Continual Learning Without Replay: New Method Leads Baselines by Up to 30 Points Across 6 Streaming Tasks
Horror-Celery4869 · reddit · 2026-11-09
A benchmark chart compares continual learning methods — "Apical Plexus" and "Apical Arbor" (both zero-replay, no raw data storage) against Experience Replay (stores raw data) and EWC/Online SGD (backprop-based) — across 6 streaming regimes with 32 seeds:
- Raw Split-MNIST: Plexus 90.7% vs EWC 20.1%
- Raw Split-CIFAR-100: Plexus 36.3%, 2.07x over Replay
- CIFAR-100 with ResNet-18 / DINOv2: Plexus leads at 22.5% / 84.8%
- CLINC150 · Qwen-2.5: Plexus 74.8%, +30.0 pp over Replay
- CLINC150 · MiniLM: Plexus 93.5%
Claimed resource profile: zero persisted data growth, <0.18ms CPU step latency, 101–429 KB static SRAM edge footprint, 3.96–8.87 MB cloud footprint.
More from Research
- MoRE: Mixture-of-Experts 3D World Model Handles Rigid, Plush and Cloth Objects for Real Robots — anand_bhattad · 2026-11-09
- wsp writes facts directly into model weights with closed-form, gradient-free solves — kastnerkyle · 2026-11-09
- Learning Abstractions: Why Immediate Concept Acquisition Is AI's Next Frontier — jmugan · 2026-11-09
- LoRA-tuned Qwen3 skips first output line whenever the caption is under ~4 tokens — prathamdmehta · 2026-11-09
- Ramez: OpenAI's Productivity Gain Per ECI Point Appears to Decline Over Time — ramez · 2026-11-09
- Betting on transformers: incremental improvements will beat radically new LLM architectures — max_paperclips · 2026-11-09