This block diagram outlines the basic steps required to be addressed when developing a machine learning solution. 

Context in source publication

Context 1

... have developed a platform for exposing high school students to machine learning techniques for signal processing problems, making use of relatively simple mathematics and engineering concepts. Along with this platform we have cre- ated two example scenarios which give motivation to the students for learning the theory underlying their solutions. The first scenario features a recycling sorting problem in which the students must setup a system so that the computer may learn the different types of objects to recycle so that it may automatically place them in the proper receptacle. The second scenario was motivated by a high school biology curriculum. The students are to develop a system that learns the different types of bacteria present in a pond sample. The system will then group the bacteria together based on similarity. One of the key strengths of this platform is that virtually any type of scenario may be built upon the concepts conveyed in this paper. This then permits student participation from a wide variety of educational motivation. Index Terms — Machine Learning, Pattern Recognition, Secondary Education, Lab Modules Machine learning (ML), a subfield of artificial intelligence, has evolved out of the need to teach computers how to automatically learn a solution to a problem. In engineering this field is referred to as pattern recognition, aptly named because the computer is extracting patterns out of data and making a decision based on the pattern identified. It is a rich field that is broadly and inherently related to signal processing most notably through data-driven learning methodologies [1, 2]. Our understanding of human learning has inspired many of the ML methods currently available. For example, take a look at the foundation of neural networks, which is based off the structure of the interconnection of multiple neurons from the brain [3]. While this class of methods undoubtedly employs coarse approximations of actual neuron function, they have shown tremendous success in several ML applications [4]. A few examples of ML applications include speech recognition aka natural language processing, image processing such as face detection, DNA sequence classification, financial analysis such as detecting credit card fraud, sports prediction and search engine algorithms which have been put into use by major household name search providers [5, 6, 7]. Many ML techniques do require a moderate mathematical background. Statistics, linear algebra, and calculus are commonplace in many of the algorithms. Fortunately, simple mathematical techniques have been shown to be quite successful on practical problems exploiting ML. With only an understanding of means and Euclidean distances, students can be shown how to instruct a computer to identify the difference between a pen and pencil, albeit with the proper assumptions. The simplicity of the math therefore permits accessibility of the field of ML to the high school student. ML not only employs mathematics for practical purposes, but also demands problem-solving skill at a fundamental level since each problem encountered requires proper tool selection and then the interfacing of the tool to the problem. Through the application of lab modules such as the one proposed in this paper, students may be exposed to multi-interdisciplinary fields simultaneously such as engineering, mathematics, computer science and the field of the problem being addressed such as biology, economics, photography, etc. There are innumerable examples with which machine learning techniques may be employed to facilitate automatic problem solving. Suppose you wish to separate quarters, nickels and dimes. What information would the computer need to dis- tinguish between these three types of coins? Think about how you would do the task yourself. It is easy to see that each type of coin has a different size. So all we need to do is supervise or rather tell the computer the circumference of each type of coin and then let the computer automatically do the sorting. This problem is exceedingly easy and does not necessarily require any complex ML technique. Now let’s say that we have a bag of unknown coins of several different currencies. We wish the computer to sort the coins by type and currency automatically. In this case the computer must learn the types for sorting and then classify each coin without our explicit input as in the previous example. This problem requires an unsu- pervised solution in that we do not know the types of coins beforehand. This example will serve as the basis for delineat- ing the basic steps involved in developing a machine learning solution as depicted in figure 1. The first step in any engineering design is to define the problem. In our example our problem is to sort a bag of unknown coins. To further motivate the use of ML techniques lets say that the bag contains 1,000,000 coins so that a man- ual solution of sorting is unrealistic for all practical purposes. We cannot simply just feed the computer the bag of coins, but must provide it with some information that it can use to make a decision about each coins type or rather its class as its known in the ML literature. This second step of the design is known as feature extraction. Feature extraction requires the user to provide the computer with information that may be used to differentiate the classes, which in our case is the types of coins. We need to insure that our features contain enough discriminatory information about the classes. For example, shape would probably be a poor choice since each coin is assumed to be round. However, if some coins were non-circular than this could be one useful feature. Many times we include multiple features to aid the algorithm. For our example the diameter of the coin is probably a useful feature, but since we have coins of multiple currencies there is a good chance that two different types of coins will have the same diameter. Therefore we chose a second feature such color of the coin. But since we are working with mathematics and computation, we need to map color to a number. We cannot simply say silver or copper, but instead find a numerical quantity that furnishes the same discriminatory information, say RGB value or perhaps luster. Let us choose luster since its arguable easier and cheaper to obtain. Some of the coins may be older than the others and therefore may not be as lus- trous as those newer of the same class. This variation could be viewed as each type of coin having a mean value of luster and a standard deviation of luster across all coins of that type. Depending on the chosen algorithm, this information could be of use for separating the coins. A third feature may be added such as weight of the coin and a fourth, albeit ex- treme example, such as the bacterial composition found on each coin with the assumption that the composition varies de- pending on the region of the world the coin originated from. From this discussion, it is therefore no surprise that if poor features are chosen, then the algorithm will perform poorly. This is known as the garbage-in, garbage-out theorem and is why feature selection is one of the most important steps in ML and is almost always carried out by the designer. There are potentially hundreds, if not thousands of different ML algorithms to choose from. In many instances, the algorithms are modified to fit a particular problem resulting in a new model, thereby growing the population of choices. High school students should not have a problem comprehend- ing Euclidean distance and mean so we suggest the use of the K-means algorithm for solving our coin sorting example and the subsequent lab activities. The details of the K-means algorithm are discussed in the next section. Of course, more advanced algorithms may be selected based on the students mathematical aptitude resulting in a highly scalable ML lab platform. No matter which algorithm is chosen it must be properly coded and executed on the computer. Many programming options are available, but we suggest the use of Matlab for rapid implementation (www.mathworks.com). This software package already includes the K-means algorithm as a one- line command thereby alleviating the students or teacher from coding the algorithm from scratch. Alternatively, coding this simple algorithm could be an excellent exercise for a student in a computer science course using any language such as C, Java, Perl, Python, etc. Weka is another great machine learning tool for algorithm implementation that is based on Java [8, 9, 10]. The bottom line is that the coding part of the ML activity can be kept as simple or scaled as complex as desired by the educator without loss of generality. Once the corresponding data has been captured based on the features selected and the algorithm has run on the data we evaluate the performance of the classifier on the dataset. In our example we expect to have groups of coins with each group containing one type of coin. In order to evaluate the performance of our ML selection we need to know the ground truth of the type of each coin. Here we may simply visually inspect the coins in each group. Under other circumstances this information could be recorded, set aside and kept unknown to the algorithm until after it has run, upon which the truth is compared to the algorithms output. There are nearly as many evaluation metrics in ML as there are algorithms. The most appropriate metric depends on the specific application and problem being solved making the selection another aspect of design. We could simply count the number of misplaced coins and use that figure as our metric. A metric known as the rand index would be suitable for our ML design [11]. This metric shown in eqn. (1) essentially as- sesses the similarity of the groupings between the algorithms output and the ground truth. It is relatively straightforward to implement the rand index in Matlab or an Excel ...