Animals can learn causal relationships between pairs of stimuli separated in time and this ability depends on the hippocampus. Such learning is believed to emerge from alterations in network connectivity, but large-scale connectivity is difficult to measure directly, especially during learning. Here, we show that area CA1 cells converge to time-locked firing sequences that bridge the two stimuli paired during training, and this phenomenon is coupled to a reorganization of network correlations. Using two-photon calcium imaging of mouse hippocampal neurons we find that co-time-tuned neurons exhibit enhanced spontaneous activity correlations that increase just prior to learning. While time-tuned cells are not spatially organized, spontaneously correlated cells do fall into distinct spatial clusters that change as a result of learning. We propose that the spatial re-organization of correlation clusters reflects global network connectivity changes that are responsible for the emergence of the sequentially-timed activity of cell-groups underlying the learned behavior. DOI: http://dx.doi.org/10.7554/eLife.01982.001.
We compared performance of recently developed silicon photomultipliers (SiPMs) to GaAsP photomultiplier tubes (PMTs) for two-photon imaging of neural activity. Despite higher dark counts, SiPMs match or exceed the signal-to-noise ratio of PMTs at photon rates encountered in typical calcium imaging experiments due to their low pulse height variability. At higher photon rates encountered during high-speed voltage imaging, SiPMs substantially outperform PMTs.
Abstract Sparse coding is thought to improve discrimination of sensory stimuli by reducing overlap between their representations. Two factors, however, can offset sparse coding’s advantages. Similar sensory stimuli have significant overlap, and responses vary across trials. To elucidate the effect of these two factors, we analyzed odor responses in the fly and mouse olfactory regions implicated in learning and discrimination — the Mushroom Body (MB) and the Piriform Cortex (PCx). In both species, we show that neuronal responses fall along a continuum from extremely reliable across trials to extremely variable or stochastic. Computationally, we show that the range of observed variability arises from probabilistic synapses in inhibitory feedback connections within central circuits rather than sensory noise, as is traditionally assumed. We propose this coding scheme to be advantageous for coarse– and fine-odor discrimination. More reliable cells enable quick discrimination between dissimilar odors. For similar odors, however, these cells overlap, and do not provide distinguishing information. By contrast, more unreliable cells are decorrelated for similar odors, providing distinguishing information, though this requires extended training with more trials. Overall, we have uncovered a stochastic coding scheme that is conserved in vertebrates and invertebrates, and we identify a candidate mechanism, based on variability in a winner-take-all inhibitory circuit, that improves discrimination with training.
Article Figures and data Abstract eLife digest Introduction Results Discussion Materials and methods References Decision letter Author response Article and author information Metrics Abstract Animals can learn causal relationships between pairs of stimuli separated in time and this ability depends on the hippocampus. Such learning is believed to emerge from alterations in network connectivity, but large-scale connectivity is difficult to measure directly, especially during learning. Here, we show that area CA1 cells converge to time-locked firing sequences that bridge the two stimuli paired during training, and this phenomenon is coupled to a reorganization of network correlations. Using two-photon calcium imaging of mouse hippocampal neurons we find that co-time-tuned neurons exhibit enhanced spontaneous activity correlations that increase just prior to learning. While time-tuned cells are not spatially organized, spontaneously correlated cells do fall into distinct spatial clusters that change as a result of learning. We propose that the spatial re-organization of correlation clusters reflects global network connectivity changes that are responsible for the emergence of the sequentially-timed activity of cell-groups underlying the learned behavior. https://doi.org/10.7554/eLife.01982.001 eLife digest Ivan Pavlov famously discovered that dogs would salivate upon hearing a bell that had previously been used to signal food, even when there was no food present. This ability to connect events that occur close together in time is known as associative learning. But how is it supported within the brain? In the late 1940s, neuroscientist Donald Hebb proposed that if one neuron persistently and repeatedly takes part in firing a second neuron, the connection between the two neurons will be strengthened. Thus, if neurons that encode the sound of a bell are active at the same time as neurons that encode receiving food, connections between the two groups will be strengthened, and this might enable the dogs to associate the two events. However, animals can also learn to associate events that do not overlap