My Birthday cake was detected as a coin!
Or common misconceptions about AI
There was a certain occasion that happened to me on my last birthday. It made me realize that anything that gets hyped rapidly, brings a lot of misconceptions with it.
On my birthday, before gathering in the main room of the office, I left my cakes on the table and left the room. Suddenly the whole office got filled with a burst of laughter. People started to tell me:
“Anna, your model says your cake is a coin.”
“Anna, tell your model that this is a cake and not a coin because it feels soft and has a specific structure.”
I thought it would be helpful to share the common misconceptions people have about the mysterious notion of “Artificial Intelligence”, so there will be less confusion.
- We do not tell our models what exactly to pay attention to. The whole beauty of the concept of Machine Learning is that the machine learns by itself. If a data scientist wants his model to learn better, it feeds it with quality data and sets it free to learn whatever it can. We do not tell our models to pay attention to the structure/color/font/curves/shape. This means, telling a data scientist to tell his model that the cake is soft therefore it’s not a coin, is not gonna work.
- The model learns only what a data scientist fed it with. I remembered the one case that happened 2 years ago. The model stopped working in the middle of the demo. The reason was the new data that potential customers gave us had a whole new set of features we(and our model) never saw before. Every time we remove or add some feature from the data, a model needs to be retrained. If a model is trained on 21 features to predict, for example, churn rate, it will need exactly those 21 features to make predictions even after 3 years.
- If a data scientist got, for instance, a good brand recognition model, it does only mean that he got a good brand recognition model. He can’t give you food/drink/emotion or loyal spouse recognition model instantly. He will need nearly as much time as for his previous project to complete another task in “something” recognition. The whole beauty of AI is that it doesn’t matter how cool of a professional you are, you never get to the point where you get a magic model that could detect and recognise anything.
We are as far from Artificial General Intelligence as we are from being called superhumans.
If one neural network architecture worked for one case it might equally fail for another task.
- We do not get a model that detects something, and then understand what we can do with it. The whole issue here is that we are missing out on the crucial part of any successful AI-based product: Well defined use case.
Data scientist’s objective is to train a model that can generalize to the unseen data that it will eventually encounter. This means that the data the model is trained on is supposed to be an unbiased representation of your use case. No machine learning, no data scientist can compensate for bad data.
As one very subtly humorous being once said: “Artificial Intelligence is like teenage sex: everyone talks about it, nobody knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it.”
Artificial intelligence has gone far for these last years, but the huge potential it has for the future makes my hands shake and be grateful that I live in this era. To get the fullest from this field it’s equally important that people have a realistic view of how it works.
