|
Zhaoze Wang I am a Ph.D. student at the UPenn GRASP Lab. I study predictive world models and compact latent representations for multimodal sequence learning, long-horizon prediction, and planning, drawing inspiration from biological memory and navigation. I previously interned at Adobe Research, where I worked on video world models, with a focus on efficient generation and long-horizon prediction. I am seeking research internships from Summer through Fall 2027, with availability for a continuous internship across both seasons. Email / Google Scholar / GitHub / LinkedIn / X |
|
Research InterestsThe brain's ability to build internal predictive models is a key inspiration behind world models. My research asks how such representations can support not just prediction, but fast updating and replanning as observations and goals change. I study compact predictive representations, recurrent memory, and learned dynamics, drawing on how the brain integrates sensory experience to guide action. My goal is to translate these principles into fast, adaptive robotic systems. I study how sensory sequences shape spatial representations and how structured predictive states emerge. I also work on memory-guided planning and efficient video world models. To support model training, I develop compact visual encoders and parallel sensory simulation. |
Selected Publications |
|
|
A Simple Model of Co-Emergence of Grid and Place Fields
arXiv preprint, 2026 Recurrent sensory-prediction models learn from masked multimodal histories and egocentric actions. Across 5,000+ architecture, masking, and noise runs, we study how structured spatial representations emerge, connecting biological navigation to learned predictive states. |
|
REMI: Reconstructing Episodic Memory During Internally Driven Path Planning
NeurIPS 2025 Zhaoze Wang, Genela Morris†, Dori Derdikman†, Pratik Chaudhari†, Vijay Balasubramanian† Paper / Project Page / Simulation Suite / Vision Encoder A neuroscience-inspired recurrent world model that binds visual observations to latent spatial states, retrieves goals from partial cues, and rolls out long-range plans in Habitat-Sim. Learned local transition dynamics generalize to shortcuts through unseen states and cue-conditioned planning. ViT-B/L masked autoencoders provide compact visual embeddings, with multi-GPU DDP/AMP rollout and evaluation pipelines. |
|
Time Makes Space: Emergence of Place Fields in Networks Encoding Temporally Continuous Sensory Experiences
NeurIPS 2024 Zhaoze Wang†, Ronald W. Di Tullio†, Spencer Rooke, Vijay Balasubramanian Paper / Code / Video / Project Page Inspired by biological spatial memory, we show that temporal continuity enables recurrent models to learn stable environment-specific latent representations and pattern-complete noisy or occluded sensory trajectories. See also Trading Place for Space. |
|
Trading Place for Space: Increasing Location Resolution Reduces Contextual Capacity in Hippocampal Codes
NeurIPS 2024 Oral Spencer Rooke, Zhaoze Wang, Ronald W. Di Tullio, Vijay Balasubramanian Paper / Video Starting from biological spatial codes, we derive geometric bounds linking spatial resolution, contextual capacity, and representation dimensionality, characterizing trade-offs in compact spatial state representations. See also Time Makes Space. |
Open Source |
|
|
BtnkMAE: Compact and Decodable Visual Representations
Source Code A bottleneck MAE that reduces each ViT frame from a patch-token sequence to a single compact embedding, then reconstructs images with learned patch queries and cross-attention. Released ViT-B/L ImageNet-1K checkpoints with configurable DDP/Hydra training and support for variable-resolution and panorama inputs. |
|
RatatouGym: GPU-Parallel Multimodal Navigation Simulation
Documentation / Source Code A PyTorch-native simulator that generates model-ready spatial and movement sensory streams from thousands of concurrent navigation trajectories, tested at 4,096 parallel environments. GPU-resident trajectory stepping and sensory encoding use batched tensor operations and torch.compile kernels for recurrent-model training. |
|
NN4N: Neural Networks for Neurosimulations
A PyTorch package for configurable continuous-time RNNs, including E/I-constrained and multi-area architectures with direct access to hidden dynamics. Implements sparsity, sign, and per-connection plasticity masks to enforce structured connectivity and scale gradients during training; shipped with documentation, tests, and CI. |
Miscellanea |
|
Service
|
Teaching Assistant, ESE 5460: Principles of Deep Learning, Fall 2025 Teaching Assistant, PHYS 5585: Comp. and Theoretical Neurosci., Spring 2026 |
|
Design and code of this website is adapted from Jon Barron's website. |