Machine Learning

From New Media Business Blog

Jump to: navigation, search

Machine Learning is the application of statistics and algorithms in an approach to allow programs to learn, think, infer, and predict. Machine learning utilizes tools from many fields including mathematics, computer science, biology, and epistemology. Much of the algorithms in machine learning are inspired to natural processes like annealing and neural networks. Most machine learning is divided into supervised and unsupervised learning. Supervised learning allows computers to learn given a set of data and their correct tags. Unsupervised learning does not provide a specific correct tag to a given data, rather an objective is provided for the program to maximize. Current machine learning is limited by hardware and software. Faster processors and more memory are required to improve learning and responding times. Software in efficient data management and process panellization will be required as machine learning improves. The increasing need for business automation will see machine learning improve by leaps. Machine learning is used in both manufacturing and services industry. Assembly lines are heavily reliant on machines that can operate within logical rules and be able to infer what actions to take. Logistics, cryptography, and banking also rely heavily on machine learning. Package routing, collision detection (for cryptography), and fraud detection are some examples. In the future, it is possible to see machine learning in autonomous vehicles and disaster response robots. Machine learning has also been the theme of many films and books including The Terminator series, iRobot, and AI.



IBM Watson

The study of human intelligence and epistemology has been around since the ancient Romans. While modern machine learning derives many of its fundamentals from ancient philosophy and mathematics, the main focus of this article is on computational machine learning. [1]

IBM Deep Blue
  • 1950: Alan Turing proposes the Turing Test for machine intelligence.
  • 1950: Isaac Asimov publishes his laws of robotics.[2]
  • 1956: The term Artificial Intelligence (AI) is coined.
  • 1997: IBM's Deep Blue chess playing computer defeats world champion Garry Kasparov.
  • 2009: Google builds a self-driving car.
  • 2011: IBM's Watson competes on Jeopardy.
  • 2011: Introduction of Apple Siri.

For its origin as a thought experiment for philosophers and mathematicians, machine learning has evolved and permeated modern life. It can be said for certain that without machine learning, many of the technologies we take for granted today would not exist. From this trajectory, it can be seen that machine learning has moved on from its tasks specific ancestors to flexible multi-use software for the everyday consumer. In its early life, humans have been concerned about the prospects of autonomous machines and such issues are only going to more prominent as machine learning becomes more pervasive.

Machine Learning

Supervised Learning

Machine Learning comes in many shapes and forms. The two more prominent are supervised and unsupervised learning. Supervised learning utilizes human knowledge to aid in learning. Unsupervised learning allows the computer to learn by itself and while this process may take longer it results in unique and novel solutions that surpass human imagination. The main task of modern machine learning is pattern recognition and classification.

Supervised Learning

Supervised learning uses human tagged solutions to teach the computer. This is similar to the way children are taught in schools. This has been a popular method due to its speed, efficiency, and accuracy. Popular uses include handwriting recognition, speech pattern analysis, spam mail detection, tumour diagnosis from patent data, and classification. Some of the more popular methods for supervised learning are logistic regression, nearest neighbours, random forest, support vector machines (SVM), naive Bayesian networks, and neural networks. On a basic level, what these methods aim to achieve is to solve an equation that seeks to reduce the error during learning phase. There are other ways to improve learning for computers. k-fold cross-validation does not require more data, rather it splices existing data into k sections then uses the permutations of the k sections as input for learning. Boosting is also another technique to make a strong machine learning method from weak ones. The premise is that multiple weak learners together are able to learn and predict with the same ability as a strong learner.

Unsupervised Learning

Unsupervised learning does not tell the computer what is right or wrong. Rather it tell it an objective to maximize and lets it loose. This is popular because it allows for the compute to use its imagination rather than conform to a human model of solutions.



For hardware, the current limitations are in size and energy efficiency. The biggest portion of current robots is batteries. Not only to move the actuators but also for the computational aspect. Current generation machine learning is stifled by the speed of computation. Some have tried to circumvent this by using server farms, but such a solution is infeasible for robotics. Machine learning responses is also limited in how much input they can take.

