jin/papers/jin_hypermodal_fair.aux

\relax 
\providecommand\hyper@newdestlabel[2]{}
\providecommand\HyField@AuxAddToFields[1]{}
\providecommand\HyField@AuxAddToCoFields[2]{}
\citation{assran2023self,bardes2024v}
\@writefile{toc}{\contentsline {section}{\numberline {1}Introduction}{1}{section.1}\protected@file@percent }
\@writefile{lof}{\contentsline {figure}{\numberline {1}{\ignorespaces JIN architecture overview. Multiple modalities are processed through hierarchical levels with bidirectional connections, unified in an energy-based embedding space.}}{1}{figure.1}\protected@file@percent }
\newlabel{fig:overview}{{1}{1}{JIN architecture overview. Multiple modalities are processed through hierarchical levels with bidirectional connections, unified in an energy-based embedding space}{figure.1}{}}
\citation{chen2020simple}
\citation{he2020momentum}
\citation{assran2023self}
\citation{bardes2024v}
\citation{radford2021learning}
\citation{jaegle2021perceiver}
\citation{lecun2006tutorial}
\@writefile{toc}{\contentsline {section}{\numberline {2}Related Work}{2}{section.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {2.1}Self-Supervised Visual Learning}{2}{subsection.2.1}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {2.2}Multimodal Learning}{2}{subsection.2.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {2.3}Energy-Based Models}{2}{subsection.2.3}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {3}Method}{2}{section.3}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {3.1}Hierarchical z-JEPA Architecture}{2}{subsection.3.1}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {3.1.1}Cross-Scale Predictive Objectives}{2}{subsubsection.3.1.1}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {3.1.2}Adaptive Computation}{2}{subsubsection.3.1.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {3.2}Hypermodal Embedding Space}{2}{subsection.3.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {3.2.1}Modality-Specific Encoders}{2}{subsubsection.3.2.1}\protected@file@percent }
\@writefile{loa}{\contentsline {algorithm}{\numberline {1}{\ignorespaces JIN Training}}{3}{algorithm.1}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {3.2.2}Energy-Based Alignment}{3}{subsubsection.3.2.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {3.3}Spatially-Aware Attention}{3}{subsection.3.3}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {3.4}Training Procedure}{3}{subsection.3.4}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {4}Experiments}{3}{section.4}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {4.1}Experimental Setup}{3}{subsection.4.1}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {4.1.1}Datasets}{3}{subsubsection.4.1.1}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {4.1.2}Implementation Details}{3}{subsubsection.4.1.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {4.2}Main Results}{3}{subsection.4.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {4.2.1}ImageNet Classification}{3}{subsubsection.4.2.1}\protected@file@percent }
\@writefile{lot}{\contentsline {table}{\numberline {1}{\ignorespaces ImageNet-1K results. JIN achieves superior accuracy with fewer FLOPs.}}{3}{table.1}\protected@file@percent }
\newlabel{tab:imagenet}{{1}{3}{ImageNet-1K results. JIN achieves superior accuracy with fewer FLOPs}{table.1}{}}
\@writefile{toc}{\contentsline {subsubsection}{\numberline {4.2.2}Multimodal Understanding}{3}{subsubsection.4.2.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {4.3}Ablation Studies}{3}{subsection.4.3}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {4.3.1}Component Analysis}{3}{subsubsection.4.3.1}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {4.3.2}Scaling Properties}{3}{subsubsection.4.3.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {4.4}Qualitative Analysis}{3}{subsection.4.4}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {4.4.1}Cross-Modal Retrieval}{3}{subsubsection.4.4.1}\protected@file@percent }
\@writefile{lot}{\contentsline {table}{\numberline {2}{\ignorespaces Multimodal benchmark results across vision-language (VQA), audio (AudioSet), and 3D understanding (ScanNet).}}{4}{table.2}\protected@file@percent }
\newlabel{tab:multimodal}{{2}{4}{Multimodal benchmark results across vision-language (VQA), audio (AudioSet), and 3D understanding (ScanNet)}{table.2}{}}
