I'll give you a practical example. So, we, in our identity verification: let's take document fraud detection. Again, we have customers from all over the world, sending us images taken on every kind of mobile device with every kind of sensor in every kind of lighting conditions and focus drifts and focus conditions.
And so as our data corpus builds, it is much more effective to ask a neural network to deal with that noise because you don't have to build symbolic rules to deal with that noise. You let the neural network learn the spread of the noise, the noise distributions. And so it can naturally deal with that.
So here's an example where something is hard to encode and as long as you get a good data corpus, let the neural networks do an amazing job. Another example would be, it's a data dependent thing. If you have a lot of data, neural networks are amazing, and we've seen this time and time and time again.
So if you've got lots of data, don't worry about a lot of symbolic stuff, right. Build in priors so that you get generalization and explainability to some extent, but use the neural networks powerfully. So it's almost like there's no one rule, right? You have to take your problem and know whether you want to go to the left bookend with neural networks or to the right bookend with symbolic rules.
And where do you sit in that depends on the problem you're solving.