in time. For example, we can associate a bout of food poisoning with a meal we consumed several hours earlier. In rodents, this type of learning is often studied using a task known as trace eyeblink conditioning, in which a tone signals the delivery of a puff of air to the eye after a short delay. Rodents eventually begin to blink in response to the tone, even thought the tone and the air puff are never presented simultaneously. Two possibilities have been proposed for how this might occur: either the neurons that encode the tone remain active until delivery of the air puff, or different groups of neurons are successively activated in a relay that spans the interval between the tone and the air puff. Now, Modi et al. have used in vivo imaging in awake mice to obtain evidence in favour of the second option. Mice were trained on the conditioning task while imaging was used to follow the activity of neurons in a region of the brain known as the hippocampus. As animals learned the task, neurons in part of the hippocampus called CA1 began to reorganize their firing patterns so that distinct groups of cells were active at each time point in the interval between the tone and the air puff. By contrast, hardly any neurons were active across the entire delay. The organized firing became particularly apparent at the same time as the mice first began to blink in response to the tone, and was only ever seen in animals that learned the task successfully. As well as providing evidence to distinguish between competing theories of associative learning across a delay, this study is the first to follow in real-time the reorganization of networks of neurons within the hippocampus during this common type of learning. https://doi.org/10.7554/eLife.01982.002 Introduction The mechanisms of memory formation have been the subject of considerable study (Morris et al., 1988; Kandel, 2001). Much evidence points to Hebbian plasticity as the neural mechanism for the association of two co-occurring stimuli (Bliss and Collingridge, 1993; Morris, 2003). However, this mechanism alone is not sufficient to account for learning under conditions where the two stimuli are separated in time by more than 100 ms (Levy and Steward, 1983), as has been commonly observed (Solomon et al., 1986; Baeg et al., 2003). One example of such a time-bridging task is trace eyeblink conditioning, where the goal is to associate a neutral tone or conditioned stimulus (CS) with a temporally separated, aversive puff of air to the eye, or unconditioned stimulus (US) (McEchron and Disterhoft, 1999). After many pairings of these stimuli, the subject learns to blink in response to the tone, even though the tone and puff never overlap in time (Tseng et al., 2004; Kalmbach et al., 2009). Lesion studies have shown that the hippocampus is required for learning the trace conditioning task, but not a related delay conditioning task, where the CS and US overlap in time (Büchel et al., 1999; Tseng et al., 2004). These observations indicate a role for the hippocampus during the association of temporally discontiguous events (Wallenstein et al., 1998), specifically during the 'trace' period separating stimulus pairs. How might a network of neurons maintain a representation of the stimulus through time? Two possible models have been proposed. The first model hypothesizes that the representation of the first stimulus is maintained by the sustained firing of stimulus-selective cells through a trace interval (Solomon et al., 1986). Such a model is supported by observations of sustained firing by neurons in the medial prefrontal cortex (Fuster, 1973; Baeg et al., 2003) and the medial entorhinal cortex (Egorov et al., 2002) during working memory tasks. An alternative model proposes that sensory representations are maintained by the sequential activation of groups of neurons (Levy et al., 2005; Howe and Levy, 2007; MacDonald et al., 2013). This view arose from modeling studies that trained simple hippocampal area CA3 network models on the trace conditioning task. Rather than observing sustained firing, the authors found that groups of neurons began to show activity in well-timed, sequential bouts. Neurons representing the CS kicked off this 'relay' of activation, which eventually activated US representing neurons at the appropriate time (Levy et al., 2005; Howe and Levy, 2007). However, this model awaits experimental verification. Sequential activity in hippocampal CA1 cells has been seen previously, albeit in some very different behavioral contexts, including temporal memory tasks (Louie and Wilson, 2001; Pastalkova et al., 2008; Gill et al., 2011). There has also been a series of recent, more closely related studies of hippocampal activity, in which rats or monkeys performed stimulus-retention tasks where they had to remember an odor or visual stimulus in order to receive a reward. Here too, hippocampal CA1 neurons were observed to be active in stimulus-triggered, time-locked sequences (MacDonald et al., 2011; Naya and Suzuki, 2011; Kraus et al., 2013; MacDonald et al., 2013). It is increasingly clear that hippocampal CA1 cells adopt sequential activity patterns when subjects are placed in a behavioral context requiring the