Posted: February 1st, 2022

Machine Learning

Machine Learning
Machine learning is the enabling of computers to perform tasks that until then belonged to human beings. Machine learning involves tasks such as the translation of speech to driving cars. Therefore, it is the driver of the capabilities of artificial intelligence because its helps the software to make sense of the unpredictable circumstances. At an advanced level, machine learning is a process that teaches computer systems to make accurate predictions after receiving data. The predictions could be stating that a piece of fruit shown in a photo is either an apple or a banana. The main different between the conventional software and machine learning is that it does not have a written code that instructs it on how to tell the difference between the fruits. The makers of the machine-learning model teach it to make predictions reliably by feeding it with large amounts of data.
In recent times, machine learning has made significant developmental strides. It is one of the methods used to achieve artificial intelligence. The early inceptions of artificial intelligence were in the 1950s (Neverova et al., 2016). The definition of AI was any machine with the capabilities of performing tasks that would otherwise require human intelligence. To that extent, AI demonstrates traits such as learning, planning, problem solving, reasoning, perception, manipulation, motion and knowledge presentation. In some instances, AI also demonstrates creativity and social intelligence. There are various approaches used to build AI and machine learning systems. One of the methods used is evolutionary computations and it involves computation and random mutations between generations by attempting to evolve or form optimal solutions. The programming of computers equips them with rules that enable them to mimic the behavior of human experts in specific domains for instance a plane flying on autopilot.
There are two distinct types of machine learning, supervised and unsupervised learning. The supervised learning approach involves the training of machines by example (Ross, Graves, Campbell, & Kim, 2013). In the teaching of supervised learning, the systems interact with significant amounts of data with labels. For instance, there may be images of showing handwritten numbers annotated to show the matching figures. After receiving sufficient examples, the system learns to recognize the cluster of pixels and the shapes that have an association with specific numbers. Eventually, the system is able to identify handwritten figures. For example, systems reliably differentiate between 4 and 9. The training of the systems requires a considerable amount of labelled data. In fact, some systems require exposure to millions of examples just to master a task.
The data sets used in teaching the systems are vast. Google open image data set has about nine million images. The video repository has close to seven million videos. Each day, the size of the training data sets continues to gro (Karatekin et al., 2019). Recently, Facebook announced it had compiled close to 3.5 billion images on Instagram with hashtags used to label each of them. According to a benchmark, set by ImageNet one billion photos can provide a system with an accuracy of about 85%. The difficult task of labelling the data sets used in the training is the role of crowdworking working services like Amazon Mechanical Turk. The service acquires a large pool of low cost labor across the world. For instance, Amazon Mechanical Turk recruited 50,000 people to assist ImageNet. However, Facebook has an approach that is not labor intensive because it does not involve manual labelling. Instead, Facebook uses data that is publicly available to teach the systems.
On the other hand, unsupervised learning involves algorithms identifying patterns in vast amounts of data. The algorithms try to spot the similarities splitting the data into categories (Tan & Zhang, 2005). For instance, a category may be a cluster of Airbnb houses available for rent in a certain neighborhood. It may also appear as google news putting together stories that have a similar topic every day. However, the design of the algorithms cannot single out distinct types of data. What the algorithm does is to assess forms of data that can fit into similar groups or certain features that stand out.
Overtime, the need to have vast sets of labelled data for machine learning training purposes may diminish. Semi-supervised learning is one of the main reason for the decline (Pahwa & Agarwal, (2019). The model incorporates both unsupervised and supervised learning. Semi-supervised learning relies on large amounts of unlabeled data and small-labelled data to teach the systems. The labelled data partially trains the machine learning models. The partially trained models then label the unlabeled sets of data in a process called pseudo labelling. Recently, semi-supervised learning models have received a boost after Generative Adversarial Networks developed machine-learning systems with the ability to use labelled data to generate a new set of data. In the event that semi-supervised machine learning become as effective as supervised learning, then the access to vast amounts of computing power may be more important than large labelled datasets.
In training a machine-learning model, a mathematical model modifies repeatedly the operations until the system is able to make accurate predictions when provided with fresh data. Before commencement of the training, it is imperative to choose the data required and decide the best data to feature (Zhang, Su, & Wang, 2007). It is also critical to note that the data should have a balance. For instance, for a distinction between wine and beer, each of the items should have an equal example. Before, the training, here is also a step of data preparation. In the step, there is normalization, error correction and deduplication. The following step involves choosing appropriate machine learning models the wide variety available. Each of the models has its own strengths depending on the type of data available. For instance, some of the data is better for text, other forms are prefer numerical data while other suite images.
The learning process involves tweaking functions of the systems until it is able to make accurate predictions. Once the process of training is complete, there is an evaluation using the data remains from the training to gauge performance. To enhance the performance further, there is room to fine-tune the parameters. One of the most common algorithms for both unsupervised and supervised machine learning is neural networks. The algorithms underlie machine learning and models such as linear regression are able to make predictions by relying on a number of data features. The neural networks assume a structure similar to that of the brain. The network is an interconnected layer of algorithms. The neurons feed data to each other using an output that preceded each layer. Each layer has the ability to asses a distinct feature of the data presented. The networks learn to identify certain components of data during the training process.

References
Karatekin, T., Sancak, S., Celik, G., Topcuoglu, S., Karatekin, G., Kirci, P., & Okatan, A. (2019, August). Interpretable Machine Learning in Healthcare through Generalized Additive Model with Pairwise Interactions (GA2M): Predicting Severe Retinopathy of Prematurity. In 2019 International Conference on Deep Learning and Machine Learning in Emerging Applications (Deep-ML) (pp. 61-66). IEEE. DOI: 10.1109/Deep-ML.2019.00020
Neverova, N., Wolf, C., Lacey, G., Fridman, L., Chandra, D., Barbello, B., & Taylor, G. (2016). Learning human identity from motion patterns. IEEE Access, 4, 1810-1820. doi: 10.1109/ACCESS.2016.2557846
Pahwa, K., & Agarwal, N. (2019, February). Stock Market Analysis using Supervised Machine Learning. In 2019 International Conference on Machine Learning, Big Data, Cloud and Parallel Computing (COMITCon) (pp. 197-200). IEEE. DOI: 10.1109/ICMLC.2009.5212413
Ross, M., Graves, C. A., Campbell, J. W., & Kim, J. H. (2013). Using support vector machines to classify student attentiveness for the development of personalized learning systems. In 2013 12th International Conference on Machine Learning and Applications (Vol. 1, pp. 325-328). IEEE. DOI: 10.1109/ICMLA.2013.66
Tan, Y., & Zhang, G. J. (2005). The application of machine learning algorithm in underwriting process. In 2005 International Conference on Machine Learning and Cybernetics (Vol. 6, pp. 3523-3527). IEEE. DOI: 10.1109/ICMLC.2005.1527552
Zhang, P. F., Su, Y. J., & Wang, C. (2007). Statistical machine learning used in integrated anti-spam system. In 2007 International Conference on Machine Learning and Cybernetics (Vol. 7, pp. 4055-4058). IEEE. DOI: 10.1109/ICMLC.2007.4370855

Order | Check Discount

Tags: Machine Learning