Introduction
The topic considered in this paper is the machine learning (ML) framework. ML is relevant since it is increasingly used by businesses to predict the trends and patterns in consumer behavior and inform them of the need to develop new products. Most big tech companies, such as Facebook and Google, consider ML a critical component of their operations (Naha et al., 2020).
However, choosing an ML framework depends mainly on the user’s preferences. The open questions related to this topic pertain to the factors that must be considered when selecting an appropriate framework. Questions such as the preferred programming language for AI and what hardware, software, and cloud services will be used for scaling are critical when selecting an ML framework. Machine learning is a branch of AI that allows computers to self-learn without necessarily being programmed. On the other hand, deep learning combines all the other types of machine learning. Lastly, neural networks are computer systems that mimic the way the human brain functions.
While this topic has a broader scope, the focus of this discussion will be to demonstrate why PyTorch is the appropriate framework. Some of the elements within this assignment’s scope include the limitations of these frameworks and their similarities and differences. However, the process of building and deploying deep machine learning using PyTorch is beyond the scope of this assignment.
Therefore, this essay aims to demonstrate why PyTorch is a suitable framework for US Mobile, as per the Crunchbase website. To achieve this, the paper will first provide an overview of machine learning and the different types of ML. This will establish the rules for determining the appropriate machine-learning framework for US Mobile. To understand how ML frameworks work, it is important to explore different ML models and algorithms, which are basically sets of commands that machines use to make decisions.
Machine Learning
Machine learning (ML), a branch of artificial intelligence (AI), enables computers to learn autonomously without explicit programming. This type of learning occurs through training data for a specified period. ML algorithms can identify patterns in a given dataset, learn from them, and later make informed predictions (Mahesh, 2020). In this traditional programming, instructions were derived from an if-then type of structure – “when certain conditions are met, the program executes a specific action” (Mahesh, 2020, p. 381). ML was introduced to eliminate the human component in the learning process.
Although AI and ML are often used interchangeably, they differ in concept. AI is a broader concept that examines how machines make decisions, acquire new skills, and solve problems like humans. ML is a subset of AI that equips intelligence systems with the ability to learn new things from data (Volkmar et al., 2020). Therefore, instead of programming machines as was initially, they can be fed with multiple labeled data, also known as training data, to guide them through calculating and identifying patterns. In essence, ML can accurately work on vast amounts of data much faster than humans (Volkmar et al., 2022). Organizations that have adopted ML have been able to save a significant amount of money needed to accomplish various tasks and analyses, such as identifying customers who are more likely to cancel their contracts.
Types of Machine Learning
Supervised Learning
Supervised learning (SL) algorithms rely heavily on training data to make predictions. Each training label must consist of an input and a desired output. SL is often used to analyze the labeled data and, in return, makes an inference – it makes an educated guess. It is one of the most popular types of machine learning since “it needs to feed manually tagged sample data to learn from” (Bao et al., 2019, p. 301).
Labeled data guides SL in establishing a clear pattern, such as data categories and similar words. A practical example is when one wants to detect spam automatically (Santoso, 2019). All it takes is feeding the machine learning algorithm with several emails classified as spam and other essential emails not labeled as spam.
The process of learning through email takes place through either classification or regression. It is essential to note that SL employs several classification algorithms in conjunction with Support Vector Machine (SVM) and Naïve Bayes (Madhukar et al., 2019). As noted above, the output value is assigned a finite number of options in classification tasks.
For instance, data can be classified as positive, negative, or neutral. A good example is when an organization is interested in analyzing customer support conversations to understand the emotions expressed. In this case, a sentiment analysis classification can capture responses systematically (Nahar et al., 2020).
Regression tasks, on the other hand, are used when the team expects the results to be a continuous member. This model is often used in predicting quantities such as “the probability an event will happen, meaning the output may have any number value within a certain range” (Bao et al., 2019, p. 302). Similarly, it can be used to predict the spread of the COVID-19 pandemic in a given area.
Unsupervised Learning
The unsupervised learning (UL) algorithm mainly establishes existing relationships in unlabeled data. Here, UL is fed input data, but the intended outcome remains unknown, allowing it to make inferences based on circumstantial evidence without any form of labeled training. As explicated in Bao et al.’s (2019) study, the model must strive to establish clear patterns without being trained with the correct answers.
