Tags: quasi-Newton Method Update mutual trust interviewer introduction differential PAP boost PKU
Two rounds of technical backbone, one round technical director, and finally CTO, then HR. A total of four rounds of technical +hr, or small harvest.
Overall feel the interview problem is flexible, the difficulty and the company business convergence of knowledge points strong. The following is the specific process:
One or one sides
1. Two-way code (time complexity)
2, K-means pseudo-code (referred to the EM)
3, the project has a challenging place, the assistant
4, explain Word2vec
5, the general discussion of the next deep learning
6, deviation-variance decomposition, over-fitting and under-fitting measures
Two or two sides
1, coding (dynamic planning, code checking and exception handling is very necessary)
2. Loss function of logistic regression and explanation
3, avoid over-fitting measures, in the field of statistics on the weight how to control (machine learning and penalties, statistical hypothesis test)
4, if there is a new stream data, how to update the model (can use a random gradient descent, directly calculate the incremental update weights)
5. How to determine the correlation between a feature item and a category (entropy) formula
6, speaking small paper
Three or three sides
Technical director, is Microsoft's. Well.. The technical Danale is quite many, according to introduce the core team also has bat,facebook backbone.
1, three sides mainly introduces the internship project, through the project to understand your business logic. This part is not a good answer.
2, finally chatted for a long time the company's business point
Four or four sides
The CTO was a graduate of the Peking University Doctor's Machine study, and a few great paper, and later at Microsoft.
1, perhaps the academic comparison strong bar, come up directly let the hand push SVM, with the relaxation factor
2, then specific to the SMO to solve the convex two planning problems of the advantages, contrast gradient descent and Newton method, quasi-Newton method
3, the last introduction of the company how to use machine learning combined with business, data mining information
To tell the truth, pushing the formula is still very difficult, the head actually to all four is very tired. Fortunately, before you push yourself, although the time is a bit long, probably the process still remember. Under the CTO's error correction, reluctantly pushed out. SMO algorithm is just a simple to understand, so the general idea to speak down, specifically involving not deep. The interviewer spoke carefully about the paper and improvements, and the people who invented the algorithm.
Talk about the pay treatment, said to communicate under the letter. This time the real experience is not important (probably still hanging), the interview experience is very valuable, thank you for the timing
The final summary of the more controllable factors in the interview:
The basic etiquette is not said
The interaction is very important, the programming question does not have the matter, but must positively to want to solve, according to the interview officer to give the mentality step by step response, does not understand on the sincere question.
Write the code must be clear whether the interviewer let you write pseudo-code, not the words it is necessary to write the exception handling and return and function head. And the algorithm you write or the time complexity you mention, you know, you ask.
Although can not be said to dominate, but this aspect is a little controllable, learn to guide the process of interview, take the initiative to say what you learned recently, you are more good at what.
4. One last question
Usually ask if you have anything to ask, be sure to ask, whether it is the company's business or self-promotion, showing interest.
Interviewing startups, algorithmic engineers.