bridging of temporally separated stimuli. However, in all the studies where this has been observed, only well-trained subjects were used, leaving the time course and mechanism of the emergence of sequentially timed activity entirely unknown. Despite sequential activity having been implicated in several temporal memory tasks, there is little experimental data on the network changes that underlie its emergence. Functional connectivity, as measured by correlations between neuronal activity in the absence of stimulus presentation, is one way to monitor such network changes (Ts'o et al., 1986; Bair et al., 2001; Fujisawa et al., 2008). In a study using two photon calcium imaging of motor cortex pyramidal neuron activity, changes in spontaneous activity correlations have been inferred to indicate learning-related circuit plasticity (Komiyama et al., 2010). In another study where rats were exploring a novel track, correlations between pairs of place cells increased with increasing exposure to the novel environment (Cheng and Frank, 2008; Dragoi and Tonegawa, 2013). With the ability to measure changes in input from upstream circuits using spontaneous activity correlations across many cells in the network, two-photon recordings allow the testing of predictions from the model proposed for the emergence of sequential activity (Levy et al., 2005; Howe and Levy, 2007). Furthermore, such recordings provide relative cell locations within the hippocampus, allowing one to examine the spatial organization of activity patterns (Hampson et al., 1999; Brivanlou et al., 2004; Kjelstrup et al., 2008). As learning progresses, one should see changes in spontaneous correlations reflecting altered inputs and changes in network connectivity. Two key questions remain unanswered in the absence of data recorded from large numbers of neurons during training on a temporal memory task – how do sequential activity representations in the hippocampus emerge during learning, and what are the underlying changes in network connectivity? In this study, we train mice on a trace eyeblink conditioning task while recording activity from populations of area CA1 neurons using two photon calcium imaging. In order to eliminate the influence of running or changing spatial position on hippocampal activity, we implemented a trace eye-blink conditioning task for head-fixed mice. Further, our activity measurements revealed CA1 network dynamics during learning, as we began with naïve mice and trained them to criterion within the recording session. We found that sequentially timed activity of groups of area CA1 cells emerged progressively during the course of learning. Additionally, mean spontaneous activity correlations at the network level rose transiently, while only the correlations between co-tuned cells remained elevated towards the end of the session. Finally, we observed spatially organized clusters of neurons that had elevated spontaneous activity correlations. These correlation clusters re-organized during learning. Results Using two-photon calcium imaging, we monitored the activity of large numbers of neurons in area CA1 of the dorsal hippocampus in awake, head-fixed mice while they were trained on a trace eyeblink conditioning task (Figure 1A). Figure 1 Download asset Open asset Behavior: trace eye-blink conditioning of mice. (A) Cartoon schematic of experimental system. Two-photon calcium imaging was carried out in area CA1 of the dorsal hippocampus in a head-fixed mouse which underwent trace eyeblink conditioning. A speaker (yellow, b) was used for tone stimulus delivery, while a nozzle (red, a) directed the aversive air-puff. A magnetometer (black, c) was used to monitor eyelid position in order to detect blinks. Scale bars at the bottom of the figure indicate times of stimulus delivery (yellow and red bars for tone [350 ms long] and puff [100 ms long] respectively, along with the gap in-between [250 ms]) as well as data acquisition (green bar, 15 s long) during a single trial. (B) Sample eyelid position signal traces from a mouse undergoing trace eyeblink conditioning. The color scale indicates eyelid position in arbitrary units, with high values indicating eyelid closure (blink). This mouse started reproducibly showing significant blinks before the air-puff (e.g., green arrow) mid-way through the session. (C and D) Eyelid position traces in response to pre-training tone presentation (C), and both tone and puff stimuli during trace eye-blink conditioning (D). Gray traces are from individual trials and black traces are averages across trials. Yellow and red bars at the bottom indicate times of delivery of tone and air-puff respectively. (E) Distributions of performance scores for trace (blue) and pseudo (green) conditioned mice. (F) Average performance score, which is the ratio of tone evoked significant blink (CR) rates to spontaneous blink rates, is plotted for all trace conditioned mice (blue) and pseudo-conditioned mice (green). Error bars indicate SEM. * indicates p<0.05. (G) Learning curves (gray lines) for the nine mice that learned the association to criterion. The learning trial identified for each mouse is marked with a red circle. The black curve shows average performance across