Iho chatbot, or how I was introduced to Machine Learning
Hello!
Did you ever get an idea that is sooo cool yet sooo ambitious?
Somehow, all my ideas end up like that.
But resisting the hard task is often as hard as the task itself!
That is a tale of how Iho turned from just a cute girl on the screen of your weather app to a multi-day research.
But first..
What is Machine Learning? 🤖
For anyone who is not in the field of computer science, it may be wayy too confusing if I go too much into details. Heck, I was also confused the first time I tried to read about it!
As to make things simpler.. Machine Learning is a bunch of algorithms and math. Most closely it is related to Statistics; they are not the same, obviously, and as I found out, it has been causing some disputes within the communities about whether they should be synonymized or not (the answer is no (i think)).
But anyway, we do calculations to find the closest fitting group for the new item out of the groups that we used for training of the model. And that is what Machine Learning is.
Yeaaah I’m sorry for that explanation 😵💫
It might have not been very good or detailed, but I tried to make it as simple as possible for the majority to understand, since I find that a lot of current resources waaaay overcomplicate things. Not every topic has to be extremely hard to grasp, sometimes you just want to get the basic idea without the marketing-oriented response! (she says as her response is likely as unexplaining).
But anyway, I got sidetracked.
Going back to our Iho, she was originally planned to be a simple image, maybe with a few interactions. But that is all.
After some thought though, making her just.. be there was not that good of an idea. Why is she there? Well, obviously, she’s introduced as some sort of helper, but what kind of helper does not help at all?
That is when the idea of a chat popped up. Since then, I knew I could not make a simple weather bot, I had to make her more diverse, able to at least react to what we say.
And to do that, I had to turn to Machine Learning.
With Machine Learning, you have a program that is able to take in the input, predict how it might fit into the existing data, and give an answer based on that.
I did not want to turn to LLMs. Training one yourself takes a lot of time, data, resources, and it would have made the program much heavier than it should be. You could base the model on an existing one, but then you would have less control of it, which was also completely not ideal.
That is why I decided to go for a simpler type, but train it myself.
As I said, after some research, it turned out to be math, math, more math, and it all was put together in some way.. Luckily, there are already libraries that help you handle the process.
What I still had to do though, was prepare training data.
I wanted the model to have two modes: professional, all about weather, and casual. That is why I had to prepare two sets of data, mostly not to pollute the professional set with too much casual info.
That is when the problems started.
Sometimes the model would not understand the topic correctly, sometimes it would give more info about weather, sometimes — more about casual things. At one point, she was fixed on talking about hobbies, when I didn’t; next moment, all her responses are solely about weather, even ones that worked well before.
Unfortunately, for anyone who is waiting for the moral or solution.. there is none, I have to solve it yet 😅
Stay tuned for when I post the updates on the topic!
Leave a Reply