The complexities and limitations of neural networks

As we’ve seen in previous sections, ANNs are extremely powerful and extensively used in different AI applications. Notwithstanding this, they are very complex machines (some of them are reaching the one trillion parameter mark), and because of this, they suffer from various limitations. The following are some of the most critical issues.

The first issue with ANNs is their “black box” nature. When we refer to an algorithm, we mean that the algorithm can get a set of inputs and return a set of outputs, but we are not entirely sure about how it achieved those results since the inner workings are not visible. It’s like looking at a black box machine, hence the name! So, if we create a neural network to distinguish between a cat and a dog, we cannot say how it managed to do it, whether it’s the color of the animal, the pointy ears, or its long tail.

Human-interpretable features are much more preferred...