Without machine learning, machines need to be specifically programmed by a computer programmer because that is the only way the machine will know the necessary steps to accomplish a task. However, with machine learning, the dependency on programmers is reduced. A programmer still has to program the machine so that it has the capability to learn, but then the machine is capable of gathering data on its own and coming up with the best steps to solving a program. A programmer no longer has to give the machine that specific ability. This concept means that not only can a machine perform more than what it was simply programmed for, but machines can be used in multiple different environments where no programmer is involved. Machine learning allows machines to be more usable by people outside of computer science. Below are some examples where machine learning can benefit the general population.
Robotic Learning Examples
HERB (Home Exploring Robotic Butler)
The people who built HERB at Carnegie Mellon University under the guidance of Professor Siddhartha Srinivasa did so with the goal of creating a "truly autonomous robotic assistant that can perform routine tasks at human speeds in cluttered environments like homes or offices" (Smith). HERB learns tasks through watching humans and then imitating the humans' actions. It also works on finding better methods for performing tasks. For example, an early version of HERB grabbed a coffee cup to place in the dishwasher with one of his thumbs down. The researchers realized that this was a more efficient method that professional bartenders use (Smith).
HERB Separates OREO
The video to the right demonstrates another example of HERB using his learning capabilities. HERB was once tasked with separating OREO cookies. He was asked to separate the cookie, remove the cream, and present the polished half to a human for their consumption. He knew what the input and the desired output were, but he still needed to figure out the steps needed to reach that outcome. Through trial and error, he eventually figured out the necessary steps. In addition, each time he smashed the cookie, thus making it impossible to finish the task, he immediately knew he needed to try a different task.
|
|
More on HERB
If you are interested in learning more about HERB, see these websites:
http://www.scientificamerican.com/article/robot-learning/
http://www.cmu.edu/herb-robot/
If you would like to see more of HERB in action, check out the videos found at
http://www.youtube.com/playlist?list=PL1HxVG_mcukv3-scZeRCnadHafor8jvzZ
http://www.scientificamerican.com/article/robot-learning/
http://www.cmu.edu/herb-robot/
If you would like to see more of HERB in action, check out the videos found at
http://www.youtube.com/playlist?list=PL1HxVG_mcukv3-scZeRCnadHafor8jvzZ
HERB's Learning Methods
Although no information is given regarding the algorithms used in HERB, some possible conjectures can be made regarding the possible algorithms based on how he completes his tasks. We came to the conclusion that for the above examples, HERB uses transductive and semi-supervised learning.
- Semi-Supervised Learning: In both the OREO and the coffee cup cases, HERB was given the input and the output and then developed the steps to reach the output.
- Transductive: HERB learned from his past experience to decide the best methods. With the coffee cup, he was shown the way humans usually do it and then further developed the idea to a more efficient method. With the OREO cookie, HERB used trial and error, and thus his past experiences, to figure out the best method to separate the OREO.
Apprentice Robots
http://video.mit.edu/watch/robots-that-learn-from-people-39/
This video focuses on creating apprentice robots. These robots are capable of watching a human perform a task and then figuring out how to mimic the motions. This allows apprentice robots to perform tasks that are difficult to program, such as folding towels. Each time the robot sees the towel, the towel will appear differently. So, how do programmers take all of these situations into account? They cannot, so they let the robot figure it out instead.
This video focuses on creating apprentice robots. These robots are capable of watching a human perform a task and then figuring out how to mimic the motions. This allows apprentice robots to perform tasks that are difficult to program, such as folding towels. Each time the robot sees the towel, the towel will appear differently. So, how do programmers take all of these situations into account? They cannot, so they let the robot figure it out instead.
Apprentice Robot Learning Methods
Once again, no information was given regarding the algorithms used for apprentice robots. From examining how they work, we came to the conclusion that two methods they most likely use are supervised learning and learning to learn.
- Supervised Learning: In order to perform a task, the robots needed to be shown how to do it first. Thus, they are given the input, output, and all the general steps in between.
- Learning to learn: After seeing the general steps, the robot needs to figure out the precise details for the steps in order to effectively complete the task.
Future Examples of Machine Learning from IBM (International Business Machines) Research
All information and videos from
http://research.ibm.com/cognitive-computing/machine-learning-applications/index.shtml#fbid=ZpOKBgZY2xO
All information and videos from
http://research.ibm.com/cognitive-computing/machine-learning-applications/index.shtml#fbid=ZpOKBgZY2xO
A classroom that understands each student
Currently, teachers have a difficult time teaching in a style that suits every student. They are limited in how much time they can spend on a class. As a result, many students do not receive the personalized instruction they could use. IBM predicts that in 5 years, there will be classrooms that will learn about each student and how each student learns. They will be able to help teachers identify students who need extra help and then assist the students in making progress.
|
|
Classroom's Learning Methods
From watching the video regarding these classrooms, we determined they most likely will use semi-supervised learning and transductive algorithms.
- Semi-Supervised Learning: The classroom knows the input is how the student is performing in the class and that the desired output is for the student to succeed. The machine classroom then needs to figure out the steps the student needs to complete to reach this desired outcome.
- Transductive: The classroom will need to learn from its past experiences helping students to determine what works well and what needs to be changed.
"A Digital Guardian to Protect You Online" (IBM)
|
Every year, there are multiple cases of identity theft through the internet. As a result, IBM predicts that in 5 years, each person will have a digital guardian that learns your character and habits. If it feels something occurred that was not typical of you and could potentially lead to identity theft, it will notify you and take action to stop the person from continuing to use your identity.
|
Guardian's Learning Methods
We determined these digital guardians will use a form of supervised learning and transductive algorithms to accomplish protecting people.
- Supervised Learning: The guardians will know that the input is the human's normal activities, the steps are to check that nothing abnormal occurs, and the output is that the human is not a victim to identity theft. Thus, the guardians are given the input, output, and steps in between.
- Transductive: If something occurs that is atypical of the human, the guardian reports it. In order to reach this conclusion, the guardian needs to have learned from its past experiences that this particular action was not typical of the human.
What do these examples mean for our future and the future of Machine Learning?
Implications of HERB and Apprentice Robots
In the future, we will most likely have robotic assistants. With HERB and apprentice robots, we would no longer need to program a robot for a specific task. Instead, we can program the robot so it has the capability to learn from humans. With this capability, we can place robots in any environment with humans, and then the robot can assist with whatever tasks the human wants it to accomplish. As mentioned above, the creators of HERB hope to make him a truly autonomous robot, although they have not quite completed him at this time.
Implications of Classrooms learning about Students
There will be more individualized instruction that allows students to learn in the way that best suits them. Currently, many brilliant students often go unrecognized because they do not learn material effectively through lectures. With these future classrooms, our schools will become more effective because the classroom machine will find a way to teach the student that allows the student to effectively learn. The machine will then monitor the students' progress to make sure they keep succeeding.
Implications of Digital Guardians
With these guardians, identity theft will hopefully go down. We will not have to worry about someone using our credit card or identity to accomplish anything to the extreme because the guardian will monitor the actions and stop the theft from being completely accomplished. Whether these guardians actually exist in five years or not, the idea that they will eventually exist means our digital lives will become safer because of machine learning.