Presentation Information
[16a-M_123-6]Ultra-wide-range temperature compensation in proton-based real-time analog reservoir computing
〇(M1C)Yoshihiro Furue1, Satoshi Hamasuna1, Kazuya Tsuruda1, Nada Besisa1, Takeaki Yajima1 (1.Kyushu Univ.)
Keywords:
Reservoir computing,Ionic Conductor,Oxide
Constructing a reservoir with protonic devices exhibiting diverse relaxation time scales enables real-time learning and inference on a timescale comparable to that of biological systems. However, analog protonic devices exhibit pronounced temperature dependence, making temperature compensation essential for stable and reliable operation of the reservoir.
Conventionally, reservoir computing has been expected to tolerate temperature fluctuations through its adaptive learning capability. However, the intrinsic temperature dependence of analog protonic devices far exceeds what can be compensated by learning alone, critically constraining the operational temperature window. Here, we introduce a simple yet powerful strategy: feeding the reservoir output back into the reservoir itself. This output feedback enables ultra-wide-range temperature compensation across temperatures from 250 K to 350 K. These results demonstrate that even analog devices with strong temperature dependence can enable reliable operation of analog reservoir computing.
Conventionally, reservoir computing has been expected to tolerate temperature fluctuations through its adaptive learning capability. However, the intrinsic temperature dependence of analog protonic devices far exceeds what can be compensated by learning alone, critically constraining the operational temperature window. Here, we introduce a simple yet powerful strategy: feeding the reservoir output back into the reservoir itself. This output feedback enables ultra-wide-range temperature compensation across temperatures from 250 K to 350 K. These results demonstrate that even analog devices with strong temperature dependence can enable reliable operation of analog reservoir computing.