Conditioned reflex

Bayliss (1915) described two fundamental mechanisms in Pavlov's methods, temporary association and the analysers. In temporary association, a stimulus given several times with food in the mouth eventually excites salivary secretion when given alone.² Bayliss also reported limits. Nocuous skin stimuli could become a conditioned stimulus for feeding, but not over bone, and not when acid in the mouth replaced food.³

Pavlov (1927) described experimental extinction, in which a conditioned stimulus is given without reinforcement. He noted that extinguishing one conditioned reflex also weakens others that were not directly extinguished.⁴ This secondary extinction reached across reflexes that shared an unconditioned reflex and across those that did not. In some cases it was profound enough to involve even the unconditioned reflexes.⁵

Abramson and Levin (2021) argued that behavioural scientists have given little attention to inconsistent definitions of conditioning procedures. Classical conditioning is the main example. They counted at least four methods, ranging from general activity conditioning to injecting the unconditioned stimulus directly into the organism. They state that no research directly compares these procedures. They doubt that the four all produce the same phenomenon.⁶ For researchers working with novel organisms, this means the name of the procedure does not fix what was done.

Biswas, Clawson and Levin (2022) tested models of biological gene networks for Pavlovian learning. They picked three nodes in each model as an unconditioned stimulus, a neutral stimulus and a response. Training turned the neutral node into a conditioned stimulus.⁷ Bongard and Levin (2023) report numerous examples in biological networks and many fewer in random control networks. They suggest evolution enriches for this property.⁸ Levin (2025) calls the result a dynamical state memory that needs no change in network topology or edge strength.⁹