For software, the current limitations are in algorithms. Some problems are very hard to solve in their nature and the only solution present are unable to achieve a solution within a feasible time. The problem of the millennium for computer scientists and mathematicians alike the the P vs. NP problem. In essence, it is the question of whether the set of problems solvable in polynomial time (as a function of its input) is the same set of the problems that are solvable in non-deterministic polynomial time. There are many pros and cons to the solution to this problem. A benefit is that is there is a solution, many problems that take a long time to solve can be solved in a fraction of the current time. As an example, Coca-Cola uses a NP problem known as 3-SAT to determine its operations and manufacturing. A downside to the solution is that many of the algorithms modern technologies rely on will have a solution. This is the biggest problem for cryptography. Cryptographic methods currently are based off the RSA algorithm that uses the NP property of prime number factorization to make passwords and other sensitive data impossible to solve in a lifetime.[1] This also applies to machine learning. For a computer being able to solve a 3-SAT problem in polynomial time means that the knowledge representation methods used today can be more complex, allowing computers to store and infer more complex problems. Coincidentally, scientists and researchers are using machine learning to find a solution to the P vs. NP problem. Machines are able to do the calculations much faster and be able to see patterns in the data and solutions that would otherwise go over the heads of humans.

For supervised learning, training requires labelled training and testing data. The dataset are usually in numbered in the hundreds of thousands of samples. For each data point multiple feature data may also be required. This usually ends up as a matrix exceeding one million entries. As it is supervised learning, the tagging of the dataset must be done by humans. Not only is this time consuming, it may also contain human errors or biases. Two researchers may tag a same picture with different targets as a typical image seldom contains one defining feature.

For unsupervised learning, the benefit of unsupervised learning is that it does not require human attention, however it typically takes much longer than supervised learning. Without guidance, the computer must compute multiple permutations, some of which are redundant or useless. In addition, without a properly define objective function, a computer may optimize to an non-optimal solution.

Business Uses


Spirit Rover

General Motors

To increase efficiency and reduce faulty products, General Motors (along with various manufacturers) have employed machine learning in conjunction with their assembly line robots. While these robots may be limited in learning capability by design, they are able to make complex, flexible, and pre-programmed decisions based on inputs. This decreases the need for constant supervision and frees up resources to be used else where. The complexity of the tasks and the design of the robot will determine the amount of machine learning it will need. Complex tasks such as multiple welding points will require the robot to accept variable measurement tolerances, while complex robot design will require the robot to calculate the best positions of its appendages.



Signals from Earth to Mars currently take +20 minutes to reach of rover drone. During that time anything could of happened to the drone or the command could come in too late. For this very reason, both the Spirit and Opportunity rovers for NASA's Mars mission have a high degree of machine learning built in.[2] In 2010, the Spirit rover lost contact with mission control due to loss of power from the harsh winter of Mars. While out of contact, the rover was able to run internal programs that rotated its solar panels and batteries towards the sun.[3]


Perhaps the mouse famous robot in the world, Toyota's ASIMO was the first robot to climb stairs and run without the assistance of supports. To perform these actions, ASIMO's internal computer must be able to collect all the data from its sensors, reviews its historical trajectory, and be able to predict where it should move in order to maintain balance.[4] If these actions were done through traditional programming with a permutation of every possible case, the list would be endless. Rather, the programmers needed ASIMO to be able to decide for itself given any situation, thereby allowing them to cover most of the possible cases and allowing ASIMO a high degree of autonomy.



Google uses machine learning in its search algorithm and its advertising targeting. Predictive search utilizes natural language processing in conjunction with Bayesian networks and Hidden Markov Models to determine what is the most likely target search string a user is looking for. Based on a user's search history, Google also uses classification algorithms to determine the best search result to show first and also which advertisements to display. It should be expected that the feature space Google is using to determine the classification has high dimensions given the amount of data they collect on users.



Apple Inc.'s Siri program is heavily reliant on machine learning. It utilizes voice recognition software to first determine what the user says. Using Natural Language Processing, it then interprets the command and associates part of speech tags like noun and verbs to the command. Based on its analysis of the speech, it will then use a data connection to the internet to find what the user needs. Siri's programming also learns preferences of the user based on multiple uses, make the results more accurate. Voice recognition requires machine learning because of the variability of voices and accents given a language, therefore the more the user speaks to it the more accurate it gets.


