Research area: Uncertainty quantification?

[kimstatistics]

Hi, I'm 1st year phd student in stat/biostat.

Since I don't have enough knowledge about fields in statistics, I would like to ask you some advice for choosing my research interests even though this can be silly question.

I don't have a specific interests until yet, but I hope it can be applied in various fields and have enough positions in academia.

High-dimensional data analysis(including functional data analysis) and uncertainty quantification(using gaussian process) come to my mind among available options.

High-dimensional study is very popular now and many statisticians do this. I guess this is in high demand but also common and more competitive. 

On the other hand, uncertainty quantification is pretty unfamiliar to me. I think it lies in Statistics and Industrial engineering and this can be advantage.

However, it is not easy to find TT faculty in this field rather than high-dimensional study. In addition, I've heard time series is on the decline recently.

Is it really on the decline? and Does it have many job opportunity especially in academia?

I know all of these are depends on my publications, but I want to know general trend.

It will be better option if there is something combined with these two fields, but I cannot imagine anything yet.

I really appreciate your advice or answers in advance!

[Stat Assistant Professor]

I think it depends more on what field you're applying uncertainty quantification to rather than broadly "uncertainty quantification." I have seen some work published in top journals related to uncertainty quantification for deep learning and causal inference. A big research area of Bayesian nonparametrics right now is studying the coverage properties of posterior credible sets (i.e. under what conditions a credible set is also a confidence set that gives the same asymptotic coverage). 

It seems like inference and uncertainty quantification are in general always seen as more important by statisticians than people in Computer Science/machine learning departments. That's why there are a lot more articles in statistics journals than machine learning conferences related to post-selection inference, construction of simultaneous confidence intervals, etc.

Getting hired in academia (at research universities) is highly dependent on your publication record and your letters of recommendations. The former depends a lot on the novelty of your work. It is possible for somebody whose work is on time series to get hired, but their contributions to the area have to be novel. FWIW, it seems like multivariate time series (such as vector autoregressive models) is still a fairly "hot" topic, and I know of somebody who was hired as an Assistant Professor at Cornell Statistics because of an Annals of Statistics paper that was a significant contribution to regularized VAR models. 

[bayessays]

As @Stat Assistant Professor says, broadly quantifying uncertainty is certainly of interest to statisticians (do statisticians do much else?) but the field of "uncertainty quantification" is something I don't see described as a research area for actual statisticians, but more in some type of niche engineering fields.  I think if you told 100 statisticians that your field of research was "uncertainty quantification", 98% would be confused as to what you actually study.

[kimstatistics]

On 1/4/2021 at 5:44 PM, bayessays said:

As @Stat Assistant Professor says, broadly quantifying uncertainty is certainly of interest to statisticians (do statisticians do much else?) but the field of "uncertainty quantification" is something I don't see described as a research area for actual statisticians, but more in some type of niche engineering fields.  I think if you told 100 statisticians that your field of research was "uncertainty quantification", 98% would be confused as to what you actually study.

 

On 1/4/2021 at 4:41 PM, Stat Assistant Professor said:

I think it depends more on what field you're applying uncertainty quantification to rather than broadly "uncertainty quantification." I have seen some work published in top journals related to uncertainty quantification for deep learning and causal inference. A big research area of Bayesian nonparametrics right now is studying the coverage properties of posterior credible sets (i.e. under what conditions a credible set is also a confidence set that gives the same asymptotic coverage). 

It seems like inference and uncertainty quantification are in general always seen as more important by statisticians than people in Computer Science/machine learning departments. That's why there are a lot more articles in statistics journals than machine learning conferences related to post-selection inference, construction of simultaneous confidence intervals, etc.

Getting hired in academia (at research universities) is highly dependent on your publication record and your letters of recommendations. The former depends a lot on the novelty of your work. It is possible for somebody whose work is on time series to get hired, but their contributions to the area have to be novel. FWIW, it seems like multivariate time series (such as vector autoregressive models) is still a fairly "hot" topic, and I know of somebody who was hired as an Assistant Professor at Cornell Statistics because of an Annals of Statistics paper that was a significant contribution to regularized VAR models. 

Thank you for your answers. I'm sorry saying too broad and unclear. It is uncertainty quantification using computer experiment. What I concern is this field is too small and has less job opportunity or not, it may not be true though. Also I cannot get this field is big or increasing. Someone said this field is not super popular in this big data generation but have become increasingly popular in the past two decades. I would really appreciate any advice.

[ZNtheory]

I once was supervised by some researchers at UCL on the topic of Gaussian Processes and its application on uncertainty quantification. I think many researchers at least at UK institute has done research in relation to UQ as well as its application in machine learning though many of them hold their position in Stat Department. As for HD statistics, I think it has been popular since the sparse models have been widely used and developed. More recently, HD statistics have been used to more broad aspects like analyzing the generalization of overparameterized models. Therefore, I think whether research on either HD statistics or UQ should be fine if you want to secure a job at academia. Just personal thought. :)