#dual-encoder

Question 1: Single choice

A classifier computes the logits $g(x) = W^\top [g_{\text{feat}}(x); 1]$ for $C$ classes and is trained with cross-entropy. Read it as dual-encoder (CLIP-style) contrastive learning between inputs and class labels, with input encoder $f_1(x) = [g_{\text{feat}}(x); 1]$ and logits $f_1(x)^\top f_2(c)$ . What is the label encoder $f_2(c)$ ?

The $c$ -th column of $W$ , a learned embedding of class $c$ 1
The one-hot vector of class $c$ 2
The average feature vector of the training inputs of class $c$ 3
The softmax probability that $x$ belongs to class $c$ 4

Timed test