A typical example of UL is clustering, in which similar data is grouped. A practical example is when the marketing department of an e-commerce company decides to use clustering to improve customer segmentation. The model relies on a specific income and spending data set to categorize customers with similar behaviors. As emphasized further by Addagarla & Amalanathan (2020), segmentation is crucial for marketers when tailoring strategies to each market. UL could help the department offer promotions and discounts to low-income customers as a reward for their continued loyalty.
Semi-Supervised Learning
Semi-supervised learning (SSL) works by dividing the learning data into a small, labeled group and a larger set of unlabeled data. Labeled data is used as input to derive inferences about unlabeled data, making it highly reliable compared to supervised learning. This model is increasingly used for tasks involving large datasets, such as image classification. The fact that SSL relies on a small amount of labeled data makes it faster to implement. Similarly, it is less costly compared to supervised learning models. Most importantly, it is best suited for businesses with vast data.
Reinforcement Learning
Reinforcement learning (RL) aims to teach a computer program how to behave in various situations to maximize its reward. In other words, RL aims to determine the approach to follow to achieve the desired results, which is achieved through trial and error. One advantage of RL lies in its ability to learn from its mistakes and, ultimately, select the best course of action to achieve the optimal solution (Ye et al., 2019). RL is a standard in robotics and video games, emphasizing the establishment of a clear connection between actions and results.
Deep Learning (DL)
DL combines all the other types of machine learning. DL is an advanced form of machine learning often used by large technology companies, such as Google, Microsoft, and Amazon. It is used to run the entire system, in addition to acting as an intelligent assistant.
DL is based on Artificial Neural Networks (ANN) – a computer that imitates how the human brain functions (Ozbayoglu et al., 2019). DL algorithms are composed of various layers of interconnected neurons, ensuring that multiple systems work uniformly and consistently. Once it receives image, text, or video data, and is commanded to perform a specific task, the data will pass through the multiple layers, allowing DL to learn progressively.
As noted in Ozbayoglu et al.’s (2019) study, DL “is used mainly in image recognition, speech recognition, and Natural Language Processing (NLP)” (p. 109). Deep learning tends to outperform other types of machine learning, particularly in solving complex problems and handling massive datasets. The only disadvantage is that they require large volumes of training data, which also takes time.
Machine Learning Frameworks
Before focusing on how a specific framework could guide US Mobile’s effort to prevent customers from canceling their mobile phone subscriptions, it is essential to provide an overview of different ML frameworks. For clarity, ML frameworks serve as interfaces that data scientists and developers use to build and deploy machine learning models with the highest efficiency (Nguyen et al., 2019). As discussed above, machine learning is utilized in various industries, including healthcare, finance, insurance, and marketing. Therefore, these industries can utilize ML frameworks to enhance their machine-learning capabilities while maintaining an effective ML lifecycle. Although some companies prefer to build their own tools, most organizations rely on an existing framework that meets their specific needs.
TensorFlow
Google introduced TensorFlow and later released it as an open-source project. It is a powerful and versatile ML framework with an extensive library that allows users to build all models, including classification, regression, and neural networks (Leon et al., 2020). Additionally, TensorFlow can customize ML algorithms according to users’ requirements. TensorFlow has become increasingly popular, as it runs on both central processing units (CPUs) and graphics processing units (GPUs) (Simmons & Holliday, 2019). The main features of TensorFlow include a wider platform for distributing CPUs and GPUs, high-level visibility in computational graphs, and parallel neural network training.
The release of TensorFlow 2.0 in 2019 introduced several essential new features. Firstly, the frameworks enabled deployment on various platforms, including mobile devices and IoT, through the SavedModel format, which allows users to transfer the framework to any platform. Secondly, TensorFlow 2.0 was introduced in response to user demand for a framework that could facilitate the process of building and running the entire computing graph (Leon et al., 2020). In other words, they were looking for a model that would facilitate eager execution that could be modified and debugged during the building process.
