A Scalable Population Code for Time in the Striatum
Read:: - [ ] Mello et al. (2015) - A Scalable Population Code for Time in the Striatum 🛫2023-10-08 !!2 rd citation todoist Print:: ❌ Zotero Link:: Zotero Files:: attachment Reading Note:: Web Rip:: url:: https://www.sciencedirect.com/science/article/pii/S0960982215002055
TABLE without id
file.link as "Related Files",
title as "Title",
type as "type"
FROM "" AND -"ZZ. planning"
WHERE citekey = "melloScalablePopulationCode2015"
SORT file.cday DESCAbstract
To guide behavior and learn from its consequences, the brain must represent time over many scales. Yet, the neural signals used to encode time in the seconds-to-minute range are not known. The striatum is a major input area of the basal ganglia associated with learning and motor function. Previous studies have also shown that the striatum is necessary for normal timing behavior. To address how striatal signals might be involved in timing, we recorded from striatal neurons in rats performing an interval timing task. We found that neurons fired at delays spanning tens of seconds and that this pattern of responding reflected the interaction between time and the animals’ ongoing sensorimotor state. Surprisingly, cells rescaled responses in time when intervals changed, indicating that striatal populations encoded relative time. Moreover, time estimates decoded from activity predicted timing behavior as animals adjusted to new intervals, and disrupting striatal function led to a decrease in timing performance. These results suggest that striatal activity forms a scalable population code for time, providing timing signals that animals use to guide their actions.
Quick Reference
Top Notes
Tasks
Further Reading
Extracted Annotations and Comments
Page 1112
Highlights d Striatal neurons fire at different times over tens of seconds during timing behavior d Response times of striatal neurons rescaled with the interval being timed d Time coding by the population predicted timing behavior from trial to trial d Striatal neurons multiplexed information about action and time
Page 1112
In Brief Time is fundamental for all behavior, yet how the brain encodes time is unknown. Mello, Soares, and Paton found that firing dynamics in populations of neurons in the rodent striatum robustly and flexibly encoded time over tens of seconds. These results supply new insight into how the basal ganglia might function during learning and action selection.
Page 1113
Multiple lines of evidence implicate the basal ganglia (BG) as a locus for the representation of such temporal information. Lesions of the striatum in rats [2], disease states that affect the BG such as Parkinson’s [3] and Huntington’s disease [4], drugs that affect dopamine (DA) signaling [5], and genetic manipulations that affect the DA system in the BG [6] all result in interval timing dysfunction. Furthermore, human fMRI studies have found that the striatum, a main input area of the BG, is activated by tasks that involve the processing of interval information [7, 8].
Page 1113
Strikingly, we found that temporal tuning stretched or contracted, rescaling with the interval being timed. Thus, striatal populations encoded relative time, flexibly adapting to the immediate demands of the environment. Finally, we ran a simple simulation of the task and show that neural responses resembling those we observe in the striatum are suitable as a basis for timing behavior.
Page 1114
Briefly, rats were placed in a behavioral box containing a lever positioned over a liquid delivery port and were trained to press the lever to receive water reward. Reward delivery triggered a timer, and reward became available again only after the timer exceeded a FI ranging from 12 s to 60 s in multiples of 12 s. Lever presses occurring after reward delivery but before the FI had elapsed were not reinforced. A FI was maintained for between 18 and 40 rewards before changing to another FI, randomly chosen from the interval set
TRIAL
Page 1114
ever, this did not reflect the pattern of responding in single trials. We asked how pressing evolved after pressing onset (pressing start times, PSTs) in each trial by aligning on the PST and averaging lever press rates across trials and within blocks of the same FI (Figure 1D).
BLOCK
Page 1114
Fig. 1
Page 1115
mental Experimental Procedures for details) shown in Figure 2A. Some cells fired just after reward delivery, others fired in the middle of the delay, and others fired leading up to the next reward (Figures 2A, S2, and S3). This produced a slow-moving ‘‘bump’’ of activity that traversed the population during each FI.
Page 1115
In theory, reading out the location of this bump in the population could provide an estimate of time within the FI. However, a core feature of interval timing behavior is that timing accuracy decreases with the magnitude of the interval being timed [9]
Page 1115
In addition, the density of peak firing rate latencies in our population decreased over time within the FI (Figure 2C). Thus, the bump in activity within the striatum population moved progressively slower as the FI wore on. Strikingly, the overall time taken by this bump to traverse the population appeared to scale with the FI (Figures 2A and S4A)
Page 1115
peak firing rate latencies
What?
Page 1115
. Of the 112 neurons recorded in all FIs, we found that 76 neurons (68%) maintained their ordinal position in time across the population (see Supplemental Experimental Procedures for details). The responses of these neurons can be observed in Figure 2A, wherein the position of cells along the y axis is the same across the panels displaying average responses in each of the FIs (for all recorded cells, see Figure S4A).
