Multi-Scale Nested Learning for Hierarchical Memory Systems in Continual Learning

TL;DR: What if we organize neural memories like Russian dolls—nested at different scales—to help AI remember both old and new things better? Concretely, let's build a system where each "level" of memory is tuned to a different timescale, and see if this improves the balance between retaining old knowledge and absorbing new information in continual learning benchmarks.

Research Question: Can a multi-scale, hierarchically nested memory system—where each memory level operates at a distinct temporal and contextual scale—enhance continual learning performance by improving retention and adaptability compared to flat or single-scale memory systems?

Hypothesis: Hierarchically nesting memory systems as per NL, with each level responsible for a different temporal granularity (e.g., short-, medium-, and long-term), will reduce catastrophic forgetting and increase flexibility in continual learning tasks, outperforming conventional memory architectures.

Experiment Plan: Design a continual learning framework based on the NL paradigm, where nested memory modules with varying update frequencies and context spans are integrated. Use benchmarks from both vision (e.g., Split CIFAR-100) and language (e.g., sequential text corpora) domains. Compare against single-scale memory systems, as well as replay-based and regularization methods (see Ma et al., 2025; Zhu et al., 2023). Key metrics: retention (accuracy on old tasks), adaptability (accuracy on new tasks), and memory usage efficiency. Analyze ablation studies to assess the contribution of each memory scale.

References:

• Behrouz, A., Razaviyayn, M., Zhong, P., & Mirrokni, V. (2025). Nested Learning: The Illusion of Deep Learning Architectures.
• Ma, Z., Ma, Y., Hong, X., Li, H., & Zhang, S. (2025). Joint Memory Optimization for Continual Learning. IEEE Transactions on Circuits and Systems for Video Technology.
• Zhu, Z., Lyu, W., Xiao, Y., & Hoiem, D. (2023). Continual Learning in Open-vocabulary Classification with Complementary Memory Systems. Transactions on Machine Learning Research.