TL;DR
This paper investigates whether the timing of a neuron's first spike (first-spike latency) can reliably encode information about a stimulus, even when the exact moment of stimulus onset isn't perfectly known. Surprisingly, they found that referencing this first-spike latency to a population-wide estimate of stimulus onset, rather than an assumed fixed point, actually increases the information conveyed by individual neurons. This suggests first-spike latency is a more robust neural code than previously thought.
What Problem Does This Paper Address?
The timing of neural responses, particularly the very first spike a neuron fires after a stimulus (first-spike latency), is known to vary with stimulus properties like sound location or visual contrast. This has led to the hypothesis that first-spike latency could be a crucial neural code. However, a major challenge for this idea is that measuring latency requires a precise reference point for stimulus onset. In biological systems, such an "independent" reference might not exist. If the brain has to estimate stimulus onset based on the collective activity of many neurons, and if these neurons' latencies are correlated, then measuring individual neuron latency relative to this estimated onset could reduce the information conveyed. This paper directly addresses this potential confound by testing whether first-spike latency information holds up when using a biologically plausible, population-derived estimate of stimulus onset.
Methods
The researchers developed a computational model to simulate how the brain might estimate stimulus onset. This "coincidence detector" model takes the combined inputs from a population of neurons. Each input spike is convolved with an exponential decay, mimicking an excitatory postsynaptic potential (EPSP). When the summed EPSP amplitude in the coincidence detector neuron crosses a threshold, it's taken as the estimated stimulus onset.
To measure information, they used the concept of mutual information (MI). They calculated MI between stimulus identity and first-spike latency in two ways:
- MImin: Using a fixed, assumed stimulus onset (5 ms after the actual stimulus).
- MIpop: Using the stimulus onset time derived from their coincidence detector model.
They also explored using a "best" constant onset time (MIbest) to see if simply shifting the reference point helped.
Datasets
The study used electrophysiological recordings from decerebrate cats responding to synthesized, broadband noise stimuli. These stimuli were designed to mimic sound localization cues like interaural time differences (ITDs) and interaural level differences (ILDs), as well as average binaural intensity (ABI). Two sets of 25 stimuli each were used, varying these cues independently. The data consisted of spike trains from 73 individual neurons in the central nucleus of the inferior colliculus (IC).
Key Insights and Ideas
- Population onset is informative: The core finding is that using a population-derived stimulus onset doesn't degrade, and often improves, the information encoded by first-spike latency. This challenges the intuition that correlated neural activity would obscure individual neuron signals.
- Latency is robust to reference uncertainty: This work demonstrates that first-spike latency can be a reliable code even when the precise stimulus onset is not explicitly known, relying instead on emergent population activity.
- Not all neurons benefit equally: While the average information increased, some neurons showed a decrease in information when referenced to the population onset. This highlights the heterogeneity in neural responses.
- Beyond the first burst: The paper notes that the timing of subsequent spike bursts can also contribute to information transfer, even if the initial burst's latency is less informative when referenced to the population onset.
Key Results
- On average, referencing first-spike latency to the population onset (MIpop) resulted in a slight increase in information compared to using a fixed 5 ms onset (MImin). The mean MIpop was 0.82 ± 0.43 bits, versus 0.70 ± 0.41 bits for MImin.
- A majority of neurons (53 out of 73) showed an increase in information when using the population-derived onset.
- Even when compared to the "best" constant onset time (MIbest), the population onset reference still provided more information for the majority of neurons.
- The study found that neurons often show a shift in their first spike latency but not in the timing of subsequent spikes, a phenomenon that has been observed previously.
Limitations and Future Work
- Noise correlations: The authors acknowledge that their study used non-simultaneous recordings, meaning they couldn't directly investigate "noise correlations" – trial-to-trial variations in first-spike latency that are correlated across neurons. They suggest that strong noise correlations could counteract their findings and reduce information. However, they argue that the precise timing of auditory onset neurons makes such massive noise correlations unlikely.
- Mechanism of population encoding: While they demonstrate the effect of population-based onset estimation, the precise neural circuits that implement this are not fully detailed. They speculate on potential mechanisms involving coincidence detection of inhibitory and excitatory inputs, similar to circuits involved in duration tuning.
- Future experiments: The authors suggest that future work should focus on exploring these potential neural circuits and investigating the role of noise correlations more directly with simultaneous recordings.
Summary updated