End-of-lesson Quiz
5 questions · First Neural Network in Keras
1
of 5
What does a single neuron actually compute?
A neuron is just
activation(w⋅x + b) — multiply each input by a weight, add them up, add a bias, then squash through an activation. That's it. A whole network is just thousands of these stacked together.
2
of 5
Why does ReLU sometimes 'die' during training, getting stuck outputting 0 forever?
Dying ReLU: ReLU is
max(0, x). If x is always negative, the output is always 0 and the gradient is 0. With no gradient, the weights can't update, so the neuron stays dead permanently. Solutions: better initialisation or LeakyReLU (small slope for negatives).
3
of 5
Your network has 44,000 parameters but only 2,000 training samples. What's most likely to happen?
Rule of thumb: keep parameters well below ~10x your training samples. With 22 parameters per sample, the network can essentially memorise every example. Either use a smaller architecture or get more data.
4
of 5
During training, the loss curve sometimes goes up for a few batches. Is this a problem?
Stochastic gradient descent samples random batches, so individual updates can briefly increase loss. The downward trend over many batches is what matters. If the trend itself is upward, the learning rate is too high.
5
of 5
Your neural network and a logistic regression baseline both have AUC ~0.92 on the same security task. Which should you ship?
If a simpler model matches the deeper one, ship the simpler one. It's faster, more interpretable, easier to maintain, and easier to debug when something goes wrong in production. Only reach for neural networks when they clearly outperform the baseline.