PyTorch
PyTorch is a machine learning framework designed based on Torch and Caffe2. This framework is ideal for any neutral design (Talloen et al., 2021). Besides being introduced to the market as an open source, PyTorch aims to facilitate cloud-based software development. The framework is integrated with Python and easily adaptable to libraries such as Numba and Cython.
PyTorch has several features that make it easily acceptable to users. First, it supports eager execution and high flexibility following the inclusion of native Python code, introduced to run on a specific processor. The second feature is that it facilitates easier switching from development to graph mode, thus making it easier to introduce changes in C++ runtime environments (Talloen et al., 2021). Thirdly, it utilizes asynchronous execution and peer-to-peer communication, which enhances the performance of machine learning in production and model environments. Lastly, PyTorch provides an end-to-end workflow that enables users to develop machine learning models in Python and deploy them on various platforms, including iOS and Android.
Sci-Kit Learn
Sci-Kit Learn was introduced to the market as an open source. It is the best option for new clients. It is very user-friendly and has detailed documentation. This framework is easier to run and troubleshoot because the developer has the authority over the algorithm’s preset parameters (Hao & Ho, 2019). According to Hao and Ho (2019), Sci-Kit Learn is one of the most widely used models for data mining and analysis. With its high pre-processing capabilities, Sci-Kit works well as a model design for clustering, classification, and regression.
Sci-Kit Learn has several features that make it popular among new users. Firstly, it can support almost all supervised learning algorithms, such as linear regression, support vector machines (SVMs), and Bayesian models. Secondly, it supports most unsupervised learning algorithms, including cluster analysis, factor analysis, and principal component analysis (PCA). Thirdly, the framework can extract features from both text and images and determine the accuracy level of models on new, unseen data.
The Appropriate Framework for US Mobile
Several aspects must be considered when choosing the appropriate machine learning framework for US Mobile, which aims to deliver exceptional customer service. Although the company has over 110 million subscribers, it must adopt an ML framework to ensure customers maintain their subscriptions. When choosing an appropriate ML framework, the main factor is evaluating the company’s needs.
The following three questions should guide the search for the right machine. Firstly, will the framework be used for deep or classic machine learning? Secondly, what is the preferred programming language for AI? Lastly, what hardware, software, and cloud services will be employed for scaling?
The first question is critical because, besides Python and R, which are the dominant languages used in machine learning, languages such as C, Java, and Scala might be suitable for the company. As the company searches for the most appropriate, it must be aware that most machine learning applications are written in Python today. There is a shift away from R because it was designed initially for statisticians (Koroniotis et al., 2020). Python is regarded as a modern programming language that offers clients concise and straightforward syntax.
The question of hardware and software is critical because many types of frameworks are being introduced to the market. Therefore, US Mobile’s hardware must meet standard patterns (Koroniotis et al., 2020). It is worth noting that GPU acceleration significantly enhances performance, particularly in the ML/AI domain. The company must be aware of the effort required for data analysis and cleanup in preparation for training on the GPU. In essence, CPU can also be used to complement the limitations of GPU, such as onboard memory (VRAM) availability.
Lastly, scalability will help address different concerns related to data and storage systems as well as operational constraints. Building machine learning starts with collecting, storing, and processing large volumes of data. For instance, scalability is concerned with the data analyzed in the training stage and the needed analysis speed. Performance in this case can be improved through hardware acceleration, specifically GPU (Koroniotis et al., 2020).
Building a reliable and accurate ML model will require a substantial amount of source data – even petabytes. The best option here is the public cloud, which meets most of the scalability requirements and also addresses the issue of cost constraints compared to Block storage, which is costly. In the deployment stage of ML, scalability looks at the number of applications that will most likely access the model concurrently. Since there are multiple requirements in the deployment stage, US Mobile would require a framework that facilitates ML development in one environment, such as in the cloud, and runs it in a different environment.
Based on the above factors, the proper framework for consideration is PyTorch – it supports both scalability and meets the requirements of development and production environments. PyTorch provides an end-to-end workflow that simplifies the transition from research to production environments, specifically for mobile devices and privacy-preserving features through deep learning. In fact, PyTorch will make it easier for the company to build deep learning applications, specifically for building and training deep neural networks.