Amazon Recommendation Engine

Netflix famously proposed a challange for computer scientists back in 2009. It needed a better recommendation system to suggest to users what they would like to watch. When someone watches something on their services, Netflix collects that data and classifies the user into a certain group. Based on the groupings, it may recommend something that is a prominent feature of the group that the user has not watched yet. Netflix put up a $1 million prize for this competition. [5]


Amazon uses support vector machines in its recommendation engine for ups-elling and cross-selling. It uses viewed and bought items to best determine what the user may like to purchase or view next. [6]


Facebook believes machine learning to be key to its future, so much so that it hired Yann LeCun, a professor of Deep Learning at NYU, to run its new artificial intelligence deparment in 2013. [7]. Facebook hopes that AI will allow it to better understand its users and allow for better targeting of advertisement. The inferences and insights it generates will also enable it to understand consumer trends and preferences, giving it an edge in predicting the competitive environment.


The massive amounts of data that world intelligence agencies collect requires massive computational powers to sift out patterns and detect anomalies. Such tasks are the work of artificial intelligence machines. These tasks range from finding and tracking a person of interest, detecting suspicious and anomalous activity, or detecting and predicting intentions. [8]

irobot Roomba


Perhaps the most known use of AI in robotics for consumer home use, irobot's Roomba sweeping robot uses machine learning in multiple ways. Spacial detection, obstacle avoidance, path finding, and automated docking are prominent features of the Roomba that use machine learning. People have been so accustomed and trusting of it that it has been used to clean a house without human supervision. It is understood that the robot will not knock over furniture or break something in the house while no human is present.


Machine learning presents many benefits for businesses in many different sectors.

For manufacturers, machine learning allows for:

  • Reduced cost for workers
  • Improved workplace safety
  • Improved product quality in quality assurance phase
  • Decreased development and manufacturing time
  • Improved customer satisfaction and customer retention using artificial intelligence in call centres[9]

For marketers, machine learning allows for:

  • Improved customer targeting through pattern recognition
  • Higher conversion through customer preference classification
  • Discovery of new customers and customer segments through classification analysis

For accountants, machine learning allow for:

  • Reduction in accounting errors through self-learning checking software
  • Improve auditing by computers and reduction in accounting fraud
  • An impartial third-party to prevent biases or briberies
  • Improve book keeping and reduction in entry time through hand writing recognition for receipts and expenses

For finance, machine learning allows for:

  • Split-second trading decisions in equity and derivatives
  • More comprehensive portfolio balancing and optimization
  • Improved hedging of assets using market volatility predictions


Machine learning, while having many benefits, also has some dangerous and potentially fatal downsides:

  • Reduction in human supervision make catching errors slow
  • May cause fatalities or faulty production in there is bug in code
  • Can run in contrary to programming if objective is ill-defined
  • Initial costs may be high, especially at an early stage of machine learning
  • High cost of electricity and data bandwidth if ran for a long time
  • Loss of human jobs with increased automation
  • Lack of responsibility or accountability on machine's part


When a business is considering using machine learning in its operation there are some things it should consider:

  • Will it provide measurable increases in profit/efficiency/safety or am I following a fad?
  • What competitive advantages does it provide?
  • How will the use of machine learning complement current operations?
  • How will machine learning affect company culture or organizational structure?

A business should not invest in machine learning if it must over-haul its operations just to use machine learning. Machine learning should complement existing operations rather than hinder it. While it is a company's objective to increase profits, it must also consider what the implications are for its company culture and workers. As with the introduction of robots to assembly lines, there may be a large increase in loss of jobs with the introduction of machine learning in the workplace.

Future Developments

Google Driver-less Car

Artificial Intelligence is one of the fastest developing markets currently. In conjunction with advances in other fields of electronics, AI is rapidly evolving. Driver-less cars like Google's utilize machine learning heavily to learn the rules of the road and how to navigate obstacles. As machine learning get more advanced, common consumers can experience machine learning in a physical medium (i.e. robots) which had been reserved for specialized tasks.

Boston Dynamics Atlas

In hardware, newer and faster processors will allow for faster computation in learning as well as calculations in predicting responses. Smaller storage mediums will allow server farms and workstations to function more efficiently as well as access data more rapidly. Advancements in hardware interfaces like PCI, eSATA, and USB will cut down on bus transfer times and dramatically improve both learning times and electrical efficiency. The development of quantum computers will provide perhaps the best benefit to AI. Calculating in q-bits will allow simultaneous k^n instruction sets rather than 2^n on a conventional binary system.