OK so even in different time scales they responded in the same order?
Page 1115
The above analyses of striatal neural responses indicate a gross correspondence between striatal activity and timing behavior across blocks of trials, suggesting that striatal activity patterns
Page 1115
Fig. 2
Kinda curious what the average jitter of the neurons were relative to their average
ok they more or less make that explicit in B
Page 1116
might guide decisions about when to begin pressing the lever during each FI.
Page 1116
decoder
Trying to figure out what this means :/
Page 1116
our decoder was constructed as follows. In each of the first seven trials of a block, we counted spikes within defined time bins and asked how likely we were to have observed that number of spikes at each time given the observed distributions of spike counts in trials 8 onward of the corresponding block.
Yeah I am not following at all
Page 1116
This generated a likelihood function for current time, given an observed spike count in each bin, for each individual cell.
Page 1116
To derive a measure of the population’s estimate of the likelihood for current time, we multiplied together the individual cells’ likelihood functions. We then took the mean of this likelihood function as our estimate for current time [17].
Page 1116
Next we asked whether such timing signals may be used by animals to guide timing behavio
Page 1116
Based on previous studies [18–20], we expected that striatal neurons would display significant modulation by behaviors during the FI
Page 1118
? A dominant view supported by a wide range of neurobiological data posits that the BG implements aspects of reinforcement learning (RL) [1, 20, 25–28], learning how an organism ought to act in order to maximize reward. However, to learn about the sometimes-delayed consequences of actions and to guide future behavior toward rewarding outcomes, it is absolutely necessary that the brain represent situations and actions through time [1, 29].
Page 1118
Indeed, temporal relations among actions and events contain the causal information that learning systems have evolved to detect through a process sometimes referred to as credit assignment [30]. Once credit for the occurrence of predictable events has been assigned, this information must be used to profitably guide the course and timing of action as situations arise. This continuous learning-behaving cycle is what RL algorithms naturally account for [29]. Yet, it is not known how the BG, the brain system most often associated with RL, represents temporal relationships over the durations necessary to explain its purported role in animal learning and behavior
Page 1118
The sequential neural states that we describe in the striatum during timing behavior can provide a unifying view of the BG’s role in timing and RL. These signals are strikingly similar to temporal basis functions proposed in existing learning models as more neurally plausible and efficient representations of time [21–23], which we show can be used to generate timing behavior similar to what we observed experimentally. Such models operate by learning a set of weights used in a weighted sum of the temporal bases to construct a moment-by-moment prediction about future events such as expected reward. In theory, a weighted combination of activity patterns in the cortical or thalamic inputs to the striatum could act as such temporal bases and modulate the responses of striatal neurons that we observed
Page 1119
Our data appear most consistent with theoretical models that suggest distributed representations of time encoded by the joint activity of populations of neurons [13]. Indeed, the decoder used in the current study assumes that time information may be present in many different neurons. However, we cannot rule out that upstream of the population we recorded in the striatum, other forms of temporal representations may exist. For instance, an accumulating process such as that contained within pacemaker accumulator models [9] might act to trigger neurons to become active at different delays as the accumulator passes a series of thresholds.
Figures (blue)
Figure
Page 1117
Figure 4
Figure 4. Single-Trial Estimates of Elapsed Time Decoded from the Population Response Correlate with True Time during Initial Trials of 12-s and 60-s FI Blocks (A) Decoded population estimates of elapsed time from reward in single trials, for the first seven trials of the 12-s FI block plotted against true time. Red traces indicate the mean of the population likelihood function, and the underlying heatmap indicates the population likelihood function. The last panel shows a seven-trial average likelihood function using the first seven trials of the 12-s block. (B) Decoded estimates of elapsed time for the first seven trials of the 12-s FI block plotted on the same axis. Curves are quadratic fits to the mean likelihood function of each individual trial (red lines in first seven panels). Red curves represent early trials, and black curves represent later trials. (C) Same description as in (A), but for the 60-s FI. (D) Same description as in (B), but for the 60-s FI. See also Figure S4.
Figure 4. Single-Trial Estimates of Elapsed Time Decoded from the Population Response Correlate with True Time during Initial Trials of 12-s and 60-s FI Blocks (A) Decoded population estimates of elapsed time from reward in single trials, for the first seven trials of the 12-s FI block plotted against true time. Red traces indicate the mean of the population likelihood function, and the underlying heatmap indicates the population likelihood function. The last panel shows a seven-trial average likelihood function using the first seven trials of the 12-s block. (B) Decoded estimates of elapsed time for the first seven trials of the 12-s FI block plotted on the same axis. Curves are quadratic fits to the mean likelihood function of each individual trial (red lines in first seven panels). Red curves represent early trials, and black curves represent later trials. (C) Same description as in (A), but for the 60-s FI. (D) Same description as in (B), but for the 60-s FI. See also Figure S4. Page 1117