papers/hanzo-jin/sections/05-experiments.tex

\section{Experiments}


\subsection{Zero-Shot Image Classification}

\begin{table}[h]
\centering
\caption{ImageNet zero-shot top-1 accuracy (\%).}
\label{tab:imagenet}
\begin{tabular}{lc}
\toprule
\textbf{Model} & \textbf{Accuracy} \\
\midrule
CLIP ViT-L/14 & 75.3 \\
ALIGN & 76.4 \\
Florence & 83.7 \\
\midrule
Jin-Base & 82.1 \\
Jin-Large & \textbf{85.2} \\
\bottomrule
\end{tabular}
\end{table}

\subsection{Audio-Text Retrieval}

\begin{table}[h]
\centering
\caption{AudioCaps retrieval recall@10.}
\label{tab:audiocaps}
\begin{tabular}{lcc}
\toprule
\textbf{Model} & \textbf{T$\rightarrow$A} & \textbf{A$\rightarrow$T} \\
\midrule
AudioCLIP & 45.2 & 43.8 \\
CLAP & 51.3 & 49.7 \\
\midrule
Jin-Base & 58.9 & 56.4 \\
Jin-Large & \textbf{63.2} & \textbf{61.8} \\
\bottomrule
\end{tabular}
\end{table}

\subsection{Cross-Modal Generation}

\begin{table}[h]
\centering
\caption{Cross-modal generation quality.}
\label{tab:generation}
\begin{tabular}{lcc}
\toprule
\textbf{Task} & \textbf{FID/FAD} & \textbf{CLIP-Score} \\
\midrule
Text $\rightarrow$ Image & 8.4 & 0.31 \\
Audio $\rightarrow$ Image & 12.1 & 0.27 \\
Image $\rightarrow$ Audio & 2.1 (FAD) & --- \\
Text $\rightarrow$ Audio & 1.8 (FAD) & --- \\
\bottomrule
\end{tabular}
\end{table}

\subsection{z-JEPA Ablation}

\begin{table}[h]
\centering
\caption{z-JEPA ablation on ImageNet linear probe.}
\label{tab:zjepa_ablation}
\begin{tabular}{lc}
\toprule
\textbf{Configuration} & \textbf{Accuracy} \\
\midrule
Single-scale JEPA (baseline) & 77.9 \\
+ Hierarchical levels & 79.4 \\
+ Cross-scale prediction & 80.8 \\
+ Adaptive zoom & 81.6 \\
+ Bidirectional feedback & \textbf{82.3} \\
\bottomrule
\end{tabular}
\end{table}

\subsection{Computational Efficiency}

\begin{table}[h]
\centering
\caption{Inference throughput (samples/sec) on A100.}
\label{tab:throughput}
\begin{tabular}{lccc}
\toprule
\textbf{Model} & \textbf{Text} & \textbf{Vision} & \textbf{Multi} \\
\midrule
Jin-Base (dense) & 890 & 234 & 156 \\
Jin-Base (MoE) & 1,240 & 312 & 223 \\
Speedup & 1.39x & 1.33x & 1.43x \\
\bottomrule
\end{tabular}
\end{table}

MoE routing provides 1.3-1.4x throughput improvement at equivalent quality.