MemFlow: A Lightweight Forward Memorizing Framework for Quick Domain Adaptive Feature Mapping
Boosts model performance up to 10% using only forward passes – no gradients needed.
A team of researchers (Jianming Lv, Chengjun Wang, Depin Liang, Qianli Ma, Wei Chen, Xueqi Cheng) has proposed MemFlow, a gradient-free forward-memorizing framework designed for quick domain-adaptive feature mapping. The core idea is to keep a pretrained visual model's backbone frozen and instead use randomly connected neurons to memorize associations between features and predictions. Inside the network, spiking signals propagate, and predictions are generated by retrieving neuron-stored memories based on confidence levels. This approach entirely bypasses gradient-backpropagation-based optimization, which is vastly more time-consuming than forward inference and makes online learning infeasible on low-power edge devices. MemFlow also supports reinforced memorization using unlabeled data, allowing rapid adaptation to new domains without requiring labeled examples.
In extensive experiments across four real-world cross-domain datasets, MemFlow achieved performance improvements of up to 10% while consuming less than 1% of the computational time required by traditional domain adaptation methods. This dramatic efficiency gain makes it practical to continuously adapt models on edge devices like smartphones, IoT sensors, and embedded cameras. The framework is particularly valuable for scenarios where test environments shift frequently (e.g., lighting changes, weather, new camera angles) and where computing power is limited. The code is available on GitHub, and the paper is published on arXiv (2402.14598).
- MemFlow uses a frozen backbone with random neurons that memorize feature-label associations via spiking signals, eliminating the need for gradient backpropagation.
- On four real-world datasets, it delivers up to 10% performance improvement while using less than 1% of the computational time of traditional domain adaptation.
- The framework supports reinforced memorization from unlabeled data, enabling rapid online adaptation on low-power edge devices.
Why It Matters
Enables real-time domain adaptation on edge devices without heavy compute – unlocking new possibilities for IoT, mobile vision, and autonomous systems.