Source
Gerstner, Lehmann, Liakoni, Corneil & Brea (2018), LIT-1494, read in full in NOTE-1130: a review of Yagishita et al. (2014), He et al. (2015), Brzosko et al. (2015, 2017) and Bittner et al. (2017). The delay windows those papers measured are THEORY-909, which this entry extends.
The claim
The reviewers read the five papers as the experimental arrival of three-factor learning rules, until then a theoretical assumption. Co-activation of pre- and postsynaptic neurons sets a decaying flag at the synapse, and “a non-zero third factor is needed to transform the eligibility trace into a change of the synapse” and to set the “direction of change” (p. 4). In striatum “the decay of the eligibility trace occurs on a time scale of 1 s” (p. 7); in cortex pre-before-post pairing “leaves an eligibility trace that decays over 5–10 s and that can be converted into LTP by the neuromodulator noradrenaline” (p. 9), with a separate LTD trace converted by serotonin; in CA1 a plateau potential within about 2 s potentiates the input. Windows of seconds are what learning from delayed outcomes in behaviour needs.
What THEORY-909 holds is the measured windows, attributed to the reviewers’ reading. In cortex the direction of change followed the pairing order as well as the modulator, so the two are not separated, and the one reversal of a single protocol’s sign, by dopamine in CA1, took about a minute and may be consolidation. Almost all of it is in slices with imposed protocols, and the reviewers concede “there is nothing—apart from conceptual elegance—to favor eligibility traces over alternative neuronal mechanisms” (p. 12).
What would settle it
The promote_when, which QUESTION-311 also asks: the seconds-scale trace shown in behaving animals, with interference in the trace’s mechanism between action and delayed outcome abolishing learning from that outcome but not from immediate ones, and independent replication of the time constants. Further slice experiments cannot settle it.