In software, more efficient algorithms will reduce learning time by utilizing more advanced mathematical principals. Software like HDFS in conjunction with a distributed storage will allow for parallel calculations across scaled machines, dramatically reducing computing times.

For consumers, this means a closer step to the realization of robots in daily lives. By reducing the size and energy consumption of these components, it would empower researchers with the tools needed to make robot helpers a reality. Current versions of robots like those from Boston Dynamics are very large and cumbersome. Their movements are slow and their movement range is rather inflexible. The most constraining factors to current robots is the need for wires and metal supports. While ASIMO may be able to walk on its own, the tasks that it can accomplish are very limited and it cannot learn new tasks. It is the vision that learning robots like Atlas (Boston Dynamics) will have the mobility of ASIMO.

For businesses, the benefit using machine learning is the lower variable cost associated with the cost of business. While a human worker has wages in addition to health benefits and social needs, a software can operate 24-hours without tiring and can be changed on the fly. The cost of an automated worker is also mostly fixed with the variable cost being electricity.

Business Implications

Business automation has been happening on a hardware scale since the 20th century with assembly lines. Machine learning now presents a way allow business automation on a software scale. This now only allows automation in manufacturing but also in services like customer service.

For new industries, this will prove new opportunities. With the machine learning revolution in the horizon, all aspects of business will feel its impact. New industries and sectors will be made of this revolution. Fully automated customer service may force the old customer service model and call centres into the shadows of history. Automated vehicles will present new allow the birth of new car manufacturers that solely focus on automation. For computer scientists, this development will be the most exciting. The demand for AI specialists will rise exponentially as human reliance of AI increases.



Governance / Accountability

Some recent events have highlighted the need for improved machine learning governance and accountability. The Flash Crash of 2008 was in part caused by an automated trading system using machine learning. Flash Crash [10]. These system are popular with big trading firms because of their speed and efficiency, however recent opponents of automated trading like Warren Buffet have caused crackdowns[11]. The dangers with automated trading is that they an trade in large sums on a whim, sometimes without reason if there is a bug. With something as important as the financial system, there needs to be improved accountability in the event of an error.

Though some people look to Asimov's 3 Laws of robotics to protect humans from an extinction scenario, critics are quick to point out the flaws in it.[12][13] While Asimov's 3 Laws may be a good starting point, the laws governing robotics and machine learning has to be more comprehensive and updated to the current technological standards.

Many prominent figures such as Elon Musk and Professor Stephen Hawkings have recently spoken out against un-fettered AI development.[14][15]. The scenario they envision is that of a superior machine intellect that replaces humans.


In recent times, many prominent academics and business people have spoken out against rampant artificial intelligence. It is their fear that without proper control and rules governing the use and development of AI, humankind could be putting itself in danger. These dystopia views are in ways similar to the Terminator series. While the future may not hold such immense dangers, it is wise to approach AI with caution. In the near future, governments and governing bodies will need to establish legislation regarding the use of AI. Many components of society like business and warfare have great hopes in AI, yet these components are the most susceptible to the dangers of AI. Beyond that, AI ethics will also be a topic of interest. It needs to be considered if human ethics should apply to machines that can think intelligently.


With systems like the automated crime detection or Facebook's user analysis, privacy become a major concern. While the prying eyes are not that of a human, users may still feel their privacy being infringed upon, as the information the machine learns and predicts will be used by humans to make real life decisions. For a crime system, people may feel the presence of a big brother always watching their actions.

Business Takeaways

For businesses, machine learning will provide many benefit. The reduction of costs or improved safety look very tantalizing. Yet, machine learning is not without its consequences. The loss of jobs or an un-supervised machine can spell the demise of a business. Businesses much use machine learning for the right reasons so that it benefits their operations. New businesses will be born of a machine learning revolution, so it is key to understand it in order to maintain or gain competitive advantage. Machine learning is only going to improve from here, it is imperative that rules and laws be established in a proactive manner rather than reactive. With so many people voicing concerns over a dystopia future in machine learning, humans should approach its research and development with trepidation.


In Popular Media

Whether they be an utopian Eden or an Orwellian dystopia, Artificial Intelligence has often been the theme of many books, movies, and plays. Popular films that directs their story around AI include the Terminator series, iRobot, her, and AI. AI is also very prominent in games and it has also been showcased to be able to play games as well.


Personal tools