using ANN hardware) would have to be able to evolve emotions and prove to humans that they are capable of independent thought alongside emotional responses, before a human would classify them as sentient. Theoretically, the Turing test, a method of determining sentience of a machine based on human-machine interaction would be possible; however, I do not believe that it would be a perfect solution to this issue. Either way, these would raise the question of what is the difference between man and machine, which cannot be easily defined. Though humans are made up of biological components, they are essentially machines themselves, very complex biological ones, but machines nonetheless, an idea originally postulated by Dr. Rodney Allen Brooks of the Massachusetts Institute of Technology.
There are specific places where humanlike ANNs could be useful, more so than either a human or a machine. One such place is theoretically in the field of accounting and finance. A machine with humanlike traits and with a lack of greed would be optimal compared to a human utilizing a machine. Additionally, utilizing ANNs for triage would be a far better idea than relying on overtaxed and overstressed humans (Eftekhar, Mohammad, Ardebili, Ghodsi, & Ketabchi, 2005). Moreover, for situations such as poker, a properly trained, humanlike ANN could provide human players with a worthy opponent, because of the element of not knowing that ANNs can excel in, and in which regular computers just cannot compute (Davidson, Billings, Schaeffer, & Szafron, 2002). Additionally, ANNs have been effective for spatial awareness in robots as exemplified by the aforementioned figure 1.2 (Pretorius, du Plessis, & B., 2009). Current state-of-the-art ANN computers are built upon organic material, generally brain cells, combined with circuitry. These are called neurochip-based computers. They are specifically designed to run ANN algorithms, and can work with multiple firing patterns at the same time (Ben-Jacob & Baruchi, 2007). Additionally, one can emulate ANN algorithms on traditional bit-computers; however, ANNs are designed to be utilized with neurochips to perform the most effectively. I see this field having quite a bright future, if the field of ANN research realizes that even humans cannot learn independently around the beginning of their lives, and that even humans require training from time to time. In contrast to some people who argue that ANNs cannot be useful due to the long training time, I suggest that ANNs be used only
where needed to supplement standard bit-computers. Additionally, not utilizing ANNs in every aspect of life would be preferred.
Bibliography
Anderson, J. (1995). An introduction to neural networks. Cambridge, Mass: MIT Press. Anderson, J. A. (1988). Neurocomputing. Cambridge, Mass: MIT Press. Basheer, I. A., & Hajmeer, M. (2000, December 1). Artificial neural networks: fundamentals, computing, design, and application. Journal of Microbiological Methods, 43(1), 3-31. Ben-Jacob, E., & Baruchi, I. (2007, May 16). Towards neuro-memory-chip: Imprinting multiple memories in cultured neural networks. Physical Review E, 75(5), 050901. Chen, L., Xue, W., & Tokuda, N. (2010, August). Classification of 2-dimensional array patterns: Assembling many small neural networks is better than using a large one. Neural Networks, 23(6), 770-781. Commons, M. L. (1991). Neural network models of conditioning and action. Hillsdale, N.J: L. Erlbaum Associates. Davidson, A., Billings, D., Schaeffer, J., & Szafron, D. (2002, January). Improved Opponant Modeling in Poker. Artificial Intelligence, 134(1-2), 201-240. Eftekhar, B., Mohammad, K., Ardebili, H. E., Ghodsi, M., & Ketabchi, E. (2005, February 15). Comparison of artificial neural network and logistic regression models for prediction of mortality in head trauma based on initial clinical data. BMC Medical Informatics and Decision Making, 5. Pretorius, C. J., du Plessis, M. C., & B., C. C. (2009). Towards an Artificial Neural Network-based Simulator for Behavioural Evolution in Evolutionary Robotics. Proceedings of the 2009 Annual Research Conference of the South African Institute of Computer Scientists and Information Technologists, 170-178. Reingold, E. (1999). Artificial Intelligence Tutorial Review for Psychology Students - Neural Networks. Retrieved August 02, 2010, from University of Toronto, Dept of Psychology Web site: http://www.psych.utoronto.ca/users/reingold/courses/ai/nn.html Roy, A. (2000, January). Artificial Neural Networks - A Science in Trouble. SIGKDD Explorations, 1(2), 33-38.
Wu, S.-i., & Ruey-Pyng, L. (1993). Combining artificial neural networks and statistics for stockmarket forecasting. Proceedings of the 1993 ACM conference on Computer science, 257 264.