GPT takes as input a sequence of the first tokens, and its corresponding output does not have anything to do with the conditional probability distribution for token , .

Match the Python libraries with their intended application.

Which statement about positional encodings/embeddings in Transformers is correct?

Let us assume that L is a list in Python, and the following line is executed:
L[1:2] = ["d", "d"]
The list L will change its length after the line has been executed.

Match the properties of the AI environments with their characteristics.

You are going to solve the NIM game (studied in our classes) using reinforcement learning. Match the components of a Markov decision process with the elements of the NIM game that would lead to a sensible implementation.

Imagine that the following simplified equation is used to implement self-attention:

,

where the symbols Q, K and V have the same meaning as in our lectures.

The question is about learnable parameters. Select one option that is correct.

Imagine that you have a reinforcement learning policy obtained using Q-learning, and your policy is optimal for the NIM game. You execute this policy with the -greedy exploration where . Would this execution lead to the selection of incorrect actions by the algorithm in some situations? That is, would the policy suggest "irrational" actions in some states?

In a standard set-up, the Transformer takes as input a matrix of word embeddings and returns a matrix of the same size as its output.