Improving Assisted Technology
With Machine Learning Techniques

Senior Seminar in Computer Information Systems
December 7, 2011
Abstract
An ever increasing population of persons with learning disabilities are continually in need of better ways to overcome the unique challenges they face in today's modern, high communication world. While Assistive Technology is making strides to close the learning gap between persons with and without learning disabilities there is still a long way to go before technology provides a level playing field for these challenged individuals. Many of the issues with existing assistive technology revolves around clumsy, inefficient interfaces that struggle to find a balance between ease of use and sufficient complexity to ensure that the proper sequence of instructions is implemented. Machine learning is on the cutting edge of programming practices and presents some significant improvement possibilities in the areas of natural language processing, pattern recognition, and interface design. Machine learning has the potential to play a significant role in allowing assistive technologies to be more adaptive to persons with diverse sets of needs. This paper will attempt to define some specific areas of assistive technology that could benefit most from the application of machine learning. We will frame the definitions by aligning specific learning disabilities with current and future assistive technologies and then examining how the implementation of machine learning could improve upon them.
Introduction
The need for assistive technologies is undeniable with as many as 8 to 10 percent of children that are under the age of 18 in the United States having one form of learning disability or another.(NINDS) Experts believe that this statistic will continue to trend up making the area of assistive technology a field of study with increasing importance. When designing learning systems and technology for traditional learners without learning disabilities a normal bell curve can be applied to interface and product design in that the product can be universally designed to maximize the effectiveness for a majority of the learners on the curve. This can be contributed to the relative normalization of efficient learning methods, as learners are grouped towards the middle of the curve. However in the area of learning disabilities there are wide ranging and complex sets of conditions that must be uniquely identified in each learner. In essence the non-traditional learners with learning disabilities typically do not fall on the normal distribution of the curve which makes universal product design very difficult. The complexity in designing universally effective assistive technologies forces us to look to new methods of interface design and processing. One possible way of addressing the complexity is through the use of machine learning and adaptive algorithms. Machine learning is a very broad term for the ability of a system to analyze patterns and essentially learn what is being processed. This type of programming eliminates the need to code elaborate stores of responses to all known questions by implementing algorithms that can identify the patterns in the input data and then make determinations based on its available knowledge base. As learners with learning disabilities can often have trouble communicating input variables in a traditional manor the use of this adaptive technology could potentially improve instruction set recognition in assistive technologies.
What Is A Learning Disability? In order to fully understand the impact of machine learning on the effectiveness and efficiency of assisted technologies we must first assess how persons with learning disabilities are classified. The National Institute of Neurological Disorders and Stroke defines learning disabilities as "disorders that affect the ability to understand or use spoken or written language, do mathematical calculations, coordinate movements, or direct attention." (NINDS) As you can see this definition can cover a wide breadth of conditions. For the purposes of this assessment we need to limit the scope of our focus to a smaller subset of the learning disability spectrum. Although the technology discussed in this paper can be applied to many facets of the learning disability spectrum we will focus our attention on individuals who interface with assistive technology through the use of voice communication. This focus will give an opportunity to explore the main topic of this paper which is analyzing input patterns using adaptive algorithms.
What Is Machine Learning? A definition that would sufficiently encompass the breadth of machine learning would be of limited use in the scope of this paper. As such the definition we will use here is the implementation of advanced programming techniques designed to change the structure of the program or data based on external inputs in such a manner that the expected future performance of the interaction improves. To expand on this definition further we would say that a system designed to face a particular interaction multiple times and produces measurable improvements over each previous instance would qualify as a learning system. This type of learning system is generally based on advanced algorithms that collect and analyze examples from each interaction and use them to better predict the outcome of future interactions with similar inputs. Tom Mitchell summarizes this characteristic as "Computer Science has focused primarily on how to manually program computers, Machine Learning focuses on the question of how to get computers to program themselves (from experience plus some initial structure)." (Mitchell, 2006) In Mitchell’s statement "programming themselves" basically mean the system doesn't require programmer or system administrator intervention to modify its program sufficiently to improve on its answer, it simply needs additional interactions to analyze.
What Is Assisted Technology? This answer when framed in the context of this paper will be loosely defined as any device or equipment that assists a learner to bypass that individual’s specific learning deficit. Some of the predominant technology being used by individuals with learning disabilities are Speech-Recognition programs, Speech Synthesizers and Screen Readers, Talking Calculators, Talking spell checkers and electronic dictionaries. Bharucha notes when talking about designing these types of devices:

