Discussion of jonnew's comment: Virtually all previous methods for decoding ensemble spiking activity have required a pre-processing phase wherein spike waveforms are sorted according to a putative cell of origin. ‘Spike sorting’ is viewed as both a de-noising step ('true' spikes are separated from random threshold crossings due to noise) and a means to cleanly assign to decoding/encoding properties (e.g. receptive fields) to individual cells. However, there is no ground truth for the validity of any spike sorting procedure. For this reason, clusters may not correspond to single cells complicating the interpretation that calculated encoding and decoding properties belong to single neurons. Spike-sorting is not standardized, and has spawned a cottage industry of algorithm development. Additionally, spike sorting procedures are biased in that they typically exclude low amplitude, low frequency spike events that are potentially informative to a decoder. In response, Koosterman et al. have develop a method ('feature decoding') for decoding an external stimulus from spiking activity without a spike-sorting step. This work is a generalization of previous techniques that use Bayesian decoding on sorted units (Brown et al, 1998, Zhang et al 1998), which can be recovered simply by mapping 'features' to a unit ID rather than a physical waveform features, such as the peak amplitude. To test the technique, apply their decoding algorithm to spiking activity obtained from multi-tetrode hippocampal recordings of freely running rats. The authors run the decoding algorithm using physical spike features (peak tetrode amplitudes, 4 dimensional), unit IDs, and multiunit firing rates to decode the rat's location as it runs on a linear maze. They find comparable performance between unit-based sorting and feature-based decoding (in terms of mean position error). Multiunit firing rate-based decoding performs poorly compared to both. Interestingly the feature-based decoding performs much better than the unit-based decoding when operating only on the unsorted 'hash' or '0 unit', indicating that there is place-specific information in these messy spike waveforms (although both perform poorly compared to when well isolated spikes are used). Most importantly, this method provides a means for accurate ensemble neural decoding that is not preceded by a spike-sorting, which is fraught with assumptions and procedures that are inconsistent lab to lab and person to person.