Fig 1.
Hypothesized relationships between stimulus discrimination and detection during adaptation.
Cartoon illustration of how the gross detection of input may differ from the ability to discriminate fine input differences during adaptation following stimulus onset. One possibility–the covarying hypothesis–is that the highly active transient response carries the most information about the stimulus regardless of whether we consider discrimination or detection. In this view, adaptive depression of synapses reduces information transmission and the critical dynamics of the steady state are too noisy for effective discrimination. Alternatively, the trade-off hypothesis is that the strong onset response is good for detection, but lacks the selectivity needed for good discrimination. This view is in line with the prediction that the critical dynamics of the steady state optimize information transmission.
Fig 2.
Adaptation tunes cortical dynamics from large-scale transient response to scale-free steady-state.
(A) Motion-enhanced movies were projected onto the retina while recording local field potential (LFP) with a microelectrode array in visual cortex of the turtle ex vivo eye-attached whole-brain preparation. (B) Shown are LFP traces from a subset of 12 electrodes. Intense population activity occurs at the onset of the movie; adaptation leads to more moderate steady-state activity. To characterize these changes in population activity, we analyze ‘avalanches’ which are defined as spatiotemporal clusters of LFP peaks beyond ±3 SD (black ticks). Avalanche size was defined as the number of LFP peaks comprising the cluster. (C) LFP peak rate time series, averaged over 80 movie repetitions. (D) Each point represents the size and time (middle of duration) of one avalanche. Avalanches from 80 trials are overlaid. Avalanches were typically very large during the transient response, but adaptation resulted in smaller and more diverse sizes. (E) Typical distributions of avalanche sizes during the transient (blue, 0–1 s after movie onset) and the baseline period (green, 2–5 s after movie onset). During the transient, the distribution exhibited a ‘bump’ at large size indicating a high likelihood of very large avalanches. During baseline, avalanche size distributions were well-described by a power-law function, in line with recent findings that adaptation tunes cortical network dynamics to criticality. Green box delineates the expected range (5–95 percentile) of probabilities for a finite sample drawn from a perfect power law. (F) To quantitatively characterize population dynamics during different time periods we computed δ, which measures deviation from the baseline distribution. Calculation of δ is based on differences between cumulative distributions like the examples shown. (G) Shown is a summary of 14 experiments. Without exception, the transient exhibited many more large avalanches than the baseline (δ>0) while the steady state period (4–5 s after movie onset) exhibited small deviations, both positive and negative, from baseline.
Fig 3.
Adaptation enhances discrimination at the cost of reduced detection: LFP rate coding.
(A) TOP: Typical rasters of LFP peaks (taken from all electrodes, randomly subsampled to 10%) showing response to 13 trials with the same background movie visual stimulus (no foreground). Detection was quantified based on the LFP peak count when the movie was off (pink box) compared to when the movie was on (blue or orange box). BOTTOM: For discrimination a foreground stimulus (4 different red dots) was presented during each of four blocks with 13 trials each, all with the same background movie stimulus. The red dot was presented either during the transient period just after background stimulus onset (blue box) or later during the steady-state (orange box). Note that the different foreground stimuli were more easily distinguished by the LFP peaks in the steady-state compared with the transient, but the presence of the background stimulus is more easily detected based on strong transient response. (B) Summary (n = 14 turtles) of how well the LFP peak count can detect the presence of the background stimulus. All turtles show a decrease in detection from transient to steady-state. (C) Summary of how well the LFP peak count can discriminate the four different foreground stimuli. Most turtles showed an increase in discrimination from transient to steady-state. (D, E) A more refined explanation of detection and discrimination is obtained by comparing to δ. Generally, lower δ resulted in enhanced discrimination and poorer detection, while higher δ exhibited the opposite trend. Thus, power-law distributed (low δ) population dynamics are associated with a functional trade-off, gaining discrimination at the cost of decreased detection. The inset is an expanded view to better show the correlation between discrimination and δ during the steady-state.
Fig 4.
Adaptation enhances discrimination at the cost of reduced detection: MUA rate coding.
