Sramana Mitra: I’m going to fight you again on this one. That’s the spirit in which I like to have these conversations, so I’m glad you’re enjoying it.
The two examples that you picked are both higher-order problems than your domain. Legal case material is far more complex. The range of vocabulary, the range of circumstances, and the range of natural language that it has to process and make sense of is far more complex.
News is far more complex than math. Math is a little domain. The only issue in everything that you said that is potentially the cause of why that justifies your approach is that the dataset isn’t quite there yet. Once you have a much larger data set, your algorithm should be able to do this on its own.
Raj Valli: I’m going to push back on your thinking a few more times. Life doesn’t fall into a beautiful Gaussian curve but a typical normal probability. Life has got a lot of cute curves all the time. That’s why you try to normalize the data before you can use the data.
There are reasons why we transform data. Otherwise they become non-usable because it doesn’t fit well into what I would call a usable curve. If I have a right skew or a left skew or any skew for that matter for a typical normal probability distribution, I’m going to have challenges interpreting and making recommendations based on the dataset.
Suppose I have a distribution of 10,000 students and I make my dataset based on the behavior of those 10,000 students. Tomorrow, for example, if you give me a hundred students who have the highest IQ, I would reject the data in the way we are building the systems because they are extremes that are going to pollute what I would call normal behaviors of the student population.
Sramana Mitra: Yes, but in your data structure you would set them up as different clusters. You would set them up to different aptitude levels of different clusters.
Raj Valli: Of course. The point I’m trying to make is that the possibility of turning a machine on and guaranteeing that a student can make learning outcomes happen is not something I can bet on.
Sramana Mitra: That’s exactly what I said. You’re not ready yet because you don’t have enough dataset and you don’t have your algorithm to be built. That’s fair.
Raj Valli: That is not true. It’s the equivalent of saying that all my customer service calls will be taken over by an AI machine. I wouldn’t do it not because the dataset is not there, it’s because I don’t want any of my customers to feel offended in 5% of the cases.
Sramana Mitra: Those are two different issues.
Raj Valli: It’s not. At what point are you going to be comfortable in having confidence in your dataset? That’s the question I’m asking you.
Sramana Mitra: The white glove approach is a completely different thing. What are you trying to address here? Are you trying to address the fact that there’s a human teacher who sends this message to a student that makes the student feel that there is a person who is taking care of him and the parents feeling that? Is that the issue?
Or are we talking about generating the exercise? Are the tutors personally involved in generating the exercise or in delivering the exercises? If delivering the exercises is the issue and if the actual generation portion is happening automatically, those are two completely different issues.
Raj Valli: Yes, so generation is not an issue; delivery is.
Sramana Mitra: Delivery is not AI, delivery is sending just a email.
Raj Valli: When I say the generation of the worksheet, the recommendation says that these are the 10 worksheets that need to be assigned to a particular student. That is a generation of that worksheet whether those 10 worksheets are going to be delivered to the student or not, whether it’s going to be passing through the tutor, and whether they’re going to be accepting or rejecting all of those 10 recommendations.
That’s where I am not comfortable having the machine completely go through this process completely.
This segment is part 5 in the series : Thought Leaders in Artificial Intelligence: Raj Valli, CEO of Thinkster
1 2 3 4 5 6