📄 other

A bio-inspired hybrid lifelong learning based on self-selective memristors

This paper proposes a bio-inspired hybrid lifelong learning system that co-designs a Drosophila-human hybrid algorithm with self-selective memristor hardware to overcome catastrophic forgetting and energy inefficiencies, achieving significant speed and energy improvements over traditional CMOS accelerators.

Original authors: Yishu Zhang, Xuemeng Fan, Guobin Zhang, Zhejia Zhang, Xinheng Mei, Zijian Wang, Wenjue Zhou, Daying Sun, Lei Deng, Mingkun Xu, Shuai Zhong, Yi Tong, Bin Yu, Rong Zhao, Dashan Shang, Xiaolei Zhu, Qing

Published 2026-07-27

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Yishu Zhang, Xuemeng Fan, Guobin Zhang, Zhejia Zhang, Xinheng Mei, Zijian Wang, Wenjue Zhou, Daying Sun, Lei Deng, Mingkun Xu, Shuai Zhong, Yi Tong, Bin Yu, Rong Zhao, Dashan Shang, Xiaolei Zhu, Qing Wan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to play every video game ever made, one after another. The problem is, when the robot learns "Space Invaders," it often forgets how to play "Pac-Man." This is called "catastrophic forgetting," and it's a huge headache for artificial intelligence. To fix this, scientists usually try to either freeze the robot's brain so it can't forget (but then it can't learn new things) or let it learn freely (but then it forgets the old stuff). Nature, however, has a better trick. Our brains don't just store everything forever; they actively forget the boring stuff to make room for the important stuff, and they have different types of memory that stick around for different amounts of time. This paper explores a new way to build AI that mimics this biological balancing act, but instead of using slow, energy-hungry computer chips, it uses a special kind of electronic component that acts like a living memory cell.

The researchers, led by Yishu Zhang and a team from institutions like Zhejiang University and Tsinghua University, have created a "hybrid lifelong learning" system. Think of it as a robot brain built from two very different parts working together. First, they designed a smart software algorithm that combines two biological ideas: "active forgetting" (inspired by fruit flies, which forget things to stay efficient) and "metaplasticity" (inspired by humans, which strengthens important memories). Second, they built a physical hardware device called a "self-selective memristor" (SSM) that naturally behaves like this software. These tiny devices can change their electrical resistance to store information, but unlike normal computer memory, they have a built-in "leak" that slowly fades away weak memories while keeping strong ones, just like a human brain.

The team didn't just simulate this on a computer; they actually built a small chip with 1,024 of these memristors (arranged in a 32x32 grid) and tested it. They found that by tuning how fast these devices "forget," they could create a team of five different "memory cores" within the chip. Some cores were set to be very forgetful and flexible (great for learning new, tricky tasks quickly), while others were set to be very stable and stubborn (great for holding onto old, important skills). When they tested this system on a series of 20 different image recognition tasks (using a dataset called CIFAR), the system successfully learned all of them without forgetting the previous ones.

The results were impressive. Compared to standard computer chips (CMOS), this new system was projected to be 3.42 times faster and use 91 times less energy. The authors suggest that this approach solves the age-old struggle between keeping old knowledge safe and learning new things, offering a path toward AI that can truly learn continuously in the real world, just like we do. They emphasize that while the core learning rules were tested on hardware, the full performance numbers for the 20-task challenge were based on simulations calibrated with their real-world hardware measurements, showing that the physics of their new devices can indeed support complex, lifelong learning.

Technical Summary: A Bio-Inspired Hybrid Lifelong Learning System Based on Self-Selective Memristors

Problem Statement
Artificial intelligence systems operating in dynamic, open-ended environments require lifelong learning capabilities—the ability to continuously acquire and retain knowledge while preserving prior skills. Current AI models face the "stability-plasticity dilemma," where learning new tasks often overwrites established representations, leading to catastrophic forgetting. Existing solutions generally fall into two divergent categories: regularization-based methods that protect knowledge at the cost of adaptability, and dynamic architecture or memory replay methods that promote flexibility but risk destabilizing learned representations. Furthermore, these software-centric approaches impose significant computational overhead and memory costs. A critical disconnect exists between complex bio-inspired algorithms and the hardware used to run them; most bio-inspired models adopt isolated biological mechanisms (e.g., active forgetting or metaplasticity) rather than synergistic integration, and they rely on von Neumann architectures that separate memory and processing, leading to prohibitive energy and latency costs for edge applications.

Methodology
The authors propose a hardware-algorithm co-design paradigm that bridges this gap by integrating complementary neurobiological principles into a hardware-native learning system.

