Machine Learning Explained

Machine learning (ML) enables computers to learn and interpret without the dependence of rules-based programming.  Computers use and process the models that are exposed to sets of new data. They adapt independently and learn from earlier computations to interpret the available data and identify hidden patterns. This requires data analysis and the automation of building analytical models using numerous ML algorithms. Powerful machine learning capabilities commission computer workstations and servers to search for and identify hidden insights, without being programmed for where to look for, and when exposed to new data sets.  Machine learning algorithms takes the automated process of parsing the data, learning from that data, and then applying what was learned.

Artificial intelligence methods and technology are not new.  The broad aspects of it is being exposed to a wider and greater set of programmers and researchers as new tools, hardware and software have made it much less costly to access and deploy.  Researchers can apply ML algorithms to develop new analytical models across numerous applications to uncover relationships, trends and patterns with minimal or no human intervention.  Machine learning has evolved due to more advanced and faster computing processors and as researchers set to discover if computational models can be developed to mimic human learning patterns from data.   As new data and gained knowledge from patterns, computations and decisions are created from similar situations. The computers iteratively make progress in learning while applying increasingly complicated mathematical calculations to huge data sets, and at accelerated rates. Driverless cars are primarily built on the basis of machine learning and processing units.  Recommendation engines in e-commerce, advertising, and media are also built using large processing power, storage and ML algorithms that is in practical use everyday.   Personal assistants through home speakers like Alexa, Google and Siri have used ML to recognize speech patterns for use in mimicking human communications. 

It is also important to discern the 3 different domains of artificial intelligence computing so that we have a more natural understanding of it from the business point of view. 

1.Machine learning systems deals with data or the information side of AI. Computer machines uses data with the aim to derive meaning.

2.Cognitive systems deals with the human aspects side of AI. Computer machines are used as a communication interfaces which include the ability to have conversations, interaction, and collaboration with humans.  This involves layers of neural networks that can make the right intelligent decisions on its own.

3.Robotic systems deals with the physical world of AI. Computer machines are used as mechanical enhancements to people and processes.

When learning, cognition, and robotics are seamlessly integrated, we get to a closer resemblance of what artificial intelligence really looks like. AI makes up the technologies that powers computing machines with human-like abilities.  

We are in the new era where we take advantage of the computational processing power and speeds to allow computers to do the work for us: identify and solve problems.  Just as such how the first hand-held calculator made it so much faster for humans to do arithmetic operations, we apply data sets beyond numbers and can take images, words, speech, and all types of data to let computers automate the process of analyzing hidden trends, patterns and predictions.  A machine learning algorithm usually follows a certain type of data and then uses the patterns hidden in that data to answer more questions. For example instructing a computer a series of photographs. A number instructions of the images f will say “this is a horse” and a number of them some say “this is not a horse“.  After this drill, you feed it more photos while it identifies if is true or false.  The correct and incorrect guess is recorded into memory. The computer gets smarter as it processes and records more results and data over time.

For the machine to learn, data scientists and researchers use the appropriate algorithms with the right tools to build the machine learning models based on iterative learning processes. Some of the most common ML training algorithms used by businesses today include:

• Decision Tree
• Linear Regression
• Logistic Regression
• Conditional random fields
• Least squares and polynomial fitting
• Support vector machine (SVM)
• Naive Bayes
• kNN (nearest neighbor)
• K-Means ( k clustering)
• Random Forest
• Principal component analysis (PCA)
• Convolutional and recurrent neural networks

Machine learning goes through the process of teaching a computer program or algorithm how to progressively and iteratively improve upon a set of tasks.  You can implement 3 types of machine learning:

1. Supervised Learning

This type of task-oriented learning is similar to teaching with flash cards.  You keep feeding example-label pairs one by one and allow the algorithm to predict the label for each of the examples correctly and incorrectly. Over time, the algorithm will learn to approximate the nature of the relationships between the examples and labels.   You’ll see this type of learning for applications in facial recognition, classifying objects and information, and matching. Examples of supervised learning algorithms: regression, decision tree, random forest, kNN, logistic regression.

2. Unsupervised Learning

In this algorithm, we do not have any target or outcome variable to predict / estimate. There are no labels to the example or data. The algorithm is fed a lot of data to understand its properties. It learns to group, cluster, or organize the data to a point which the data starts to make sense in a new way. You’’ll see unsupervised learning in applications such as recommendation engines, purchasing behaviors and customer segmentation, intelligent customer service, support and issue resolution.  Examples of unsupervised learning algorithms: Apriori and K-means.

