author : Wang Chao


I'll have one, too 《 Excellent algorithm engineers don't need deep learning 》

Homage to the divine 《 Good girls don't have sex <https://zhuanlan.zhihu.com/p/31956362>》

thank @Manjusaka <https://www.zhihu.com/people/25bd46cbd3f60fb99aee8b0d76e7b41d>《
Good algorithm scientists don't need to be unique <https://zhuanlan.zhihu.com/p/32130515>》 inspiration .

I interviewed one the other day C9 Fresh master , Pattern recognition major , Failed to answer several professional questions in a row .

Beyond embarrassment , I asked him :「 Do you have any ideals ? What's your most desired thing now ?」

He twists his big eyes , No false thinking cableway :「 take kaiming Big Resnet Extend to 10 Ten thousand stories , hold kitti,COCO Database detection and recognition task promotion 20 More than points 」

I didn't expect this kind of operation in the interview .

I ask why this can be the most desired thing at this stage , He asked 「 Don't you LeCun,Bengio and Hinton Are you moved by the persistent spirit of ? Don't you envy ILSVRC2012
AlexNet Is it brilliant ? Don't you get googlenet,Resnet Are you convinced by his profound thoughts ?」

It makes sense. I can't argue .

Engineers who know the market so well , Must be a rare talent !

therefore , I decided : Don't admit him .

These years , Deep learning in CV The field is very popular , Whether it is detection,segmentation,classification, still stereo
matching,pose estimation, In-depth study of the traditional state of the
art Method of flying . nowadays , Algorithm engineers don't know classical networks , I'm too embarrassed to say hello to others .


This phenomenon is only limited to the little white just entering the pit . But for immersion 5 No matter more than years bat Algorithm manager or senior human feature design engineer , This kind of thing is just a joke to them . in the course of time , I found a cruel common ground ——

They only use traditional methods .

Isn't the salary high ? The lowest monthly salary 20K+, And corporate options and shares .

Is the technology not good ?Paper Soft hair , Just compile the code once .

I asked one of them : I think you focus on different tasks all day , Manual design features , Classifier is not tired , Don't want to try cnn Method ?

He said : crap , I think so .

I asked : Then why don't you try LeNet,AlexNet What about ,caffe Are there any examples under the framework ?

He sighed : No , Too busy .

He means a lot : Excellent algorithm engineers don't need deep learning methods .

The man I was talking to , He's a great God ,BS CMU,MS Stanford,doctor stay MIT, Three years to complete five years course , During reading the blog 100 Multiple articles sci,h index
40 quite a few , Return to China as chief scientist of start-up company , Nothing to do Paper play , And the citations are considerable .

Follow interest input cv Research 10 several years , As early as 01 year Paul
Viola propose Haar And cascade adaboost Time , Minor modification of features and classifiers , Recognition rate improved 0.1%, In the field of face detection state of the
art. Later, I couldn't stand the discrimination of foreign universities on Chinese academic , Resolutely return to China , Financial freedom has long been achieved .

This kind of algorithmic guy , I want to come out , Never buy a good video card to run deep learning , Buying a video card should also be used to eat chicken .

I still know another algorithm engineer .

<https://shop162567423.taobao.com/>
Favorite traditional features and classifiers , image gabor wave filter ,LBP features ,adaboost algorithm ,SVM classification ,random
forest It's a treasure of nature , Always marvel at harr Feature in face detection ,hog In human detection ,LBP The success of face recognition brings tears to our eyes . But also because of obsession , Every day :
Show , Legendary , May Fourth leader , Tihuazhixiu . Dismissive of deep learning , We can make a comparison cnn more work Traditional features + Classifier method .

This talent is a clear stream in the circle .


His biggest hobby is video surveillance rgb On the foreground image generated by Gaussian background modeling , Using meter scale to measure the display to determine the aspect ratio of objects , Remove leaf jitter , Water waves , False alarm caused by kowtow machine, etc . In the eyes of Xiaobai, who just entered the pit, these divine operations , It's so out of place . Sometimes he was advised , at present cnn By pruning , Compression model and other technologies , stay 1080ti Real time on , You can try it, too , Reduce the workload .

He will not be affected .

I asked him : How on earth do you keep a normal heart , Others use deep learning to achieve better generalization effect , You design features and classifiers by hand for each scene , Not tired ?

He smiles , say : What's wrong with being tired , Not willing to let deep learning run at the low end arm Try it on the platform ?

I get it right away .

The same person , Using traditional methods , Better effect than deep learning in specific scene , And through 5 Optimization of wheel algorithm , Just transplant the algorithm to the low-end platform , Every way for the rest of the company 2K Cost of , It's a spirit of craftsmanship .

Last week with a CEO Friends out for dinner , He said he recruited some vegetable chicken Algorithm Engineers , Always want to buy more than one 1080ti Video card , Improving R & D efficiency with deep learning method .

He told me ,「 But you're different , You will always remind me , Keep me away from the comfort zone , I can't settle for the status quo , I need to be crisis conscious , You seem to be particularly interested in reusing traditional methods to solve problems .」

People need independence .

Still young forever , Tears in my eyes forever .

Still want to be full of expectation of beauty all the time , Yearning for the future , Awe of classics .

Remember , Know the world, but not the world , Far away from the Jianghu , It's the kindest maturity .