Detection, classification, and superposition resolution of action potentials in multiunit single-channel recordings by an on-line real-time neural network
read more
Citations
A review of methods for spike sorting: the detection and classification of neural action potentials.
Online detection and sorting of extracellularly recorded action potentials in human medial temporal lobe recordings, in vivo☆
Evaluation of spike-detection algorithms fora brain-machine interface application
Neural spike sorting under nearly 0-dB signal-to-noise ratio using nonlinear energy operator and artificial neural-network classifier
A 100-channel system for real time detection and storage of extracellular spike waveforms.
References
Learning representations by back-propagating errors
Sensitivity of MST neurons to optic flow stimuli. I. A continuum of response selectivity to large-field stimuli.
Digital filters and signal processing
Multispike train analysis
Bayesian modeling and classification of neural signals
Related Papers (5)
A review of methods for spike sorting: the detection and classification of neural action potentials.
Neural spike sorting under nearly 0-dB signal-to-noise ratio using nonlinear energy operator and artificial neural-network classifier
Unsupervised spike detection and sorting with wavelets and superparamagnetic clustering
Frequently Asked Questions (18)
Q2. What are the future works in "Detection, classification, and superposition resolution of action potentials in multiunit single-channel recordings by an on-line real-time neural network" ?
Future work will be needed to find ways for the NN to track amplitude changes, and to adapt to slowly changing template shapes caused by electrode movement. Since this is a simultaneous technique, each detected waveform is a potential initiator waveform for a cluster.
Q3. How did the authors test the network at higher sampling rates?
The authors also tested the system at higher sampling rates to determine if a sampling rate of 24 KHz is sufficient to overcome the problem of time quantization.
Q4. Why are there no false positives for the matched filter?
The reason there are no false positives for the matched filter when the scaling factor is 1.0 (one mightexpect a smaller spike to be identified as a larger template) is because as the vectors in multidimensional space are extended, the distance between them increases.
Q5. What were the changes that were incorporated in the training?
Modifications that were incorpo-rated in the training were an adaptive learning rate, weight momentum, weight decay, and weight annealing.
Q6. What is the effect of asynchronous sampling?
If the sampling rate and buffer size are simultaneously increased, the effect of asynchronous sampling would be reduced and classification performance should increase.
Q7. What is the advantage of using an ANN?
With the use of a high-speed digital signal processor and an IBM-PC, an automated real-time system based on the NN was implemented without the need for off-line post-processing.
Q8. What is the advantage of a distributed processing algorithm?
With distributed processing, the algorithm can be readily scaled to identify more units from one electrode or to process data from multiple electrodes.
Q9. How was the coefficient of the recursive filter determined?
The coefficients of the recursive filter were determined by fitting a single exponential to the decay of the autocorrelation function [41].
Q10. What was the effectiveness of the NN in classifying single units in a multiunit recording?
The effectiveness of the NN in classifying single units in a multiunit recording was comparable to the MTF and importantly, the network also resolved most superpositions with overlap less than 0.5 ms.
Q11. What is the threshold for each waveform?
3) Each waveform initiates a cluster containing other waveforms that satisfyfor and, thresholdwhere is the Euclidean distance between vectors and and the threshold is determined from a purenoise segment.
Q12. What is the way to train a network to respond to a spike?
In the case of superpositions that are samples apart, ideally the network would respond to the first spike and then respond again n samples later to the second spike.
Q13. How does the NN perform when no superpositions are present?
When no superpositions are present [Fig. 5(b)], both discriminators attain at least 80% correct classification at SNR’s of 5.0 and above.
Q14. What is the step of the algorithm?
The steps of the algorithm are as follows.1) All waveforms whose peaks exceed three standard deviations of noise (amplitude detection) are located and are centered about their point of maximum slope in a window of 24 samples.
Q15. What is the reason why the performance of the network dropped?
The observed performance dropped marginally because the network would occasionally classify the event one sample before or after the actual time of the event (an error of 1/24th of a ms).
Q16. How many hidden units were used for the NN?
The maximum size of the network that could be implemented with a 24- sample input buffer and a 24-KHz sampling rate was eight hidden units and three output units.
Q17. How can the NN overcome the alignment problem?
This alignment problem can be overcome by sampling at a higher rate, or by interpolating the waveform (provided the sampling rate meets the Nyquist criterion).
Q18. What is the difference between the digitized waveform and the MTF?
The digitized waveform of a spike includes a variable offset (between 0.5 sampling interval) because the spikes occur randomly with respect to the periodic sampling, i.e., they are not phase locked to the A/D conversion of the data.