Scholar
Zhixuan Lin
Google Scholar ID: BiyrJeMAAAAJ
University of Montreal; Mila
Citations & Impact
Citations
512
H-index
4
i10-index
4
Publications
6
Co-authors
0
Publications
Resume
Academic Achievements
- Adaptive Computation Pruning for the Forgetting Transformer, COLM 2025 (First author).
- Forgetting Transformer: Softmax Attention with a Forget Gate, ICLR 2025 (First author).
- The Curse of Diversity in Ensemble-Based Exploration, ICLR 2024 (First author).
- Improving Generative Imagination in Object-Centric World Models, ICML 2020 (First author).
- SPACE: Unsupervised Object-Oriented Scene Representation via Spatial Attention and Decomposition, ICLR 2020 (Co-first author, marked with *).
- GIFT: Learning Transformation Invariant Dense Visual Descriptors via Group CNNs, NeurIPS 2019 (Collaborator).
Background
- Third-year Ph.D. student at Mila and the University of Montreal, advised by Professor Aaron Courville.
- Research goal is to understand and build general intelligence; currently focused on long-context sequence models (especially linear-complexity models) and their applications in reinforcement learning (RL).
- Views an agent fundamentally as a sequence model, emphasizing the temporal and sequential nature of agent-environment interaction (memory, experience stream, learning, credit assignment, etc.) as central to intelligence.
- Believes RL is likely necessary for superhuman intelligence, but simple future prediction—not just reward-based learning—may remain essential due to limited learning signals.
- Intrigued by the inefficiency of human thought and reasoning in natural language, and questions whether neural networks can develop mathematical or high-dimensional geometric intuition beyond human capabilities.