What is machine learning?

What is machine learning? 

A subfield of computer science and artificial intelligence (AI) called "machine learning" focuses on using data and algorithms to simulate human learning processes and progressively increase their accuracy.

IBM and machine learning have a long history together. One of its own, Arthur Samuel, is credited with creating the phrase "machine learning" through his study on the game of checkers (link goes outside of ibm.com). In 1962, the self-described checkers master, Robert Nealey, competed against an IBM 7094 computer in a game of checkers, losing. This seems like a small accomplishment in the grand scheme of things, yet it represents a significant turning point in the development of artificial intelligence.

Machine learning-based products like Netflix's recommendation engine and self-driving cars are now viable thanks to advancements in storage and processing power over the last few decades.

Machine learning is a key component of the rapidly developing field of data science. Statistical techniques are used to train algorithms to identify relevant information in data mining projects and to generate predictions or classifications. These insights then guide decisions on applications and business, which ideally have an impact on critical growth indicators. As big data continues to expand and thrive, there will be a larger need for data scientists. They will have to help decide which business inquiries are the most important to answer and what data is required to do so.

Typically, frameworks like TensorFlow and PyTorch, which speed up solution development, are used to design machine learning algorithms.

Machine Learning vs. Deep Learning vs. Neural Networks

Given how frequently deep learning and machine learning are used interchangeably, it's important to understand their differences. Artificial intelligence encompasses machine learning, deep learning, and neural networks as subfields. But in reality, deep learning is a subfield of neural networks, and neural networks is a subfield of machine learning.

The method that each algorithm learns is how deep learning and machine learning are different from one another. Labeled datasets, often referred to as supervised learning, are a useful tool for "deep" machine learning algorithms, however they are not always necessary. Deep learning has the ability to automatically identify the collection of characteristics that differentiate various data categories from one another. It can also ingest unstructured material in its raw form, such as text or photos. This allows for the utilization of bigger data sets and reduces the need for some human interaction. As Lex Fridman points out in this MIT lecture (link lives outside of ibm.com), you can think of deep learning as "scalable machine learning".

Traditional machine learning, often known as "non-deep" learning, relies more on human input to acquire knowledge. The set of attributes that human experts need to distinguish between different data inputs is determined; often, this requires more structured data to learn.

Artificial neural networks (ANNs), also known as neural networks, are made up of node layers that have an output layer, one or more hidden layers, and an input layer. Every node, or artificial neuron, has a weight and threshold that are connected to one another. A node is activated and sends data to the following layer of the network if its output exceeds a given threshold value. If not, that node does not forward any data to the following tier of the network. The number of layers in a neural network is all that is meant to be implied by the term "deep" in deep learning. A deep learning algorithm or deep neural network can be defined as a neural network with more than three layers, encompassing both the input and the output. A three-layer neural network is considered a basic neural network.

Neural networks and deep learning are recognized for having accelerated advances in computer vision, speech recognition, and natural language processing, among other fields.

For a detailed examination of the relationships between the various ideas, see the blog post "AI vs. Machine Learning vs. Deep Learning vs. Neural Networks: What's the Difference?"

How machine learning works

Three primary components make up the learning system of a machine learning algorithm, according to UC Berkeley (link outside of ibm.com).

• A Decision Process: Predictions and classifications are typically made using machine learning algorithms. An estimate of a pattern in the data will be generated by your algorithm based on certain input data, which may or may not be labeled.

• An error function is a tool used to assess a model's prediction. An error function can compare known examples in order to evaluate the model's correctness.

• Weights are changed to lessen the difference between the model estimate and the known example if the model fits the training set's data points more closely. This process is known as model optimization. This "evaluate and optimize" procedure will be repeated by the algorithm, which will update weights on its own until an accuracy level is reached.  

Machine learning methods

There are three main types of machine learning models.

Supervised machine learning            

Supervised machine learning, or supervised learning, is characterized by the use of labeled datasets to train algorithms for precise outcome prediction or data classification. The model modifies its weights when input data is entered until a satisfactory fit is achieved. This happens during the cross-validation phase, which makes sure the model doesn't overfit or underfit. Sorting spam into a different folder from your email is just one example of the many real-world problems that supervised learning helps enterprises solve at scale. Neural networks, naïve bayes, logistic regression, random forest, linear regression, and support vector machines (SVM) are a few techniques used in supervised learning.