mice. The vertical, dotted line indicates the mean learning trial across mice (trial 26). Each learning curve was obtained by first obtaining a binary list of significant response trials for each mouse, and then using a previously described expectation maximization algorithm to calculate the CR probability on each trial, for individual mice. The horizontal, red line indicates the probability of CRs by chance. https://doi.org/10.7554/eLife.01982.003 Mice learn a trace eyeblink conditioning task within a single session Head-restrained mice were trained on a trace eyeblink conditioning task (Tseng et al., 2004), where they learned to associate a neutral tone stimulus (Conditioned Stimulus–CS) with an aversive puff of air to the eye (Unconditioned Stimulus–US) within a single session ('Materials and methods-Behavioral training'). Tone and puff were non-overlapping and separated by a 250 ms interval, thus requiring the subject to maintain a representation or 'trace' of the tone (CS) in order to associate it with the puff (US). Behavioral responses were measured by recording deflections in eyelid position (Figure 1B). We found that naïve mice responded to tone presentation with small, but distinct and measurable eyelid movements early in training, even on trials prior to the introduction of the puff stimulus (Figure 1C). Following repeated pairings of tone and puff, however, blink responses to tone increased significantly (>2 standard deviations [SD]) in amplitude and duration, (n = 9 of 18 mice, Figure 1B,D). Conditioned Response (CR) trials were defined as those trials in the training session whose area under the eyelid–position curve, during the interval spanning tone onset to puff onset, was significantly larger than the pre-training baseline ('Materials and methods-Behavioral training'). We next established a performance score by measuring the ratio of CR blink rates to spontaneous blink rates ('Materials and methods-Data analysis', Figure 1E,F). As a learning control, we randomized the relative timing of tone and puff from trial to trial, for a different set of mice. These pseudo-conditioned mice did not demonstrate an increase in blink amplitude or duration (performance scores, trace = 8.74 ± 2.73, pseudo = 1.96 ± 0.44, mean ± standard error of the mean (SEM); Figure 1E,F, two-sample t test, p=0.012). 9 of 18 trace conditioned mice had performance scores higher than those of the pseudo-conditioned mice. We next examined the performance of each mouse, to assess which individual mice learned the task, and if so, at what trial number in the training session. We used a previously described expectation maximization algorithm to assess whether each individual mouse had learned the association, and to obtain learning curves (Smith et al., 2004). Briefly, the algorithm uses the list of CR trials for a given mouse, along with the chance probability of the occurrence of a well-timed, significant blink to estimate the probability of CR production at each trial in the session (Figure 1G). Furthermore, the trial at which a mouse has learned the task is statistically defined. It is the first trial when the lower 95% confidence interval of the probability of CR production rises and remains above chance. As per this criterion, 9 out of 18 trace conditioned mice learned the association. The mean of the individual learning trials was 26 ± 5 (mean ± SD, n = 9 learners; for six of these learner mice, imaging data was also acquired, Video 1 shows high-speed video of mouse blinks before and after learning). These were the same nine mice that also had higher performance scores than pseudo-conditioned controls. Video 1 Download asset This video cannot be played in place because your browser does support HTML5 video. You may still download the video for offline viewing. Download as MPEG-4 Download as WebM Download as Ogg Mouse Behavior. Video of mouse eyeblink behavior acquired at 100 frames per second (fps) and played back at 10 fps (i.e. 0.1x speed). A yellow spot in the bottom-right corner indicates when tone is being delivered, and a red spot indicates when the air-puff is being delivered. Frames from three trials, at different points in the session are shown, depicting behavior early in the session, prior to learning, and late in the session after learning. The last portion of the video shows eyeblink behavior in a probe-trial, where the tone was presented but no air-puff was delivered. https://doi.org/10.7554/eLife.01982.004 Half the mice trained on the trace eyeblink conditioning task failed to learn to criterion. In part, this was because we were restricted to training mice for only a single session due to the limited residence time of calcium indicator dye in cells (Stosiek et al., 2003). These non-learners most likely represent a heterogeneous population of mice at different stages of learning involving multiple brain regions (Kalmbach et al., 2009). In other words, given more training sessions, many of these mice would likely have learned this task to criterion. Consequently, the interpretation of area CA1 calcium imaging data from 'non-learners' is complicated by the uncertain state of learning of each mouse. Hence, in most analyses, we have not used data