Introduction
This blog contains various random terms we encounter in our research activities. Some of the terms may need extensive elaboration to explain properly. I will may be refer them on my other section like trending topics or some terms.
Latency machine learning
Nice Introduction and the original source and finally the impact of different values
In short, its the time required by the model to infer from one data sample. It can be think of as a response time for batch size of 1. Obviously the smaller the better means faster response. There is another terms related to the latency which is Throughput [look at the unit image per seconds for CV application].
Well it may seem that throughput is directly inverse to the latency, which is true up to inversely proportional. Just remember that sometimes operation can be shared and paralleled which will increase the throughput not the latency. We can measure the throughput by considering the memory and parallelism more.
Stop Gradient
Gradient Reversal
Poisson distribution
good enough source Discrete probability. Its parameterized by lambda parameter (Mean value). The PMF (source) <img src = https://wikimedia.org/api/rest_v1/media/math/render/svg/c22cb4461e100a6db5f815de1f44b1747f160048>
This means given the number of event lambda (for the considered time), the probability of k events in that time.
Generalized by event rate r and time frame t, probability of k events in the t time is defined by poisson’s distribution
Now just expand by simple math {summation} that how many event’s probability we want. Add the probability for each of the considered events.
Well, separate this from the exponential distribution. This is continuous analogous to the geometric distribution [first time something occurred]. good enough source. Also can be though of the distribution of the time interval between the events in the poisson arrival process. This is continuous variable (x) distribution so we get PDF [caution about actual probability, needs integration over a range] as following
simply, distribution of the exact arrival time (the random variable itself) from given last known time position. [this lambda is not the same as in the earlier equation]
The whole things make sense under integration by asking the probability of different events. Mind the event and inter-event time distribution.