Unsupervised machine learning

Unsupervised learning, sometimes referred to as unsupervised machine learning, is the process of analyzing and grouping unlabeled datasets using machine learning algorithms. These algorithms find hidden relationships or patterns in the data without requiring human assistance. This approach is perfect for consumer segmentation, cross-selling tactics, exploratory data analysis, and pattern and image recognition since it can identify patterns and similarities in data. It can also be applied to dimensionality reduction, which lowers the amount of features in a model. Two popular methods for this are singular value decomposition (SVD) and principal component analysis (PCA). Neural networks, probabilistic clustering techniques, and k-means clustering are among more algorithms utilized in unsupervised learning.

Semi-supervised learning 

A satisfying middle ground between supervised and unsupervised learning is provided by semi-supervised learning. It guides categorization and feature extraction from a larger, unlabeled data set during training by using a smaller, labeled data set. The issue of insufficient labeled data for a supervised learning system can be resolved through semi-supervised learning. It's also beneficial if labeling sufficient data would be too expensive. 

Reinforcement machine learning

Similar to supervised learning, reinforcement learning is a type of machine learning in which the algorithm is not trained on sample data. Through trial and error, this model picks up new skills as it goes. To create the ideal suggestion or policy for a certain issue, a series of effective results will be reinforced.

An excellent example is the IBM Watson® system that prevailed in the 2011 Jeopardy! competition. In order to determine which square on the board to choose, when to try an answer (or question, as it were), and how much to bet (particularly on daily doubles), the system employed reinforcement learning.

Common machine learning algorithms

Many machine learning algorithms are frequently employed. These consist of:

• Neural networks: With a vast network of interconnected processing nodes, neural networks mimic the structure and functions of the human brain. Because of their propensity for pattern recognition, neural networks are widely used in speech, picture, natural language translation, and image production applications.

• Linear regression: Using a linear relationship between various values as a basis, the linear regression process is used to forecast numerical values. The method might be applied, for instance, to forecast home values in the neighborhood using past data.

• Logistic regression: The supervised learning process known as logistic regression is used to forecast categorical response variables, such as "yes" or "no" responses to inquiries. Applications like spam classification and production line quality control can make advantage of it.

• Clustering: To group data, clustering algorithms use unsupervised learning to find patterns in the data. By pointing out distinctions between data items that humans have missed, computers can assist data scientists.

• Decision trees: Decision trees are useful for categorizing data and for regression, which is the prediction of numerical values. A tree diagram can be used to illustrate the branching sequence of linked decisions used in decision trees. Unlike neural networks, which are a black box, decision trees can be easily validated and audited. This is one of their advantages.

• Random forests: The machine learning approach uses a random forest to predict a value or category by aggregating the output of several decision trees.

Real-world machine learning use cases

Here are a few instances of machine learning that you may come across on a daily basis:

Speech recognition: Also referred to as computer speech recognition, speech-to-text, or automatic speech recognition (ASR), speech recognition is a technology that converts spoken words from humans into written form by utilizing natural language processing (NLP). Speech recognition is a feature that many mobile devices have built into their systems to enable voice search (like Siri) or increase messaging accessibility.

Customer service: As human agents are supplanted by online chatbots across the customer experience, our understanding of consumer involvement on websites and social media platforms is evolving. Chatbots can offer individualized advise, cross-sell products, recommend sizes for customers, and respond to commonly asked questions (FAQs) on subjects like shipping. Examples include chatbots on Facebook Messenger and Slack, virtual agents on e-commerce sites, and duties often performed by voice and virtual assistants.

Computer vision: Computer vision is a branch of artificial intelligence (AI) that allows computers to interpret and act upon meaningful data derived from digital photos, videos, and other visual inputs. Convolutional neural networks are the engine of computer vision, which finds use in self-driving cars in the automotive sector, radiological imaging in healthcare, and photo tagging on social media. 

Recommendation engines: AI algorithms can assist in identifying data trends that can be utilized to create more successful cross-selling tactics by utilizing historical consumption behavior data. Online shops utilize this strategy to suggest suitable products to customers when they are checking out.

Automated stock trading: AI-powered high-frequency trading platforms execute hundreds or even millions of deals daily without the need for human participation, all in the name of optimizing stock portfolios.

Fraud detection: Machine learning can be used by banks and other financial organizations to identify transactions that seem suspicious. A model can be trained via supervised learning with data from known fraudulent transactions. Anomaly detection is able to recognize transactions that are abnormal and should be examined further.

