Manifestro
DREAM
Dynamic Recall and Elastic Adaptive Memory — a neural architecture for continuous-time adaptation with surprise-driven plasticity.
Overview
DREAM is a continuous-time recurrent neural network that adapts during inference. Unlike static architectures, it modulates plasticity and integration speeds based on prediction error — enabling real-time learning without gradient updates.
Architecture
Surprise-Driven Plasticity
Learning rates modulated by prediction error surprise using Hebbian weight updates.
Liquid Time-Constants
Integration speeds adapt continuously based on input novelty and temporal context.
Fast Weights
Low-rank weight decomposition enables efficient meta-learning with minimal overhead.
Sleep Consolidation
Memory stabilization during low-surprise periods prevents catastrophic forgetting.
Benchmarks
| Model | Parameters | Improvement | Adaptation |
|---|---|---|---|
| DREAM | 82K | 99.9% | 0 steps |
| LSTM | 893K | 93.9% | 0 steps |
| Transformer | 551K | 92.6% | 0 steps |
Applications
Online Learning
Real-time adaptation to streaming data
Speech Recognition
Acoustic modeling with temporal dynamics
Time Series
Forecasting with regime change detection
Few-Shot Learning
Rapid task adaptation
Anomaly Detection
Reconstruction-based outlier detection
Non-Stationary Data
Handle distribution shift