C. Myers
Synopsis
The octopus is a lower animal which displays an impressive ability for learning visual discrimination of shapes and objects. This requires maintenance of the original stimulus image in memory during approach and attack until, for example, reinforcing taste sensations occur. OVSIM (Octopus vulgaris simulation) achieves this type of delay learning through attention-driven buffering (ADB) of inputs to a neural network capable of reinforcement learning; it is applied to discrimination tasks which models one basic aspect of discriminative learning in the octopus.
Experiments with OVSIM mimic those with octopus in such features as shape of learning curve, fall in delay to attack with training, and mutual interference of positive and negative memories. Further, when the ADB mechanisms in OVSIM are damaged, it begins to malfunction similarly to octopus after ablation of learning centres.
These correspondences suggest that the manner in which OVSIM learns - by adjusting probabilities of acting and by ADB - may well be related to the octopus's strategy for maintaining stimulus images during the delay until reinforcement arrives.
Key words:
neural networks, reinforcement learning, delay learning, adaptive learning, animal learning models, octopus