Challenges of machine learning

Without a doubt, the advancement of machine learning technology has made our lives easier. But using machine learning in the workplace has also brought up a lot of ethical questions around AI. A few of these consist of:

The singularity of technologies

Although there is a lot of public interest in this subject, many researchers do not believe that artificial intelligence will soon surpass human intelligence. Superintelligence or powerful artificial intelligence are other terms for technological singularity. "Any intellect that vastly outperforms the best human brains in practically every field, including scientific creativity, general wisdom, and social skills," is how philosopher Nick Bostrum describes superintelligence. Although superintelligence is not a reality anytime soon, the concept of it poses some thought-provoking issues when it comes to the application of autonomous systems, such as self-driving automobiles. Although it is impractical to expect a driverless car to never have an accident, in those cases, who bears liability and responsibility? Should we continue to develop driverless cars, or should we stop at vehicles that are only partially autonomous and aid in safe driving? These kinds of ethical discussions are emerging as cutting-edge AI technology advances, albeit the jury is yet out on this.

AI impact on jobs

Although the loss of jobs is a major concern in the public's understanding of artificial intelligence, this worry should definitely be reframed. Every disruptive new technology causes a shift in the market demand for particular job roles. For instance, if we look at the automobile sector, a lot of companies—like General Motors—are starting to concentrate on producing electric vehicles in order to support environmental initiatives. The energy sector is here to stay, but the source of that energy is changing from gasoline to electric.

Similar to this, professions in other fields will become less in demand due to artificial intelligence. People will be required to assist in managing AI systems. In the sectors of the economy that are most likely to see changes in the demand for workers, such customer service, there will still be a need for personnel to handle increasingly complicated issues. The largest obstacle posed by artificial intelligence and its impact on the labor market will be assisting individuals in making the shift to more in-demand occupations.

Privacy

Data privacy, data security, and data protection are frequently brought up while talking about privacy. In recent years, these worries have enabled governments to take more action. For instance, the GDPR law was developed in 2016 to provide individuals greater control over their data while safeguarding the personal information of citizens in the European Union and European Economic Area. Individual states in the US are creating regulations; one such law is the California Consumer Privacy Act (CCPA), which was unveiled in 2018 and mandates that companies notify customers when their data is being collected. Such laws have compelled businesses to reconsider the way they handle and maintain personally identifiable information (PII).

Prejudice and insensitivity

Many ethical concerns about the application of artificial intelligence have been brought up by instances of prejudice and discrimination in a variety of machine learning systems. Given that biased human processes may have produced the training data itself, how can we protect against prejudice and discrimination? Reuters (link lives outside ibm.com) outlines some of the unintended implications of adopting AI into recruiting methods, even though most organizations have good intentions when they automate. Amazon accidentally discriminated against job candidates based on their gender for technical tasks in an attempt to automate and streamline a process; as a result, the business had to abandon the initiative. The Harvard Business Review (link is external to ibm.com) has posed additional insightful queries regarding the use of AI to hiring procedures, including what information ought to be available for use in the candidate evaluation process.

Discrimination and bias aren't just present in the human resources department; they may also be discovered in a variety of applications, such as social media algorithms and facial recognition software.

Businesses are participating more actively in the conversation on AI ethics and values as they become more conscious of the hazards associated with AI. IBM, for instance, has discontinued its all-purpose facial recognition and analysis software. In a statement, IBM CEO Arvind Krishna stated that the company "firmly opposes and will not condone uses of any technology, including facial recognition technology offered by other vendors, for any purpose which is not consistent with our values and Principles of Trust and Transparency, or for mass surveillance, racial profiling, or violations of basic human rights and freedoms."

Responsibility

There is no true enforcement mechanism to guarantee that ethical AI is used because there isn't any meaningful regulation governing AI practices. The present financial consequences of an unethical AI system serve as an incentive for businesses to act morally. In order to close this gap, ethicists and researchers have worked together to develop ethical frameworks that regulate the creation and application of AI models in society. But for now, these are just meant to be used as a guide. According to some study (link sits outside of ibm.com), it is not possible to prevent harm to society when there is a lack of awareness of potential effects combined with divided accountability.

How to choose and build the right machine learning model

It can be difficult to create the ideal machine learning model to handle a given problem. Building an ML model involves experimentation, ingenuity, and diligence, as outlined in the seven-step strategy that is summarized here.

1. Recognize the issue facing the company and establish success standards. The purpose is to translate the group's understanding of the project's goals and the business problem into a problem formulation that is appropriate for machine learning. It is important to ask why machine learning is needed for the project, what kind of algorithm best suits the challenge, if transparency and bias reduction are necessary, and what the expected inputs and outputs are.

2. Recognize and define the data requirements. Assess the data to see if it is suitable for model ingestion and what is needed to build the model. How much data will be required, how the data will be divided into training and test sets, and whether or not a pre-trained machine learning model may be applied are all important questions to ask.

