Fig. 2: Hybrid neural networks based on corticohippocampus circuits and metaplasticity mechanisms.
From: Hybrid neural networks for continual learning inspired by corticohippocampal circuits
a The hybrid neural network comprises an ANN (depicted in pink) and a SNN (depicted in green). The ANN is trained on the similarities among image samples to generate episode-related regularities for each task, which modulate the SNN. The SNN is tailored to learn sequential specific tasks, thereby generating specific memories. b The learning process within the SNN is modulated by metaplasticity mechanism. Large synaptic spines, depicted in green, have stored substantial amounts of memory and learn at a slower rate in subsequent learning. Small synaptic spines, colored in pink, store less memories and are capable of learning additional knowledge. The size of synaptic spines varies across different episodes. c The impact of metaplasticity mechanism on learning dynamics across different spine sizes, illustrating the decline in learning capability as the absolute value of neural weights increases.
