\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{amsmath,amssymb}
\usepackage{enumitem}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[hidelinks]{hyperref}
\hypersetup{pdftitle={CLIP: Oral Knowledge Questions},pdfsubject={Architecture, contrastive training, zero-shot classification, and limitations}}
\setlength{\parindent}{0pt}
\setlength{\emergencystretch}{2em}
\setlist[enumerate]{leftmargin=*,itemsep=1.2em,topsep=1em}

\begin{document}
\raggedright
{\LARGE\bfseries CLIP: Oral Knowledge Questions\par}
\vspace{0.6em}
Explain each answer in your own words and justify the underlying mechanism.

\begin{enumerate}
\item \textbf{CLIP's two encoders and shared space.}

Describe how CLIP converts an image and a caption into comparable vectors.
What are the roles of the image encoder, text encoder, linear projections,
and L2 normalization? Do the image and text interact through cross-attention
during encoding?

\newpage
\item \textbf{Why CLIP's text encoder does not generate text.}

CLIP's text encoder uses causal Transformer blocks, which are also used in
autoregressive language models. Why does this not make CLIP a caption
generator? Explain what CLIP extracts from the end-of-sequence position,
how that output is trained, and how it is used.

\newpage
\item \textbf{Positive pairs, negatives, and collapse.}

In a batch of matched images and captions, what counts as a positive pair
and what counts as a negative pair? Why could simply making matched
embeddings similar lead to every input receiving the same vector? Explain
why that representation would perform poorly on CLIP's contrastive
matching task.

\newpage
\item \textbf{Zero-shot classification versus training a classifier.}

How can CLIP classify images into a new set of categories using only the
category names and text prompts? Contrast this with training a linear
classifier on frozen image features. Which parameters change in each
approach, and why can prompt wording affect zero-shot predictions?

\newpage
\item \textbf{Recognizing objects versus understanding relationships.}

Consider the captions ``a dog under a table'' and ``a dog on a table.''
Why might CLIP recognize the objects yet struggle to distinguish these
captions? Explain how training with captions that contain the same objects
but incorrect relationships could encourage better relational understanding.
\end{enumerate}
\end{document}