“the complexity of designing devices that are commensurate with an individual's needs and preferences, critical research gaps also exist in the engineering and computer science domains. These include a) efficient collection and storage of voluminous real-time continuous data from multimodal sources (basic and advanced sensors, audio and video technologies, etc.) that lend themselves to user-friendly access and analysis" (Bharucha) Speech-Recognition programs are programs that collect the individuals input via voice command then analyze the patterns in the signal and attempt to generate the appropriate command signals to the device. This type of device is of primary concern in the structure of this paper due to the advanced algorithms employed in the analysis of the captured input. Speech Synthesizers are another major component and is closely linked with the use of speech recognition programs. Speech Synthesizers, sometimes called screen readers, analyze the data that is normally available to an individual via the computers display and output its contents as audio only speech. Talking calculators are use similar processes as Speech Synthesizers to assist visually impaired learners perform calculator functions by audibly describing the contents of the visual display. Talking spell checkers are quite simply screen readers with a very specific purpose, to allow visually impaired learners the functionality of using a spell checkers through audio representation of the display screen. The impact of learning disabilities on individuals can be catastrophic if not recognized and mitigated properly. Currently there are several main weaknesses with standard speech processing system. The first major weakness is the inability of the current crop of programs to adapt to a variety of different speech patterns and conversation styles. Many of the speech processing programs are very good at interoperating a small percentage of speech patterns but fail in many instances simply because the user is using a speech pattern that is different enough not to be recognized by the system. This is a major issue considering the vast array of stylized speech patterns in the U.S. alone. To compensate for this programmers have tried to program a wider array of styles into the structure of the program in an attempt to identify a larger cross-section of users. This however is still limited to the pre-programmed styles and can in no way accommodate every style and generally have a limited vocabulary bank. Another issue with the current crop of speech recognition programs is that these programs typically require that the user spend a considerable amount of time learning how to communicate with the system. During this time the computer is also trying to identify words based on the responses of the user, a limited attempt at understanding the speech pattern of the user. The training that is needed in order for the system to begin to understand the user is in large part one reason that frustration sets in on the part of the user. The use of speech recognition has come a long way in improving the lives of individuals with learning disabilities however there is still a large gap to bridge. The current crop of speech recognition tools that are available today have one other major issue. After the user has spent the time training the software to recognize his/her speech pattern the user must always use a similar speech pattern to interact with the system efficiently. An example of when this could be a problem is if you spent a considerable amount of time training your speech recognition software to understand your voice and then you suddenly came down with an illness that caused you speech pattern to change considerably. At this point the currently available speech recognition technology would have difficulty adapting to your new pattern and fail to efficiently complete your requests. This is all due to the way the tools are currently created. Each tool is pre-programmed with a set vocabulary and very rigid rules as to how it can make modifications to its internal structure. Kinber notes that "restricting the data available to the inference machine also reduces its learning potential" (Kinber) Through the use of machine learning techniques and algorithms speech recognition could reach a much broader audience with greater efficiency. The underlying issue with most speech recognition tools is that its main goal is to identify and replicate exactly what the user is saying rather and attempting to understand what the user is implying. For instance if you are trying to instruct a piece of assistive technology to perform a task like sending an email to your sibling at a particular email address but your specific disability keeps you from properly pronouncing the word email, or send, the common speech recognition software programs will have a greater chance of not accepting the command. To attempt to overcome this issue experts are working on a new technology that when combined with speech recognition programming can understand the intent of the request rather than a just a literal translation of the words. This technology is called natural language processing or NLP. Natural language processing uses machine learning algorithms to analyze patterns in the recorded data and attempt to classify them into groups of conditions. This process is called pattern recognition and it differs from the traditional speech recognition program algorithms that are based on pattern matching. Pattern matching looks for exact pre-programmed copies in the input text whereas the pattern recognition used in NLP attempts to attain a match based on a patterns located in the input stream that can be assigned to a category. The primary types of pattern recognition used in NLP falls into a few general categories. The first category is Supervised Learning which is basically looking for patterns in a pre-set dataset with previously labeled categories and outputs that have been verified. This type of learning is limited to the extent of the data set provided during setup however can still provide a substantial increase in pattern recognition and usability over pattern matching systems. The second category of pattern recognition in NLP is called Semi-Supervised Learning. Maria-Florina Balcan describes semi-supervised learning as "semi-supervised learning—using unlabeled data together with labeled data—since any useful information that reduces the amount of labeled data needed can be a significant benefit."(Balcan) This type of pattern recognition uses a sample dataset and the user input as training data. This increases the sample size for comparison and thus increases the chance of a successful outcome. The last category of pattern recognition in NLP is called Unsupervised Learning. In this type of pattern recognition there is no preprogrammed sample dataset to be compared too. The tool simply uses the incoming data stream and assigns labels on the fly. The use of one particular category of pattern recognition or another is simply based on the type of data stream that needs to be analyzed. If the data stream is relatively small and the outcomes are well know then a supervised approach may be the most efficient implementation, however if the input stream is largely unknown a semi-supervised or unsupervised approach may be necessary. The use of natural language processing to improve the assistive technologies like speech recognition can dramatically improve the ability of the user to overcome their specific disability by allowing a more natural speech pattern to be recognized. The use of NLP is not limited to speech recognition however, there are implications in a wide range of assistive technologies which will someday allow individuals with learning disabilities to better communication and interact with their assisted technology devices.
Conclusion
The use of assistive technology can have an amazing impact on the lives of individuals with learning disabilities. The implementation of those technologies and the efficiency of their operation can be the difference between an individual that is able to actively participate in learning or one that becomes dejected and falls behind other learners. Processing input commands efficiently is a key step to proving the real time analysis and processing needed to bridge the gap that the individuals learning deficit has created. The use of machine learning algorithms in these assistive technologies could significantly reduce or eliminate the need for custom designed programming tailored to each specific deficit by creating a package that can adjust on its own to the specific needs of the user through analysis of the interactions between the device and the user. The true power of technology in regards to assistive technology is summed up in this quote:

