Machine Learning is already a common say at the end of 2018. But though there are still lots of ground to cover moving forward, looking back there are some basics to remind ourselves of.
This article is an improvement on how to approach machine learning problemsand would answer some questions that were left unattended to.
Let’s get right into it
You still remember that Machine Learning is trying to build intelligence that cannot be coded line by line into a computer system. Whether it is recognizing image, playing a game or composing a new song.
But before machine learning path is taken, I believe we have some questions to answer both the board and yourself:
One, do I have the right data to evaluate a machine learning algorithm for the problem I wanna solve?
Yeah, there are a lot of data out there, much more being generated by the minute….blah blah blah.
We get it, there is a lot of data.
But what are you solving. Are you trying to predict the pattern of rainfall in the suburbs of somewhere where no rainfall data has been taken in years? Are you trying to computerized the old letters in your museum and though your cause is noble, the red tape is titanium-thick?
So, before you ever embark on the machine learning -path, think, do I have the data I need?
Two, after I have implemented my solution, what is the business impact?
If you company turns over a tens of thousands yearly, and your algorithm – after you have consumed more resources (time and money) – would save the same company a few hundred in that same space of time. Please Don’t!
Cos at the end, your job is at stake and you’ve got an angry board
Three, before taking the path, ask yourself if you would be carefully considering the human in the loop
For a machine learning problem that wants to distinguish between a cat and a dog, mistaking one for the other is a trivial mistake.
But when a machine learning algorithm mistakes a human for a tree in a self-driving car algorithm or cannot recognize the human, then we have a big problem on our hands.
What we are saying…
When humans are in the loop our model accuracy must be closer to 100% to save lives or we would cost more than before the machine learning algorithm.
One last question, when our model is our there in the world how do we test and improve it?
Before taking your model to the real world, the machine learning engineer has to determine if the model has stopped learning (Batch learning) or would still continue to learn as users interact with the system (Online learning).
When one of the two has been chosen, the machine learning engineer must then factor it into the model before launching it for the world to see.
At this point also, the machine learning engineer utilizes his lessons in software engineering – testing, quality assurance etc.
The cloud is basically the place where all machine learning models finally “go to rest” and there are a lots of options – Amazon Web Services, Google Cloud, Azure, Paperspace etc.
The journey of a machine learning engineer starts from trying to solve a problem but does not end when it is safely in the cloud like many software applications.
The journey still continues. The machine learning engineer has to go over a lot to ensure that the problem is being solved using the best possible combination and that intelligent is actually built.
If the system is truly intelligent, we applaud. If not, we bemoan
And that’s the plight of the machine learning engineer.
Though there are lots of questions to be answer before embarking on the ml path, I just believe these ones would set thing straight before complications arises
Thank you for reading