Who does the learning when it comes to Deep Learning Models
“The essence of the independent mind lies not in what it thinks, but in how it thinks.” ― Christopher Hitchens,
I can’t offer an expert opinion on the limitations of Machine Learning models. I’m new to this field and I hardly know anything about AI and Data Science. The level of my ignorance is only surpassed by my determination to obliterate said ignorance.
Still, I’m writing this post because you might find some value in a beginner’s point of view.
To me, the main limitation of Machine Learning is that there is no learning. There’s only super fast calculating.
AI is a huge network of mini-calculators (the nodes), which pass their results forward and backward in a way that enables the network to compute extremely fast the best combination of parameters for a function that needs to ignore outliers and exceptions.
The process can be summed up in two steps:
- AI takes in huge amounts of data that was processed by humans (a lot of processing in the case of structured data, less processing in the case of unstructured data);
- Using different algorithms and hyper-parameters like accuracy level, AI goes through billions and billions of combinations of weights for the different features of the data, to find a function that uses those weights to calculate a most likely output.
AI however doesn’t (and can’t) do a lot of things that true learning involves. To keep this post within the recommended 300 word limit, I’ll only mention two:
- AI can’t evaluate the quality of its data. If I decide to label cat pictures as mice, we’ll probably hear in the news that pictures of mice are the most shared animal pictures on the web. 😃
- AI can’t put anything in context (not the inputs, not the outputs, not the algorithms). This can lead to big problems like the armageddon some worry about, or some smaller ones like perpetuating bias.
Having just mentioned some of the bad things about AI, I have to say that, overall, I’m fascinated by its potential. At this point, I believe it’s best to have AI than not to have it. It is an impressive tool that can help us find patterns and relationships other tools can’t (because they are too slow.)
All the learning, however, is done by humans. 😃
By: Bianca Aguglia (theresarttodatascience.com)