3. Gather and ready the information needed to train the model. Cleaning and labeling the data, replacing inaccurate or missing data, augmenting and improving the data, eliminating ambiguity and noise, anonymizing personal data, and dividing the data into training, test, and validation sets are some of the actions that are taken.

4. Ascertain the features of the model and train it. Make the appropriate algorithm and approach choices. After the model has been trained and validated, optimize it by adjusting its hyperparameters. Machine learning algorithms can include natural language understanding features like recurrent neural networks or transformers made for NLP tasks, depending on the type of business problem. Decision tree models can also be optimized through the use of boosting algorithms.

What are the advantages and disadvantages of machine learning?

Businesses that use machine learning well can gain a competitive edge over those that only use conventional statistics or human intelligence. This is because machine learning can spot trends and anticipate outcomes with more precision. Businesses can gain from machine learning in a number of ways:

• examining past data in order to keep clients.

• introducing recommender systems to increase income.

• enhancing forecasts and planning.

• evaluating trends to find fraud.

• reducing expenses and increasing efficiency.

However, machine learning has drawbacks as well. It can be costly, first and foremost. Most machine learning projects are led by highly compensated data scientists, or data scientists. Additionally, software infrastructure—which can be costly—is needed for these projects. And companies may face a plethora of additional difficulties.

The issue of bias in machine learning exists. Inaccurate models of the world that, at best, fail and, at worst, are discriminatory might result from algorithms that were trained on data sets that omit specific populations or contain errors. Enterprises that build their key business processes on biased models run the risk of facing regulatory action and damage to their brand.

Importance of human-interpretable machine learning

When an ML model is sophisticated, it can be difficult to explain how it operates. Because it's critical for the company to provide an explanation for each decision it makes, data scientists in certain vertical industries are forced to employ basic machine learning models. This is particularly true for sectors like banking and insurance where there are significant compliance requirements. Data scientists frequently have to choose between a model's efficacy and accuracy and its transparency. Although complex models are capable of producing precise forecasts, it can be challenging to explain how an output was derived to a novice or even an expert.

Machine learning examples in industry

Several industries have embraced machine learning. The following industries are adopting machine learning to suit market demands:

• monetary services. Financial services organizations use machine learning in risk assessment, algorithmic trading, client service, and personalized banking. For instance, Capital One used machine learning (ML) for credit card protection, which falls under the larger umbrella of anomaly detection.

• prescription drugs. ML is used by pharmaceutical companies in clinical trials, drug discovery, and medication production. For instance, Eli Lilly has developed AI and ML models to identify the ideal locations for clinical trials and increase participant diversity. The company claims that the models have significantly shortened the duration of clinical trials.

• production. Use cases for predictive maintenance are common in the manufacturing sector, where costly production delays can result from equipment failure. Furthermore, machine learning's computer vision component may check products as they leave a manufacturing line to provide quality control.

• Coverage. Recommendation engines are able to provide options to customers based on both their needs and the experiences of other users with particular insurance products. Processing claims and underwriting both benefit from machine learning.

• shopping. Retailers utilize computer vision for inventory management, personalizing, and designing the styles and colors of a particular fashion line in addition to recommendation systems. One such important use case is demand forecasting.

What is the future of machine learning?

Driven by copious amounts of research conducted by corporations, academic institutions, and governments worldwide, machine learning is a rapidly evolving field. AI and ML breakthroughs appear to occur every day, making conventional wisdom nearly instantly out of date. About machine learning, one thing is certain: it will remain a major force in the twenty-first century, revolutionizing both the way we live and conduct business. 

In NLP, enhanced infrastructure and algorithms will result in more conversational AI that is more fluent, more versatile machine learning models that can adjust to new tasks, and customized language models that are tailored to specific business requirements.

It is anticipated that the rapidly developing field of computer vision will have a significant impact on a wide range of fields, including environmental science, software engineering, healthcare, and virtual and augmented reality. As technology advances, computer vision will become more crucial for diagnosis and monitoring in these fields.

Machine learning platforms are one of the most competitive areas of enterprise technology in the near future. Prominent providers such as Amazon, Google, Microsoft, IBM, and OpenAI are vying for clients' business by offering automated machine learning platform services that encompass a wide range of machine learning tasks, such as data gathering, data cleansing, data classification, model development, training, and application implementation.

Despite the excitement, businesses will encounter many of the same difficulties as those posed by earlier cutting-edge, quickly developing technology. Adapting legacy infrastructure to machine learning systems, minimizing ML bias, and figuring out how to best leverage these incredible new AI capabilities to produce profits for businesses despite the associated expenses are some of the new hurdles.