from these mice. However, for completeness, we have included data from non-learners in key figures. Cells in hippocampal area CA1 imaged from awake, behaving mice exhibit temporal tuning We surgically exposed the left, dorsal hippocampus of naïve mice and bolus loaded a synthetic calcium indicator dye. We then implanted cranial windows through which we imaged calcium responses from cells in area CA1 of the hippocampus (Figure 2A, Figure 2—figure supplement 1A, Video 2 shows calcium responses from a sample field of view). Image acquisition was carried out within a field of view covering 96 ± 29 (mean ± SD) cells per mouse, imaged at frame rates ranging from 11 Hz to 16 Hz. In parallel with calcium imaging, we simultaneously measured eyeblink responses of mice over the entire pre-training and training sections of the conditioning protocol (n = 14 trace conditioned and six pseudo-conditioned mice; 'Materials and methods-Awake, two photon calcium imaging of area CA1 cells'; Figure 2—figure supplement 1B). Cell bodies were imaged from the visually identified stratum pyramidale layer of the hippocampus (Figure 2—figure supplement 1C), at depths ranging between 135–150 µm below the hippocampal surface (supporting Video 2 shows imaged calcium responses and Video 3 shows a stack of optical, z-section images with the densely labeled cell-body layer visible). To ensure the reliability of our calcium fluorescence data, we carried out two checks. First, the frequency of calcium transients was observed to be 1.3 ± 0.2 Hz (mean ± SD). While this need not accurately reflect spike rates in our experimental system, it falls well within the range of spontaneous spike rates previously observed in CA1 cells, and was un-changed between early and late trials in the session (Figure 2—figure supplement 1D; Czurkó et al., 1999). Second, the summed area under the calcium curve was also calculated for all datasets, and those datasets that showed a significant drift between early and late trials were discarded (n = 1 of 21; Figure 2—figure supplement 1E). Figure 2 with 1 supplement see all Download asset Open asset Two-photon imaging of calcium-responses in area CA1 neurons from awake mice. (A) Schematic of the imaging preparation. o–objective lens, s–skull, cs–cover slip, hb–head bar, ag–agarose. (B) Histogram of neuron response widths in ms, calculated as the time for which a given neuron's trial-averaged, ΔF/F trace remains above 50% of the peak value. The red, dotted line indicates a response width of 600 ms, which is the time of interest between tone-onset and puff-onset. (C) Area CA1 cell responses show sequentially timed activity peaks after learning. Calcium response (ΔF/F) traces for six exemplar neurons from a single mouse, for sets of three trials before (panel on the left), and after (panel on the right) task learning. Neurons have been sorted as per the timing of the peak in the averaged trace. The yellow and red bars at the bottom represent the times of delivery of tone and puff respectively. The gray shading to the left covers the period of spontaneous activity prior to the onset of the tone. The red asterisks indicate the peak in each individual trace. Scale bars indicate 0.4 ΔF/F and 500 ms along the time axis. (D) Area CA1 cell activity peaks tile the entire CS-on to US-off interval. Area CA1 calcium response traces from an example dataset, sorted by the peak times of the responses. Each response trace has been averaged over all trials following the learning trial (Figure 1G), and has been normalized to the peak ΔF/F response value for each neuron. The yellow and red bars below indicate times of delivery of tone and air-puff respectively. 50% of the neurons from the field of view, with the most reliably timed responses have been shown. This is to make this plot comparable to the ones from subsequent analyses, where neurons have been similarly chosen. (E and F) Cell activity peak timings change during learning. In E, pseudo-colored ΔF/F traces for the period of interest during and after tone delivery, are plotted using data acquired during the pre-training session, where tones without air-puff were delivered. Cells have been sorted as per the timings of peaks in pre-training session data. The yellow bar at the bottom indicates time of delivery of tone (350 ms). In F, the same averaged activity traces as in E have been re-ordered according to each cell's activity peak timing after learning has occurred, as shown in (D) Plotted in Figure 2—figure supplement 1, are panels depicting the surgical preparation, the numbers of mice from each treatment group, images of dye-loaded tissue taken at multiple depths and basic data quality control analyses. https://doi.org/10.7554/eLife.01982.005 Video 2 Download asset This video cannot be played in place because your browser does support HTML5 video. You may still download the video for offline viewing. Download as MPEG-4 Download as WebM Download as Ogg Calcium responses. Video showing a time-series of images of a single field of view of area CA1 cells. Brighter colors in the gray-scale indicate higher fluorescence intensities. Flashes visible are calcium responses. https://doi.org/10.7554/eLife.01982.007 Video 3 Download