Algorithmic Framework (Hybrid Lifelong Learning - HLL):
The HLL model fuses two distinct biological mechanisms:
- Drosophila-inspired Dynamic Active Forgetting (DAF): Mimicking the Drosophila mushroom body, this mechanism prunes redundant synapses to optimize resource allocation and enhance plasticity. It involves an "active forgetting rate" ( $\alpha$ ) where weakly potentiated inputs decay rapidly while strongly potentiated inputs are retained.
- Human Brain (HB)-inspired Metaplasticity and Consolidation: Drawing from human brain mechanisms, this component safeguards consolidated knowledge and promotes forward transfer. It utilizes a global gain parameter ( $\beta$ ) to modulate synaptic plasticity based on task-specific expectations and errors, analogous to dopamine-mediated metaplasticity.
Hardware Implementation (Self-Selective Memristors - SSM):
To physically instantiate these dynamics, the authors fabricated 1-kb (32×32) crossbar arrays of ITO/InOx/GaOx/Pt self-selective memristors.
- Device Physics: The bilayer heterostructure features a GaOx layer with low oxygen vacancy concentration (7.6%) acting as a rectifying barrier, and an InOx layer with high oxygen vacancy concentration (65.33%) facilitating resistive switching. This asymmetry yields high self-rectification (SR > 10⁶) and nonlinearity (NL > 10⁶), eliminating the need for external selectors and preventing cross-talk.
- Intrinsic Dynamics: The devices exhibit tunable conductance decay, emulating active forgetting. The decay rate is stimulus-dependent, allowing for programmable retention ratios. The devices also support essential synaptic functions like STDP, PPF, and PPD.
- System Architecture: The system integrates five parallel Synaptic Processing Cores (SPCs). Each SPC is programmed with distinct, tunable weight-update dynamics (via pulse amplitude and frequency) to create a gradient of plasticity-stability. SPC1 is configured for high plasticity (rapid adaptation, ~92% forgetting rate), while SPC5 is configured for high stability (knowledge retention, ~37% forgetting rate).
Co-Design Strategy:
The system operates via a distributed processing architecture where all five cores process incoming tasks in parallel. The core whose plasticity-stability profile best matches the current task becomes the dominant attractor. A hardware-compatible gain parameter ( $\beta$ ) bridges algorithmic rules and device physics, modulating pulse parameters to implement neuromodulation and sparse gating.

Key Contributions

Synergistic Biological Integration: The work moves beyond isolated bio-mechanisms by integrating Drosophila-inspired active forgetting with human-derived metaplasticity into a single, unified learning model.
Hardware-Native Emulation: The SSM devices physically instantiate the stability-plasticity continuum. The intrinsic conductance decay and self-rectifying properties of the hardware directly execute the learning rules (forgetting and consolidation) without relying on software-based regularization or external selectors.
Scalable Architecture: The design achieves terabit-scale storage capacity (up to 17.1 Tb with 1/3V scheme) under a 10% read margin, enabled by the high self-selectivity of the SSMs which suppresses sneak-path currents.
Gradient Plasticity-Stability Continuum: The five-core architecture creates a hardware-native continuum from rapid reconfigurability to long-term retention, allowing emergent collaborative task partitioning without explicit core selection logic.

Results
The system was evaluated on 20 sequential CIFAR-variant tasks (S-CIFAR-100, R-CIFAR-100, R-CIFAR-10/100) using hardware-calibrated simulations validated by per-tile hardware testing.

Performance: The system achieved an average test accuracy of 68.5%, comparable to CPU (70%) and GPU (70.3%) implementations, while significantly outperforming DAF-only models.
Efficiency: Compared to conventional CMOS accelerators, the SSM-based system demonstrated a 3.42× speedup and a 91× improvement in energy efficiency.
Stability-Plasticity Balance: The system successfully balanced forward transfer (FT) and backward transfer (BWT). The optimal active forgetting rate ( $\alpha$ ) range was found to be around 120%; values below this led to insufficient transfer, while values above caused catastrophic performance degradation.
Robustness: The system maintained performance despite device-to-device (D2D) and cycle-to-cycle (C2C) variations, which were minimized through robust fabrication (D2D variation in LRS was 2.91%).
Hardware Validation: Readout errors across the 1-kb array were constrained within ±0.06%, and tile-level relative errors were 0.08%, confirming high computational fidelity.

Significance
The paper claims to establish a scalable, hardware-algorithm-integrated foundation for the efficient deployment of lifelong learning in real-world systems. By resolving the stability-plasticity trade-off through a multi-mechanism synergy embedded directly in memristor physics, the work bridges the long-standing divide between complex bio-inspired algorithms and practical hardware implementation. It demonstrates that neuromorphic AI can move beyond passive memory substrates to active, adaptive computing platforms capable of continuous, resource-efficient adaptation in dynamic environments. The authors position this co-design paradigm as a critical step toward autonomous systems that can operate effectively in open-ended, real-world scenarios without the energy and latency penalties of traditional von Neumann architectures.

Technical Summary: A Bio-Inspired Hybrid Lifelong Learning System Based on Self-Selective Memristors

More like this