Do tiny neural spikes really matter? Uncovering the role of low‑amplitude spikes

In the past decade scientists have built algorithms that read brain signals without first sorting them into neuron groups. These methods, based on marked point‑process theory, treat all spikes—the brief electrical blips—as a single information stream. In this paper the authors asked how much each individual spike contributes to the brain's overall message. They compared two approaches: the traditional one that uses only sorted spikes (the easy‑to‑detect ones) and the new one that also includes low‑amplitude spikes, often discarded because they are hard to cluster. Analyzing rat hippocampus data—known for coding spatial position—they measured entropy reduction, i.e., how much uncertainty drops when a new spike is added. Looking at isolated spikes, high‑amplitude ones convey more information than weak spikes. Yet when all previous spikes are taken into account, the difference vanishes: low‑amplitude spikes, embedded in the context, transmit the same amount of data. This shows that, although low‑amplitude spikes seem useless alone, they become valuable when the brain combines them with its activity history. The “clusterless” approach thus not only simplifies analysis but also reveals how the brain exploits every tiny signal to map space.

Marked point processes are a class of statistical models describing discrete events (like spikes) associated with a label or "mark" (e.g., amplitude or waveform shape). In neurophysiology each spike is represented by a pair (t, m), where t is the occurrence time and m is the feature vector of the signal. The clusterless model bypasses spike sorting, treating the entire mark distribution as a continuous density in feature space. Entropy H(X) quantifies the average uncertainty of X; entropy reduction ΔH = H(prior) – H(posterior) measures how much information a new datum provides. In the study, the prior is the rat's position distribution given spikes observed up to time t‑1, while the posterior also incorporates the spike at time t. The ΔH difference between high‑ and low‑amplitude spikes is significant only when ΔH is computed for a single spike (prior = uniform distribution). When computing ΔH conditioned on the whole spike history, the conditional probability P(position|spike_1,…,spike_t) converges to the same distribution regardless of the new spike's amplitude, because cumulative information dominates. From an information‑theoretic perspective, low‑amplitude spikes have high mutual information I(S;X|H_{t-1}) when H_{t-1} is data‑rich, but low I(S;X) in isolation. Moreover, the analysis employs extended Kalman filters to estimate the spatial trajectory, integrating spike observations into a linear‑nonlinear state‑space model. Bayesian updating recursively refines the probability distribution, exploiting the Markovian structure of the rat's path. In summary, the work shows that the information each spike carries depends heavily on temporal context, and that clusterless methods combined with information‑theoretic metrics provide a fuller picture of neural coding.