\@writefile{lot}{\contentsline {table}{\numberline {3}{\ignorespaces Ablation study showing the contribution of each component.}}{4}{table.3}\protected@file@percent }
\newlabel{tab:ablation}{{3}{4}{Ablation study showing the contribution of each component}{table.3}{}}
\@writefile{toc}{\contentsline {subsubsection}{\numberline {4.4.2}Attention Visualization}{4}{subsubsection.4.4.2}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {5}Analysis}{4}{section.5}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {5.1}Computational Efficiency}{4}{subsection.5.1}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {5.2}Emergent Properties}{4}{subsection.5.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsubsection}{\numberline {5.2.1}Scale Invariance}{4}{subsubsection.5.2.1}\protected@file@percent }
\@writefile{lof}{\contentsline {figure}{\numberline {2}{\ignorespaces Scaling behavior of JIN compared to ViT. JIN shows better scaling efficiency.}}{4}{figure.2}\protected@file@percent }
\newlabel{fig:scaling}{{2}{4}{Scaling behavior of JIN compared to ViT. JIN shows better scaling efficiency}{figure.2}{}}
\@writefile{lof}{\contentsline {figure}{\numberline {3}{\ignorespaces Cross-modal retrieval examples. Top: text→image, Middle: audio→video, Bottom: image→3D.}}{4}{figure.3}\protected@file@percent }
\newlabel{fig:retrieval}{{3}{4}{Cross-modal retrieval examples. Top: text→image, Middle: audio→video, Bottom: image→3D}{figure.3}{}}
\@writefile{toc}{\contentsline {subsubsection}{\numberline {5.2.2}Modality Dropout}{4}{subsubsection.5.2.2}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {5.3}Energy Landscape Analysis}{4}{subsection.5.3}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {6}Discussion}{4}{section.6}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {6.1}Theoretical Insights}{4}{subsection.6.1}\protected@file@percent }
\bibstyle{plain}
\bibdata{references}
\bibcite{assran2023self}{{1}{}{{}}{{}}}
\bibcite{bardes2024v}{{2}{}{{}}{{}}}
\bibcite{chen2020simple}{{3}{}{{}}{{}}}
\bibcite{he2020momentum}{{4}{}{{}}{{}}}
\@writefile{lof}{\contentsline {figure}{\numberline {4}{\ignorespaces Hierarchical attention maps showing focus at different scales. Higher levels attend to larger semantic regions.}}{5}{figure.4}\protected@file@percent }
\newlabel{fig:attention}{{4}{5}{Hierarchical attention maps showing focus at different scales. Higher levels attend to larger semantic regions}{figure.4}{}}
\@writefile{lot}{\contentsline {table}{\numberline {4}{\ignorespaces Robustness to scale variations on ImageNet.}}{5}{table.4}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {6.2}Limitations and Future Work}{5}{subsection.6.2}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {7}Conclusion}{5}{section.7}\protected@file@percent }
\@writefile{lof}{\contentsline {figure}{\numberline {5}{\ignorespaces Performance degradation with modality dropout. JIN maintains reasonable performance even with single modality.}}{5}{figure.5}\protected@file@percent }
\bibcite{radford2021learning}{{5}{}{{}}{{}}}
\providecommand\NAT@force@numbers{}\NAT@force@numbers
\@writefile{toc}{\contentsline {section}{\numberline {A}Implementation Details}{6}{appendix.A}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {A.1}Architecture Specifications}{6}{subsection.A.1}\protected@file@percent }
\@writefile{lot}{\contentsline {table}{\numberline {5}{\ignorespaces Model configurations for different JIN variants.}}{6}{table.5}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {A.2}Training Hyperparameters}{6}{subsection.A.2}\protected@file@percent }
\@writefile{lot}{\contentsline {table}{\numberline {6}{\ignorespaces Training hyperparameters for JIN.}}{6}{table.6}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {A.3}Data Preprocessing}{6}{subsection.A.3}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {B}Additional Results}{6}{appendix.B}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {B.1}Transfer Learning}{6}{subsection.B.1}\protected@file@percent }
\@writefile{lot}{\contentsline {table}{\numberline {7}{\ignorespaces Transfer learning results with linear probing.}}{6}{table.7}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {B.2}Zero-Shot Evaluation}{6}{subsection.B.2}\protected@file@percent }
\@writefile{lot}{\contentsline {table}{\numberline {8}{\ignorespaces Zero-shot evaluation without task-specific fine-tuning.}}{6}{table.8}\protected@file@percent }
\gdef \@abspage@last{6}