1

Topic: Criticism ML with NIPS

Yesterday on NIPS there was an interesting presentation with criticism ML. tl; dr; Today the ML is faster Alchemy than a science. We use any approximations of type of stochastic graded-index descent which cannot give the guaranteed result.  batch norm (as on-Russian?) which like accelerates but again  guarantees nothing. And generally nobody understands as this batch-norm works. For services everyone there with photos it , and for serious areas of type of medicine is "pipe". To look with 11 minutes. https://www.youtube.com/watch?v=Qi1Yry33TQE

2

Re: Criticism ML with NIPS

DP> Yesterday on NIPS there was an interesting presentation with criticism ML. DP> tl; dr; Today the ML is faster Alchemy than a science. We use any approximations of type of stochastic graded-index descent which cannot give the guaranteed result.  batch norm (as on-Russian?) which like accelerates but again  guarantees nothing. And generally nobody understands as this batch-norm works. For services everyone there with photos it , and for serious areas of type  is "pipe". In my opinion, this serious obstacle to advancement ML. Hardly probable, say, traffic control (air, ), will be entrusted the system which is a black box. And any other more or less responsible application.

3

Re: Criticism ML with NIPS

Hello, D. Petrov, you wrote: DP> tl; dr; Today the ML is faster Alchemy than a science. We use any approximations of type of stochastic graded-index descent which cannot give the guaranteed result.  batch norm (as on-Russian?) which like accelerates but again  guarantees nothing. And generally nobody understands as this batch-norm works. For services everyone there with photos it , and for serious areas of type  is "pipe". Much truth, but as a whole blackness of a box is explicitly exaggerated." Nobody understands as this batch-norm works "- is too strongly told. ML is a normal statistics, but applicable to the big arrays  the data. I suggest to consider a question: whether it is possible to trust statistics? Here all realize that cars and planes create from details, quality and which reliability not 100 %. But there is a reliability theory and norms of spoilage, cars and planes all the same do, by them go and fly, to them somehow trust.

4

Re: Criticism ML with NIPS

Hello, Nuzhny, you wrote: N> Much truth, but as a whole blackness of a box is explicitly exaggerated. "Nobody understands as this batch-norm works" - is too strongly told. ML is a normal statistics, but applicable to the big arrays  the data. I suggest to consider a question: whether it is possible to trust statistics? About that that and speech... If you do all "on a science" that and optimization will do by exact methods of type of reversal of matrixes. But where that on dimensionality 3 the order such  is lost and it is necessary to use approximate methods. Modern ML it also is ability to work with these  methods by their combination and . Here also touch with a science is lost and the alchemy begins. PS: I do not try to prove that ML - heresy. It is just necessary to remember such moments if seriously  this .

5

Re: Criticism ML with NIPS

Hello, TMU_1, you wrote: TMU> In my opinion, this serious obstacle to advancement ML. Hardly probable, say, traffic control (air, ), will be entrusted the system which is a black box. And any other more or less responsible application. Well so therefore also there was a direction under title Explainable AI.

6

Re: Criticism ML with NIPS

Hello, Nuzhny, you wrote: N> Much truth, but as a whole blackness of a box is explicitly exaggerated. "Nobody understands as this batch-norm works" - is too strongly told. ML is a normal statistics, but applicable to the big arrays  the data. I suggest to consider a question: whether it is possible to trust statistics? On mine, for serious disciplines we cannot trust statistics. Be more exact at us the adequate model which describes the validity should. Here we can check this model both on statistican, and on the artificial data. Should check correct operation of model in extreme points, etc. Differently each new case which is not coming across in statistican, will lead to failure.

7

Re: Criticism ML with NIPS

Hello, D. Petrov, you wrote: DP> About that that and speech... If you do all "on a science" that and optimization will do by exact methods of type of reversal of matrixes. But where that on dimensionality 3 the order such  is lost and it is necessary to use approximate methods. Modern ML it also is ability to work with these  methods by their combination and . Here also touch with a science is lost and the alchemy begins. Actually it is not important exact or optimization if it is used is not precisely fulfilled the model is incorrect. And the schedule on last points also will be in the majority incorrect model.

8

Re: Criticism ML with NIPS

Hello, Nuzhny, you wrote: N> Much truth, but as a whole blackness of a box is explicitly exaggerated. "Nobody understands as this batch-norm works" - is too strongly told. ML is a normal statistics, but applicable to the big arrays  the data. I suggest to consider a question: whether it is possible to trust statistics? N> here all realize that cars and planes create from details, quality and which reliability not 100 %. But there is a reliability theory and norms of spoilage, cars and planes all the same do, by them go and fly, to them somehow trust. Here other problem - quality and reliability are reached for the account "", i.e. it is iterated converging process. , on how many I know, it is impossible to debug - why  this or that  etc.

9

Re: Criticism ML with NIPS

Hello, kl, you wrote: TMU>> In my opinion, this serious obstacle to advancement ML. Hardly probable, say, traffic control (air, ), will be entrusted the system which is a black box. And any other more or less responsible application. kl> well so therefore also there was a direction under title Explainable AI. Irony that not-explainable tenserflow now on two orders more popular direction.

10

Re: Criticism ML with NIPS

Hello, Sharov, you wrote: S> Here other problem - quality and reliability are reached for the account "", i.e. it is iterated converging process. , on how many I know, it is impossible to debug - why  this or that  etc. Moreover, different  (including  with different coefficients) can yield identical result on last . The data. And here on the future data can yield various result. I.e. it simply reminds the schoolboy who knows the answer, but does not know the decision. It adjusts the decision to the answer. Whether and correct it or not, works on other task of such class - does not know.

11

Re: Criticism ML with NIPS

Hello, D. Petrov, you wrote: kl>> Well so therefore also there was a direction under title Explainable AI. DP> Irony that not-explainable tenserflow now on two orders more popular direction. Well so it yields practical results, in it investments that attracts serious people - anything surprising leak. Generally on a subject quite good public discussion in FB Yann Lecun (the director of AI in ) was tore. Its message that historically there is nothing strange or surprising that practice overtakes the theory in separate directions and it not an occasion to cease to practise and wait while theoretical tools will be ready. Truth in it, of course, is, but the completing appeal to Ali that supposedly if suffices the theory go work over it, instead of complain, it is ridiculous enough.

12

Re: Criticism ML with NIPS

Hello, D. Petrov, you wrote: DP> Yesterday on NIPS there was an interesting presentation with criticism ML. By the way, here one of these days still the interesting piece appeared, began to be considered: The Case for Learned Index Structures. Now not only algorithms will be a black box, but also data structures (in this case for indexes in a DB)