asset This video cannot be played in place because your browser does support HTML5 video. You may still download the video for offline viewing. Download as MPEG-4 Download as WebM Download as Ogg z-stack of dye-loaded hippocampal tissue. Video showing optical sections through a typical, dye-loaded imaging preparation of the dorsal hippocampus. Optical sections were acquired beginning at the dorsal hippocampal surface and moving ventrally in steps of 2 µm/frame. The scale bar represents 50 µm. The densely-packed cell-bodies of the curved, stratum pyramidale cell-body layer appear near the 7 s time-point in the video. The frames of this video show some motion—they were not motion corrected as each frame was taken at a different depth and thus, differed from the others. https://doi.org/10.7554/eLife.01982.008 We first looked for evidence of sustained or sequentially timed neuronal activity in calcium fluorescence traces for individual cells (ΔF/F,'Materials and methods: Data analysis'), on trials following the CR learning peak. First, we determined whether or not cells showed sustained activity through the entire period of interest by measuring how wide each cell's response was at half the amplitude of the peak of the response. The mean response width measured in this manner was found to be 120 ms ± 81 ms (mean ± SD, n = 542 cells, Figure 2B). Only three cells (∼0.5%, of all cells) had response widths greater than 600 ms, the length of the period from tone-onset to puff-onset. Based on this analysis, we concluded that sustained cell activation is not the mechanism used by the hippocampus to maintain a stimulus representation. We next looked for indications of sequential activation. Prior to learning, calcium response peaks were not reliably timed across trials, relative to tone onset (Figure 2C, left panel). Post learning, however, we observed that area CA1 cells had activity peaks at fixed time-points relative to tone onset, seen consistently across multiple trials (Figure 2C, right panel). We further characterized this by averaging these ΔF/F activity traces over trials and identifying activity peaks within the time window from tone onset to 200 milliseconds after puff onset. We rank ordered cells based on the timings of their activity peaks and found that small groups of cells firing at each time-point, tiled the entire interval of interest (Figure 2D). When the same procedure (averaging followed by sorting by timing) was carried out on traces from the pre-training dataset, tiling was markedly skewed with most cells (66% in this case) showing activity peaks during the tone period (Figure 2E). Furthermore, if neurons in the pre-training dataset were ordered as per timing of peaks in the training session, no clear tiling was visible (Figure 2F). This indicated that timings of peak activity of area CA1 cells, relative to the onset of tone stimulus, changed during training. Furthermore, at the population level, these peak times appeared to tile the interval of interest between tone and puff. After learning, area CA1 cells show reliably timed, sequential calcium-responses Having observed that area CA1 cells were active at fixed time-points relative to tone onset, we next wanted to quantify the reliability with which cells fired at these times. For each cell, we defined the timing of the peak in its averaged ΔF/F trace as its peak response time (PT). We then quantified the reliability with which cells fired at their respective PTs, from trial to trial (Figure 3A shows a cell reliably firing at a particular PT). First, we reasoned that if cells show time-aligned activity, the peak of the trial-averaged ΔF/F trace would be higher than if activity were not reliably timed. Hence, we computed the peaks of the averaged ΔF/F traces during the tone-onset to puff-onset period (Figure 3—figure supplement 1A). The mean peak amplitude for cells from trace conditioned mice was significantly higher than for cells from pseudo-conditioned mice or for spontaneous activity data (trace = 0.022 ± 0.007, pseudo = 0.017 ± 0.009, spontaneous = 0.015 ± 0.008; mean ± SEM; one way Analysis of Variance (ANOVA), followed by Tukey Kramer honest significant difference (h.s.d.) p<0.01). The difference seen was a small one, but this measure does not control for differences in cell response sizes or average activity. Hence, to more rigorously characterize activity timing reliability, we computed a reliability score for each cell. Figure 3 with 2 supplements see all Download asset Open asset After training, area CA1 cells show reliably timed, sequential calcium responses. (A) Calcium response traces for an example neuron, aligned to the time of stimulus delivery (tone and puff indicated by yellow and red bars at bottom respectively). Warmer colors in the traces indicate higher ΔF/F values. The blue rectangle on the 'trial number' axis indicates the trial averaging window comprising all trials after the learning trial. (B) The same data as in A, except with each trial's ΔF/F trace given a random time offset. (C) Averaged calcium response curves obtained from aligned, as well as random time-offset traces. The averaged curve from time-aligned traces (blue curve) has a prominent peak, which