A classic example of this phenomenon is the face cell in the inferotemporal cortex of the primate brain. Moreover, the behavioral response is similar regardless of the position of the nearby stimulus in the upper visual field.
Electrode pooling
One recent report showed that the escape probability varies with the contrast of the threatening stimulus [9]. Placing the electrode closer to the neuron also helps, because most of the voltage drop occurs right at the cell body where the capacitive currents of the dendrites accumulate.
SIFTING OF VISUAL INFORMATION IN THE SUPERIOR COLLICULUS
Abstract
Introduction
Effectively, the alarm circuits in the mouse's visual system extract from the overall visual display only one or two bits of information needed to initiate action. In general, neural processing in the SC identifies salient points in the environment and coordinates the animal's orientation toward or away from such locations.
Results
Indeed, compared to superficial SC, neurons in the deep SC became more selective for the expanding dark disc (Figure 2.2D). Neurons in the superficial layers typically produced a peak burst at a similar firing rate with each repetition of the stimulus (Figure 2.1C).
Discussion SummarySummary
The present study focused on stimuli presented in the upper visual field and recordings made from the corresponding medial region of the SC. This contrasts with our observations of neurons in the superficial SC showing no site-specific habits (Figure 2.4B).
Figures
Supplementary Figure S2.2: Distribution of the threat selectivity index for neurons in the superficial and deep SC versus population summary of the variability in the timing of the first peak from the experiment in (A). The response of an example dSC neuron to threatening stimuli presented at the beginning (left) and end (right) of a recording.
Other stimuli (e.g., chessboard) were presented to the animal in the ~50 minutes separating these two blocks. This suggests that the strong response at the initial presentation of the threatening stimulus is not simply a result of motor output or change in arousal level in the absence of visual threat. Linear decoders were trained with simultaneously registered sSC and dSC neurons to predict the location (left) and novelty (that is, whether the stimulus appeared at a location for the first time) (right) of stimuli in the experiment being conducted. described in figure 2.3.
Supplementary Figure S2.6: One of several putative local detectors identified in the superficial SC (A), with a local receptive field (B) and selectivity for the nearby stimulus without significant habituation to repeated stimuli (C).
Methods
The moving dark disk moved at ∼40–70 °/s, with the reaction zone at the center of the trajectory. We defined the latency as the timing of the first spike during the stimulus period. In the case of the location decoder, the labels were multi-class and ranged from 1 to 25 (one for each stimulus location).
In the case of the novelty decoder, the labels were binary (stimuli that were novel, i.e. . . . first to appear at a location were 1; others were 0). A circuit for the detection of interaural time differences in the brainstem of the spotted owl. Journal of Neuroscience. Responsiveness to sensory stimulation of units in the superior colliculus and inferior tectotegmental regions of the rabbit.
Neurons in the most superficial lamina of the mouse Superior Colliculus are highly selective for the direction of the stimulus.
ELECTRODE POOLING
Abstract
Introduction
Clearly, we need a way to increase the number of neurons recorded while avoiding a concomitant increase in the number of wires entering the brain. At the other end of the line, a synchronized switch or sampling system can demultiplex the signals again. The cycle rate of the switch is limited by the sampling theorem [29]: it must be at least twice the highest frequency component present in the signal.
Therefore, an essential element of any such multiplexing scheme is an analog low-pass filter associated with each electrode. One obstacle is the physical size of the anti-aliasing filter associated with each electrode. In that case, aliasing of high-frequency thermal fluctuations will increase the noise power in the recording by a factor equal to the number of electrodes 𝑁 being multiplexed.
This way, the low-pass filtering and amplification can take place at the other end of the wire, outside the brain, where the base of the shaft expands to provide virtually unlimited silicon space.
Results
An action potential that appears on only one of the 𝑀 electrodes is therefore attenuated by a factor of 𝑀1 in the combined signal. This can be achieved by using the split-mode recordings from the early sampling phase of the experiment. We then matched each unit in pooled mode with the split-mode unit that had the most similar waveform, based on the cosine similarity of their waveform vectors (Figure 3.6B-C).
This suggests that one can jump start the sorting of the total signal by incorporating the prior knowledge from the sorting of the split-mode recordings. The total signal is a weighted average of the signals at the two sites, with the pooling coefficients determined by the relative impedances (Eqn3.2) and summed to 1. At low concentrations, the noise actually increased dramatically (Figure 3.8A), and this reflects the thermal noise 𝑁ele.
