Neural Network Trivia Questions

Text and images are converted into numerical representations (word embeddings, pixel values) before being fed into networks. They can handle any data type with proper encoding.

‘Deep’ refers to multiple hidden layers. A network with zero hidden layers is just a linear model, not deep at all—depth requires at least one hidden layer.

Early neural networks, like the perceptron, were modeled after biological neurons. However, modern deep networks are far simpler and less realistic than actual brain wiring.

Generative Adversarial Networks pit a generator (creating fake data) against a discriminator (detecting fakes). This adversarial training leads to highly realistic outputs.

A single-layer perceptron can only solve linearly separable problems. XOR requires a hidden layer—this was famously proven by Minsky and Papert in 1969.

Dropout works by temporarily ‘dropping out’ a random subset of neurons each training step, forcing the network to learn more robust features. It’s a regularization method.

Unsupervised learning methods, like autoencoders or self-supervised learning, train networks without labels. They find patterns in data itself.

Sigmoid and tanh activations can cause gradients to shrink exponentially in deep networks, making early layers learn slowly. ReLU and skip connections help fix this.