3. Reinforcement Learning

Reinforcement learning algorithms are designed to learn from experience or mistakes. It is analogous to teaching through rewards and punishment. Actions that yield the best rewards are identified by algorithms that use trial and error methods.  The algorithm will make positive signals to positive behaviors, and make negative signals to negative actions.  There are three major components in reinforcement learning, namely, the agent, the actions and the environment. The agent in this case is the decision maker, the actions are what an agent does, and the environment is anything that an agent interacts with. The computer is exposed to an environment where it trains itself continuously through trial and error and rewards.  We can also reinforce the algorithm to prefer bad actions to make more mistakes – rather than good actions and less mistakes.   It captures the knowledge gained.  You’ll see this applied in the area of video games, assembly lines with robotic automation, and power management in data centers and self-healing networks.  Markov Decision Process, Q-learning based on the Bellman equation, State-action-reward-state-action (SARSA) are examples of reinforcement learning methods.

After you get a handle on the algorithms, ML requires discipline in working with the appropriate tools and processes:

• Collection of quality data based on the identified problems that you aim to solve
• Preparing the data and choosing a measure of success.
• Choosing a model for the different tasks and how to evaluate the basis of which it can be improved
• Training the model with increasingly accurate data
• Evaluating the model including the selection of the measure
• Tuning and tweaking the model parameters
• Testing models with additional data sets and estimating how the model will perform to address questions and predictions.

Why is Machine Learning and AI (in general) increasingly important in business today?

In summary, information, the speed of information, and the speed in which action and response has always been sources of competitive advantage for most businesses.   It can also lead to greater efficiencies where humans do not to see, understand, or react on their own, and to take the actions at the right time – with higher levels of accuracy.   When Benjamin Franklin when said ‘time is money’ in 1748, he was referring to time in the sense of labor, wages and opportunity costs. In the modern information age, it still holds true.  Computers have helped humans and businesses to reduce the time it takes to do things, crunch data, make decisions, and power machinery– which all amount to time and money.  Businesses can apply machine learning in all functional areas and operations to gain increases or reductions.  Research has shown that companies that leverage data to the fullest have higher levels of performance compared to companies that don’t. ML gives the job to computers that can constantly and  automatically process/analyze the data with no sleep and fatigue.  The results could mean increased accuracy, predictions, increased savings and sales, or even better customer experiences.    Most of the industries dealing with huge amounts of data have now recognized the value of machine learning. By gleaning hidden insights from this data, businesses can work more efficiently and can also gain a competitive edge from its competitors or entering new markets.  Affordable and easy computational processing and cost-effective data storage options have made it feasible to develop models that quickly and accurately analyze lakes of complex data.

Machine learning is now being implemented by many organizations across the globe. With machine learning applications gaining prominence, here’s what it takes to develop effective machine learning systems for your business:

• Start with a strategy and vision of the problems you want to solve. Think big, but do small.
• Don’t over promise. Set a feasible set of expectations, and commit.
• Acquire superior data collection, cleaning, data preparation and visualization capabilities
• Bring in staff or consultants with skills and knowledge in advanced algorithms and mathematics
• Use tools that have a track record of success or maturity.
• Put in the scalable computing infrastructure and resources in place (ahead of time) so that data scientists can do their job with efficient resource capacity to minimize wait times.
• Establish a business improvement manager or team that understands the specified areas of business functions, processes, and industry.
• Get the right department staff involved in the process. Spend time on the problem with them – not the solution.
• Work with the team to define at least 1 technical measure and 1 business related measure for a specific problem. Don’t define metrics that can solve world hunger here.
• Evaluate progress and expectations. Re-evaluate feasibility and scope of projects and deliverables.
• Understand that projects will run into problems and challenges, so fail fast. Test and test again. Be sensitive to false positives.
• Communicate. Communicate, and communicate.

Depending on the purpose of the business, business executives will commonly ask where is ML being applied?  What are the top use cases for machine learning beyond eliminating minor repetitive tasks?

Marketing and Sales

Discover and predict behaviors, next best actions, offers, and forecasts. Provide personalized product recommendations for purchase or likes. Automate journeys in products, services and experiences.  Hyper-target advertising

Financial Services and insurance

Identify key insights into financial data. Prevent financial fraud for companies and individuals. Identify opportunities for investments and trade. Automate claims processing. 

Information security

Machine learning helps businesses improve their ability to analyze threats, stop attacks, and respond to malicious activities faster. Protect confidential data. Analyze networks and vulnerabilities.  

Healthcare and health science

Machine learning helps the healthcare and science industry accelerate insights. ML is being used to monitor and predict health threats, analyze diagnosis and treatments. Use data from external data factors such as weather, pollution, exposure, living and economic environment to predict medical treatment and disease development and identify patent risks.  Genetic and molecular drug discovery, development and simulations. Improve diagnosis. Reduce the costs of medical record keeping. Use advanced 3D radiological images for radiotherapy and surgical planning.  

Transportation

Predict potential problems that could arise such as traffic management and congestion. Monitor driver behavior, safety and security. Automate driver and passenger experiences in self-driving vehicles. Predict maintenance of commercial vehicles. Improve cargo and logistics. Detect defects and erosions of train tracks to improve locomotive efficiency and reduce failure rates due to environment factors.

Government and public sector

Implement simulations across a number of fields of government such as public services, criminal investigations, energy, and defense. Analyze underground minerals, geology, and discover new energy sources, including environmental impact.