We investigate the discrimination capacity of a single neuron model including the integrate-and-fire (IF) model and the IF-FHN model Neural Networks, 14 (2001) 955, both theoretically and ...numerically. Both magnitude and frequency discrimination tasks are considered. It is found that for the magnitude discrimination task, with a small fraction of coherent inputs, an IF model and an IF-FHN model with inhibitory inputs can tell one input from the other. The total probability of misclassifications (TPM) is considerably reduced with the increasing of inhibitory inputs. However, the IF model and the IF-FHN model are incapable of performing the frequency discrimination task.
It is generally believed that the support vector machine (SVM) optimizes the generalization error and outperforms other learning machines. We show analytically, by concrete examples in the one ...dimensional case, that the SVM does improve the mean and standard deviation of the generalization error by a constant factor, compared to the worst learning machine. Our approach is in terms of the extreme value theory and both the mean and variance of the generalization errors are calculated exactly for all the cases considered. We propose a new version of the SVM , called the scaled SVM, which can further reduce the mean of the generalization error of the SVM.
From an observation of efferent interspike intervals of a neuron, we consider how to decode the input temporal information. It is found that the integrate-and-fire model is blind in the temporal ...domain due to the fact that its efferent firing rate is independent of the input temporal frequency. The conclusion is then confirmed for the integrate-and-fire model with correlated inputs, with reversal potentials, with a nonlinear leakage and with a subthreshold oscillation. For the Hodgkin-Huxley model, however, in terms of efferent firing rates alone, it is possible to read out the input temporal information.
We consider how correlated inputs can ensure a group of neurons (the integrate-and-fire model or the Hodgkin–Huxley model) to synchronize their activities. Without interactions, the shortest ...synchronization time occurs when the coefficient of variation of efferent spike trains of single neuron is greater than 0.5. The application of synchronization driven by correlated inputs to signal detection is also taken into account.
We find that adding certain amounts of inhibitory inputs to a neuron improves its capability of accurately decoding the input information. The optimal ratio
r of inhibitory to excitatory inputs for ...decoding the input information from an observation of the efferent interspike intervals is calculated. Surprisingly, the Fisher information could be zero for certain values of the ratio, seemingly implying that it is impossible to read out the encoded information at these values. By analyzing the maximum likelihood estimate of the input information, it is then concluded that the input information is, in fact, most easily estimated at the points where the Fisher information vanishes.
Most recent semantic segmentation methods adopt a fully-convolutional network (FCN) with an encoder-decoder architecture. The encoder progressively reduces the spatial resolution and learns more ...abstract/semantic visual concepts with larger receptive fields. Since context modeling is critical for segmentation, the latest efforts have been focused on increasing the receptive field, through either dilated/atrous convolutions or inserting attention modules. However, the encoder-decoder based FCN architecture remains unchanged. In this paper, we aim to provide an alternative perspective by treating semantic segmentation as a sequence-to-sequence prediction task. Specifically, we deploy a pure transformer (ie, without convolution and resolution reduction) to encode an image as a sequence of patches. With the global context modeled in every layer of the transformer, this encoder can be combined with a simple decoder to provide a powerful segmentation model, termed SEgmentation TRansformer (SETR). Extensive experiments show that SETR achieves new state of the art on ADE20K (50.28% mIoU), Pascal Context (55.83% mIoU) and competitive results on Cityscapes. Particularly, we achieve the first position in the highly competitive ADE20K test server leaderboard on the day of submission.
We consider how the output of the perfect integrate-and-fire (I&F) model of a single neuron is affected by the properties of the input, first of all by the distribution of afferent excitatory and ...inhibitory postsynaptic potential (EPSP, IPSP) inter-arrival times, discriminating particularly between short- and long-tailed forms, and by the degree of balance between excitation and inhibition (as measured by the ratio, r, between the numbers of inhibitory and excitatory inputs). We find that the coefficient of variation (CV; standard deviation divided by mean) of efferent interspike interval (ISI) is an increasing function of the length of the tail of the distribution of EPSP inter-arrival times and the ratio r. There is a range of values of r in which the CV of output ISIs is between 0.5 and 1. Too tight a balance between EPSPs and IPSPs will cause the model to produce a CV outside the interval considered to correspond to the physiological range. Going to the extreme, an exact balance between EPSPs and IPSPs as considered in 24 ensures a long-tailed ISI output distribution for which the moments such as mean and variance cannot be defined. In this case it is meaningless to consider quantities like output jitter, CV, etc. of the efferent ISIs. The longer the tail of the input inter-arrival time distribution, the less is the requirement for balance between EPSPs and IPSPs in order to evoke output spike trains with a CV between 0.5 and 1. For a given short-tailed input distribution, the range of values of r in which the CV of efferent ISIs is between 0.5 and 1 is almost completely inside the range in which output jitter (standard deviation of efferent ISI) is greater than input jitter. Only when the CV is smaller than 0.5 or the input distribution is a long-tailed one is output less than input jitter 21. The I&F model tends to enlarge low input jitter and reduce high input jitter. We also provide a novel theoretical framework, based upon extreme value theory in statistics, for estimating output jitter, CV and mean firing time.