For most people technology makes things easier. For people with disabilities, however, technology makes things possible. In some cases, especially in the workplace, technology becomes the great equalizer and provides the person with a disability a level playing field on which to compete. (Bailey)

Works Cited
Bailey, Nell. "Assistive Technology, Accommodations, and the Americans with Disabilities Act." Assistive Technology, Accommodations, and the Americans with Disabilities Act. Ed. Susanne M. Bruyere. Cornell University, Dec. 2000. Web. 5 Dec. 2011. .
Balcon, Maria-Florina, and Avrim Blum. "A Discriminative Model for Semi-supervised Learning." Citation. Web. 08 Nov. 2011. .
Bharucha, Ashok J., Vivek Anand, and Jodi Forlizzi. "Intelligent Assistive Technology Applications to Dementia Care: Current Capabilities, Limitations, and Future Challenges." Intelligent Assistive Technology Applications to Dementia Care: Current Capabilities, Limitations, and Future Challenges. National Institute of Health, 27 Oct. 2009. Web. 09 Nov. 2011. .
Jorge, Joaquim. "Adaptive Tools for the Elderly." Citation. Web. 08 Nov. 2011. .

Kinber, Efim, Rusins Freivalds, and Carl Smith. "On the Impact of Forgetting on Learning Machines." Citation. Web. 08 Nov. 2011. .
"Learning Disabilities Information Page." National Institute of Neurological Disorders and Stroke (NINDS). Web. 07 Dec. 2011. .
Mitchell, Tom M. "The Discipline of Machine Learning." Carnegie Mellon University ML-06-108, July 2006. Web. 1 Dec. 2011. .
Turkington, Carol, and Joseph Harris. Understanding Learning Disabilities: the Sourcebook for Causes, Disorders, and Treatments. New York: Checkmark, 2003. Print.