is absent in the random time offsets case (red curve). This indicates that the neuron fires reliably at a fixed time relative to stimulus delivery. The area under the shaded region (peak ± 1 frame) was used to calculate the reliability score. (D) Area CA1 cells from trace learners show significantly higher activity-timing reliability scores. Average reliability scores for all neurons in the entire dataset for learners of the trace-conditioning task (blue), pseudo-conditioned mice (green) spontaneous activity data (red) and data from non-learners (cyan; * indicates p<0.01). (E) Change in reliability score with learning for neurons from trace conditioned (blue), pseudo-conditioned (green) and spontaneous activity data (red) respectively. Reliability of firing at the final peak response time (PT) gradually increased over the training session. Reliability scores were computed in five-trial bins. The shaded regions represent SEM. (F) Change in reliability scores between early and late blocks of training trials. The increase over the training session for trace-learners was significantly higher than for spontaneous activity data or for pseudo-conditioned mice (* indicates p<0.01). This increase was calculated by subtracting reliability scores of the first half of the session from those of the second half. (G and H) Distributions of single-cell peak response times (PT) at different stages of learning: pre-training (G) and early in
Abstract Silicon photomultipliers (SiPMs) are a class of inexpensive and robust single-pixel detectors with applications similar to photomultiplier tubes (PMTs). We performed side-by-side comparisons of recently-developed SiPMs and a GaAsP PMT for two-photon fluorescence imaging of neural activity. Despite higher dark counts, which limit their performance at low photon rates (<1μs), SiPMs matched the signal-to-noise ratio of the GaAsP PMT at photon rates encountered in typical calcium imaging experiments due to their much lower pulse height variability. At higher photon rates and dynamic ranges encountered during high-speed two-photon voltage imaging, SiPMs significantly outperformed the GaAsP PMT.
Sparse coding can improve discrimination of sensory stimuli by reducing overlap between their representations. Two factors, however, can offset sparse coding’s benefits: similar sensory stimuli have significant overlap and responses vary across trials. To elucidate the effects of these 2 factors, we analyzed odor responses in the fly and mouse olfactory regions implicated in learning and discrimination—the mushroom body (MB) and the piriform cortex (PCx). We found that neuronal responses fall along a continuum from extremely reliable across trials to extremely variable or stochastic. Computationally, we show that the observed variability arises from noise within central circuits rather than sensory noise. We propose this coding scheme to be advantageous for coarse- and fine-odor discrimination. More reliable cells enable quick discrimination between dissimilar odors. For similar odors, however, these cells overlap and do not provide distinguishing information. By contrast, more unreliable cells are decorrelated for similar odors, providing distinguishing information, though these benefits only accrue with extended training with more trials. Overall, we have uncovered a conserved, stochastic coding scheme in vertebrates and invertebrates, and we identify a candidate mechanism, based on variability in a winner-take-all (WTA) inhibitory circuit, that improves discrimination with training.
Abstract Memory guides behavior across widely varying environments and must therefore be both sufficiently specific and general. A memory too specific will be useless in even a slightly different environment, while an overly general memory may lead to suboptimal choices. Animals successfully learn to both distinguish between very similar stimuli and generalize across cues. Rather than forming memories that strike a balance between specificity and generality, Drosophila can flexibly categorize a given stimulus into different groups depending on the options available. We asked how this flexibility manifests itself in the well-characterized learning and memory pathways of the fruit fly. We show that flexible categorization in neuronal activity as well as behavior depends on the order and identity of the perceived stimuli. Our results identify the neural correlates of flexible stimulus-categorization in the fruit fly. Impact Statement Flies can optimally recall a memory with high specificity by comparing options close in time, or default to generalization when they cannot.
'%3e%3ccircle r='20' fill='%23008'/%3e%3ccircle r='17.5' fill='%23fff'/%3e%3ccircle r='3.5' fill='%23008'/%3e%3cg id='d'%3e%3cg id='c'%3e%3cg id='b'%3e%3cg id='a' fill='%23008'%3e%3ccircle r='.875' transform='rotate(7.5 -8.75 133.5)'/%3e%3cpath d='M0 17.5.6 7 0 2l-.6 5L0 17.5z'/%3e%3c/g%3e%3cuse xlink:href='%23a' transform='rotate(15)'/%3e%3c/g%3e%3cuse xlink:href='%23b' transform='rotate(30)'/%3e%3c/g%3e%3cuse xlink:href='%23c' transform='rotate(60)'/%3e%3c/g%3e%3cuse xlink:href='%23d' transform='rotate(120)'/%3e%3cuse xlink:href='%23d' transform='rotate(-120)'/%3e%3c/g%3e%3c/svg%3e)