Figure 3.8F shows a distribution of the biological RMS noise, which had different amplitudes on the two different banks (approximately 3.8 mm apart) and exceeded both the thermal and electronic noise.
Discussion Summary of resultsSummary of results
In the current version of the Neuropixels device [12], the ratio of electrodes to wires is only 2.5, so there is little practical benefit to be gained from joining electrodes together. In today's Neuropixels, each electrode has only one associated switch, and thus only one candidate wire. This is a simple CMOS circuit that changes the sign of the waveform [3] depending on a local switch setting.
If half of the electrodes in a pool use the inverter, this helps distinguish the spike shapes of different neurons. We have implemented this method in KiloSort2 and shown that it can greatly increase the number of split-mode cells found in merged records (Figure 3.6). Indeed, knowing what waveforms to look for in the recording would help any point classifier.
Indeed, they are fully complementary: static switching exploits the redundancy in the neural signal across neighboring electrodes; time division multiplexing uses the fact that a wire has much higher bandwidth than a neuron; and electrode pooling exploits the sparsity of spiking neural signals on the time axis.
Figures
In Section 3.3 we outline some heuristic rules that one can follow, but turning this into an effective algorithm, making use of the full noise and impedance specifications of the device, remains an open problem. An equivalent circuit model for two electrodes connected to a common wire along with downstream components of the signal chain, such as the amplifier, multiplexer and digitizer. Recording locations (black squares, numbered from 1 to 4) in the same relative location of each bank can be combined into a single wire by closing the switches (yellow).
The number of cells from split-mode recordings found matches in the merged recording is plotted for three different sorting conditions: sorting all recordings by KiloSort1 followed by manual management (left), sorting all recordings by KiloSort2 (middle) and sorting the pooled recording by KiloSort2 with templates initialized based on the split recordings (right). The numbers next to the waveform at each recording location indicate the pooling coefficients, which are the amplitude ratio of the poled signal to the corresponding split signal. The electronic noise, defined as the noise level at 10X PBS in (A) at both Bank 0 and Bank 1. C) Prediction of the pooled noise in PBS based on noise measured at each bank prior to pooling.
The spike trains of sorted units are then matched to the ground truth and given an accuracy score.
Methods
In these cases, we assigned the average of these values as the pool coefficient of the survey site. From this we can check whether their pool coefficients add up to 1, as expected from Section 3.3 (Figure 3.7B). The distribution in Figure 3.7C only includes pool coefficients from these sites (50 pairs in Banks 0 and 1).
To determine the distribution of thermal noise due to 𝑅bat, the noise at 10X PBS was subtracted from the noise at 1X PBS in quadrature (Figure 3.8D). The estimated electronic and thermal noise from Figures 3.8B and D were subtracted from the noise at these electrodes to obtain the biological noise distribution (Figure 3.8F). Each of the three noise sources - 𝑁bio, 𝑁ele, 𝑁com - was simulated as Gaussian white noise.
According to field methods [ 4 , 18 ], the spike times of the classified units were matched to those of the ground truth units via a confusion matrix.
CONCLUDING REMARKS
Visual sifting in the SC
One line of investigation is a more in-depth analysis of the circuit described in Chapter 2. Given its access to the entire retinal output [1], the SC is well equipped to perform many other visual computations. The presence of a mutant mouse line that lacks the cerebral cortex developmentally [4] may be useful for dissecting SC function from reciprocally connected cortical areas.
It would also be interesting to examine how the SC processes more natural visual stimuli instead of the artificial ones that have been used to test its function so far. Finally, we could continue to monitor the neural processing of the threat reaction and ask: how the visual information screened by the SC is translated into behavior. This question is particularly interesting because the animal creates a memory of the location of the nest after just one visit [6].
The interplay between the SC and other brain regions such as the hippocampus would be interesting to study in this context.
Electrode pooling
How the animal incorporates prior knowledge of its environment (e.g. the presence of a nest) and makes this decision in a few tenths of a second is unknown. In the case of flight, how does the brain use the collicular output to plan and execute a trajectory back to the nest. Despite intense research, the neural basis of navigation remains largely unexplained, and few studies have investigated such rapid homing behavior in the mouse.