What are the limitations of ML?
Welcome back to the second episode of “Towards Machine Learning” and we are going to discuss the limitations of Machine Learning and some stories around them.
For all its resourcefulness even ML, in its approach, has some limitations too. It is not a silver bullet that provides answers to all the business problems. Let’s go back to the example of Google’s self-driving car Waymo that we discussed in our first session.
Stop!!
Suppose you’re the lead engineer of Google’s self-driving car project. While teaching the car how to drive you’ve never come across a stop sign. This means the machine never learned about the stop sign. The car never learned what it should do when it encounters a stop sign. So, the car when deployed in real life does not “see” the stop sign and will not obey the traffic rule. It will run over the stop sign. Here’s a video of a similar situation. So, one of the limitations of ML is that your machine will not perform well in real life scenarios if the machine did not come across such scenarios in the past (training).
Lack of data…
Another limitation of ML is the amount of data required to train a machine learning algorithm. Complex machine learning algorithms such as neural networks need a lot of data to train Sometimes finding huge data is not easy. Machine learning algorithms learn the same way kids do, kids need a lot of practice (analogous to training an ML algorithm) to distinguish the faces of people. So, in a nutshell, ML algorithms need to see a lot of examples to work really well in real life scenarios.
One shot learning
Machine Learning algorithms need a lot of data to do well in real life examples. As of now, ML algorithms are not fully capable of learning in one shot or with fewer examples. Imagine the amount of time that can be saved when training a software program with fewer examples.
This is a very novel research area in Machine Learning/Deep Learning. It works as follows; if I show the computer a picture of a car, the machine should be able to recognize any other cars. This has a lot of applications where collecting training examples is a tedious job or when there are only limited training examples such as drug testing.
These are some of the limitations of Machine Learning. Apart from these, Machine Learning is a time-consuming process and requires lots of expertise to design and run the algorithms. Having said that they are also quite fun to learn. Now, that you have the knowledge of limitations of Machine learning, you will be in a better position to judge the use cases of ML.
Hope you enjoyed reading this issue and coming up in the next issue, what is training and testing of a Machine Learning algorithm?
In case you missed our first issues, it’s here, Towards Machine Learning: Episode-1.
If you love to learn more here is our third episode- How to become an analyst in 3 months?
Ciao for now!
