Supplemental Information: Adaptation can explain evidence for encoding of probabilistic information in macaque inferior temporal cortex Kasper Vinken and Rufin Vogels Supplemental Experimental Procedures We simulated responses as a Poisson process of which the rate for a particular stimulus presentation r t (i.e. a cue or choice stimulus) is determined by a simple model incorporating firing rate dependent fatigue, firing rate dependent recovery, and stimulus-specific adaptation. Specifically, r t is a combination of a fixed rate for the baseline r base and the unadapted rate associated with the cue stimulus r stim (referring to either a low noise face r face, a low noise fruit r fruit, or a high noise cue response r noise ), which can be suppressed according to a stimulus-specific adaptation variable A t that captures the state of adaptation for that stimulus at that time (0 = complete suppression, 1 = no adaptation). In addition, a fatigue variable F t captures the state of neural fatigue at a particular trial. Note that we also enforced a stronger response to choice stimuli by multiplying r stim by a factor of 1.2. = +. A t is determined by stimulus-specific resource variables R t face parameter δ. For example, in case of a face stimulus: or R t fruit and a lower bound = + 1
Both R face fruit t and R t start at a value of 1 and are updated after every stimulus presentation following a simple resource decay model [S1]. That is, after a presentation of a low noise face cue R face decreases and R fruit increases (or vice versa after the presentation of a fruit cue): = = + 1 Similarly, R face t and R fruit t are updated after the presentation of a choice stimulus. Since there were no stimulus specific effects for high noise cue stimuli in the data reported by Bell et al., we considered them as neither face, nor fruit and let both R face and R fruit increase. The interstimulus interval between a choice stimulus and the next cue (mean = 1350 ms, not counting time to initiate a new trial) is longer than that between the cue and choice within a trial (mean = 350 ms, not counting response time). We account for this difference by attenuating the decay of the resource variable after a choice stimulus by a factor of 3. For example, for a face choice that would be: = 1 1 3 To update the fatigue variable F after every stimulus presentation t, we used the net rate for that stimulus, normalized by the maximum unadapted net rate for the face and fruit choice stimuli (r fruit and r face multiplied by a factor of 1.2): = max,. F is then updated with the fatigue parameter α and recovery parameter β:
= + 1+. The value of F is constrained between 0 and 1. If the net firing rate to the previous choice stimulus decreases, the amount of recovery will eventually become higher than the amount of response fatigue (if it is 0, F will increase with β). We accounted for the difference in interstimulus interval mentioned earlier by attenuating the amount of recovery after a cue stimulus by a factor of 3 (i.e. using β/3). Populations of neurons were simulated by generating firing rate values r base, r face, r fruit, and r noise from lognormal distributions. The average baseline firing rate was 10 Hz (SD = 5) and the average net response was 15 Hz (SD = 15) to faces, 6 Hz (SD = 6) to fruit, and.5 Hz (SD =.5) to high noise stimuli. This produced for simulation C response levels similar to those shown in Figure 3A of [S2]. Fatigue and recovery parameters α (mean =.12, SD =.15) and β (mean =.08, SD =.1) and stimulus specific adaptation parameters γ (mean γ =.5, SD =.1) and δ (mean δ =.5, SD =.1) were generated using beta distributions. The experimental design was as described by [S2], except that a block had a fixed number of 50 trials (the mean block length in [S2]). For each simulated neuron, each of the five block-types (0%, 25%, 50%, 75%, 100% faces) was repeated three times in random order, resulting in a total of 750 trials. Behavioral responses were randomly generated to approximate the mean behavioral performances per p(face) as reported in Figure 1B of [S2]. Supplemental Results Contributions of cue and choice stimuli
To examine whether the effect is driven by the probability of a face cue or that of a face choice, one could use the following regression model (C. Summerfield, personal communication): = + + + + h + h + h +, where p(face cue ) is identical to the p(face) of Model 1, and p(face choice ) is based on the history of choices instead of cues. For our simulated data, the effect for the cue is clearly much stronger than for the choice (β 2 = -0.05, versus β 3 = -0.01). This could lead one to conclude that it is indeed the expectation of a face cue that drives the effect (which is false because we did not simulate expectation). Decoding forthcoming cue identity Bell et al. [S2] could decode the upcoming cue from baseline activity on trials were expectation and the stimulus were congruent: faces in high p(face) blocks versus fruits in low p(face) blocks. While they argued that neural expectation signals carried information about its likely identity, it cannot be excluded that these decoding results arose from temporally correlated slow fluctuations in baseline activity, which were unlikely to be equally spread across high p(face) and low p(face) blocks within a session. We confirmed this possibility by first arbitrarily dividing real monkey IT spiking data in blocks, and then decoding block membership from baseline activity, leading to an above chance accuracy similar in magnitude to theirs (data not shown). Supplemental References S1. Mill, R. (2014). Stimulus-Specific Adaptation, Models. In, Encyclopedia of
Computational Neuroscience. D. Jaeger, and R. Jung, ed. (New York: Springer New York), pp. 1 7. S2. Bell, A.H., Summerfield, C., Morin, E.L., Malecek, N.J., and Ungerleider, L.G. (2016). Encoding of Stimulus Probability in Macaque Inferior Temporal Cortex. Curr. Biol. 26, 2280 2290.