Conclusion
Most major tech companies, such as Google and Facebook, have already integrated machine learning into their operations. As a branch of AI, ML allows computers to self-learn without mandatory programming. As evidenced above, there are several types of ML, such as supervised learning, unsupervised learning, reinforcement learning, semi-supervised learning, and deep learning. For example, reinforcement learning aims to demonstrate how a computer program can act in different situations to achieve the highest possible reward.
Machine learning frameworks exist to facilitate the process of building and deploying ML models. Common ML frameworks examples include TensorFlow, PyTorch, and Sci-Kit Learn. Several factors must be considered when selecting the proper framework, including scalability and evaluating the company’s specific needs. For instance, scalability will help address various concerns and operational constraints related to data and storage systems. Similarly, the company should regard programming languages such as C, Java, and Scala as essential when selecting a machine learning framework.
References
Addagarla, S. K., &Amalanathan, A. (2020). Probabilistic unsupervised machine learning approach for a similar image recommender system for E-commerce. Symmetry, 12(11), 17-83. Web.
Bao, W., Lianju, N., & Yue, K. (2019). Integration of unsupervised and supervised machine learning algorithms for credit risk assessment. Expert Systems with Applications, 128, 301-315. Web.
Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., &Raffel, C. A. (2019). Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 3(2), 134-200. Web.
Hao, J., & Ho, T. K. (2019). Machine learning made easy: A review of scikit-learn package in python programming language. Journal of Educational and Behavioral Statistics, 44(3), 348-361. Web.
Koroniotis, N., Moustafa, N., &Sitnikova, E. (2020). A new network forensic framework based on deep learning for Internet of Things networks: A particle deep framework. Future Generation Computer Systems, 11(3), 91-106. Web.
Leon, V., Mouselinos, S., Koliogeorgi, K., Xydis, S., Soudris, D., &Pekmestzi, K. (2020). A tensorflow extension framework for optimized generation of hardware cnn inference engines. Technologies, 8(1), 6-15. Web.
Madhukar, N. S., Khade, P. K., Huang, L., Gayvert, K., Galletti, G., Stogniew, M., &Elemento, O. (2019). A Bayesian machine learning approach for drug target identification using diverse data types. Nature Communications, 10(1), 1-14.
Mahesh, B. (2020). Machine learning algorithms-A review. International Journal of Science and Research (IJSR).[Internet], 9, 381-386. Web.
Nahar, K. M., Jaradat, A., Atoum, M. S., & Ibrahim, F. (2020). Sentiment analysis and classification of arabjordanianfacebook comments for jordanian telecom companies using lexicon-based approach and machine learning. Jordanian Journal of Computers and Information Technology, 6(3), 45-99. Web.
Nguyen, G., Dlugolinsky, S., Bobák, M., Tran, V., López García, Á., Heredia, I., &Hluchý, L. (2019). Machine learning and deep learning frameworks and libraries for large-scale data mining: A survey. Artificial Intelligence Review, 52(1), 77-124. https://doi.org/10.1007/s10462-018-09679-z
Ozbayoglu, A. M., Gudelek, M. U., &Sezer, O. B. (2020). Deep learning for financial applications: A survey. Applied Soft Computing, 93, 106-384. Web.
Santoso, B. (2019). An Analysis of Spam Email Detection Performance Assessment Using Machine Learning. Jurnal Online Informatika, 4(1), 53-56. Web.
Simmons, C., & Holliday, M. A. (2019). A comparison of two popular machine learning frameworks. Journal of Computing Sciences in Colleges, 35(4), 20-25.
Talloen, J., Dambre, J., &Vandesompele, A. (2021). PyTorch-Hebbian: facilitating local learning in a deep learning framework. arXiv.org e-Print archive, 21(2) 60-88. Web.
Volkmar, G., Fischer, P. M., & Reinecke, S. (2022). Artificial Intelligence and Machine Learning: Exploring drivers, barriers, and future developments in marketing management. Journal of Business Research, 149, 599-614. Web.
Ye, H., Li, G. Y., &Juang, B. H. F. (2019). Deep reinforcement learning based resource allocation for V2V communications. IEEE Transactions on Vehicular Technology, 68(4), 3163-3173. Web.