(A) Based on the entire recording for each turtle, the LFP count per 1 second time window is strongly correlated with the MUA spike count per 1 second. MUA was generally lower rate than LFP. Each shaded region corresponds to one experiment. The vertical extents of the shaded regions delineate quartiles. (B) Similar to our observations based on LFP peak response (Fig 3), we found that when response was defined based on MUA spike counts, detection is highest during the transient response and decreases as adaptation progresses to a steady state. The decrease in detection during the steady-state compared with LFP-based detection is likely due to the lower activity rate of MUA relative to LFP. (C) MUA spike response also exhibited increased discrimination during the steady state compared with the transient.
Fig 5.
Adaptation enhances discrimination at the cost of reduced detection: Temporal coding.
(A) The time course of population response during the 1 s (0.2 s resolution) following the four different foreground stimuli (red dots) showed little difference during the transient response. Black lines indicate response averaged over repeated trials. Gray lines indicate individual trials. Each response was subtracted by its time average and normalized by its variances to emphasize effects of rate coding. One example turtle shown. White dots indicate the 5 bins used to compute the 5 bit temporal response. (B) During the steady-state, different foreground stimuli evoked differing temporal structure of responses. Thus, temporal structure carries useful information for discrimination. (C) Summary of discrimination based on temporal structure for 14 experiments. Most turtles showed an increase in discrimination from transient to steady-state. (D) Variability in discrimination is better explained when changes in δ are accounted for. Similar to rate-based discrimination, power-law distributed (low δ) population dynamics are associated with optimal temporal discrimination. (E) Summary of detection based on temporal structure for 14 experiments. Detection typically decreased from transient to steady-state. (F) Variability in detection is better explained when changes in δ are accounted for. Similar to rate-based detection, power-law distributed (low δ) population dynamics are associated with low detection.
Fig 6.
A network model with short-term depression reproduces the discrimination-detection trade-off.
(A) The model network is driven by a slowly varying ‘background’ stimulus that turns on at t = 500 timesteps. We interpret 1 time step to be approximately 1 ms. (B) Avalanches with a tendency for very large sizes occur during the transient following stimulus onset (blue). Smaller avalanches occur during the baseline period (green), which includes the later time period labeled ‘steady-state’ (orange). Avalanches are overlaid from 80 repetitions of the same background stimulus. (C) Synapse strength (averaged over all except the input synapses) drops at stimulus onset and fluctuates due to short term depression. Gray lines indicate single trials (n = 80); black line is the cross-trial average. (D) Avalanches are distributed according to a power-law during the baseline (green) and have a high likelihood of very large avalanches during the transient (blue). Inset: cumulative distributions of the same data reveal that δ>0 for the transient. (E) A ‘foreground’ stimulus (red) is applied at two times: during the transient and later during the steady-state. (F) Raster of model spikes (from all neurons, subsampled) including 80 trials, broken into four blocks of 20, each with a different intensity of foreground stimulus. Response was defined as the spike count during the transient (blue) or the steady-state (orange). (G, H) Consistent with our experiments, discrimination of foreground stimuli was inversely proportional to δ while detection was proportional to δ.
Fig 7.
Extreme synaptic depression results in small-scale dynamics and decreased discrimination.
(A) Shown are results from our model with more extreme depression with τd decreased by a factor of 100 compared to the model results in Fig 6. All other model parameters including the stimulus paradigm are unchanged. (B) Subsampled spike rasters for 20 trials for each of 4 different foreground stimulus levels. Note that the steady state spike rate is lower than that in the model results of Fig 6. Spike response does not vary strongly with changing foreground during the steady state. (C) The extreme depression implemented here results in about 50% reduction in cortex synapse strengths. This is a large reduction compared to the ~10% reduction for the model results in Fig 6. (D) During the baseline period (same definition as shown in Fig 6), avalanche sizes are not distributed according to a power-law (green), indicating that the dynamics are not at criticality. Rather, the curvature of the distribution and the lack of large avalanches are consistent with subcritical dynamics. In contrast, the avalanche size distribution during the transient is close to a power-law because the synapses have not yet reached a low enough level to strongly deviate from criticality. (E) In contrast with the results in Fig 6, the transient period (blue) exhibits high discrimination as expected for critical dynamics. The subcritical steady state (solid, orange) exhibits decreased discrimination compared with the critical steady state of Fig 6 (open, orange). Box vertically spans the two quartiles around the median (middle line). Whiskers indicate the range of the data.