Where to start

Deep Learning & Neural Networks

A short check to suggest a starting chapter. It's advice only: your answers aren't saved, nothing gets unlocked, and it doesn't count as evidence of a skill. It hasn't been tested on learners yet, and the questions are in English.

What does reverse-mode differentiation compute efficiently?

The gradient of one output with respect to many inputs

Only second derivatives

Derivatives of functions with no computational graph

Training loss jumps around and then becomes NaN. What's the most likely first thing to try?

Add more layers

Lower the learning rate

Remove the validation set

How should dropout behave at evaluation time?

Drop the same units as in training

Drop twice as many units

It's turned off, so that predictions are deterministic

In a PyTorch training loop, what happens if you never call optimizer.zero_grad()?

Gradients accumulate across batches

The weights are reset at each step

PyTorch raises an error

A 3×3 convolution with no padding is applied to a 32×32 image. What's the output's height and width?

32×32

30×30

29×29

In scaled dot-product attention, why divide by √d?

To keep large dot products from pushing softmax into regions with tiny gradients

To make the matrix square

To remove the need for a mask

Understand

Learn by doing. Understand why.

Learn

Explore the programmes Higher education High school Programming Data Science & AI Search

Understand

The method Create an account Log in

Help and privacy

Privacy Contact