Background: 985 bachelor’s degree, 985 master’s degree

Direction: Microservices and Software Architecture

Position: Autonomous algorithm engineer

To introduce myself

Conventional content

Project introduction

  • See you have done the rubik’s cube project, please give me a general introduction

    The content of the project is introduced, including project objectives, solutions, difficulties and ideas

  • Why did you use LSTM to solve this problem

    Since this project is essentially a classification problem of a set of sequential data, I thought of LSTM in NLP to deal with this problem

  • Why use deep learning instead of thresholds or rules

    Threshold and rules are prone to oscillations, that is, the classification results are unstable near the threshold, so we hope to get more robust results by using machine learning methods

  • You also used reinforcement learning. What’s your reward

    Is the distance between the current state and the target state, and I’m calculating the mean square error

  • I see you did a good job on the written test, but you didn’t pass all the test cases on the first question

    (I got two out of five questions, and it was not bad?) It looked like they all passed when I saw them

Attention is DevOps

  • Gee, it says here you got a perfect score, but you didn’t pass all the test cases. Weird. Tell me the difference between RCNN and Attention

    Attention mainly solves the problem of long-term dependence of RCNN. In the traditional Seq2Seq architecture, RCNN only uses the last hidden vector as output when decoding. When the source sequence is too long, the data at the front end of the sequence will be ignored, and the data station at the back end of the sequence is too heavy, so it is impossible to use the whole sequence to output. Attention provides Attention mechanism for this problem, which is essentially a query mechanism. Q is used to query the distance between Q and K of each K-V pair, and different weights are assigned to different V according to this distance. In Seq2Seq architecture, it can be expressed as: using the current implicit vector when decoding, Query the distance between it and the encoding implicit vector, and assign a higher weight to the close implicit vector. Finally, the encoding implicit vector is weighted and summed as an input in the current decoding stage to participate in the decoding process.

  • The HR of another company called in the middle, but didn’t dare to hang up. After talking to HR, she answered the phone. I came back feeling that HR was a little unhappy

  • You’ve also done DevOps, so tell me about your research

    He introduced himself to his work

Their self-introduction

  • Ok, let me introduce our situation to you. Their direction is to use the physical data (pedestrian coordinates, speed, vehicle coordinates, speed, etc.) to predict the movement and position of people and vehicles in the following period of time.

    I would like to ask what is the difficulty of our task (a little cerebral palsy, I should not ask this question, so that HR thinks I think their task is not difficult).

  • As I said, it’s the interaction between the person and the vehicle, and different actions may occur in different situations

    At present, I think our prediction is a little difficult, because the physical data are limited, and the behavior of people and vehicles is subject to a lot of subjective factors. For example, the owner of the vehicle may be in a bad mood today, so he will not yield to pedestrians.

  • You know, in academia there are things that you do, you know, you analyze the emotions of the car owners and you make a judgment, but we’re still experimenting

Ok.

  • So do you think you’d be interested in our work

    I think maybe I have more experience in CV and am interested in it. Maybe I am more interested in direct perception. But I just heard you said that we also have a map recognition work, so I am more interested in doing it.

  • (The interviewer is slightly displeased) We are working in Beijing, is that ok with you

    Is there any base anywhere else

  • B: No, only Beijing

    Oh, well, Beijing, I think I can handle it

  • Well, we’ll let you know if there’s another interview

feeling

This interview made me realize that I should not be too straight. Although I did not understand their research direction to some extent, the geographical conditions of working in Beijing did not particularly meet my current wish. But at least this is a good opportunity to learn. Meituan is a big factory anyway, and this opportunity is very rare anyway. I should have been more positive in the last few questions, but IT may be because of my personality that I am not good at lying or hiding my emotions. However, there are pros and cons. In the future, I should try to seize as many opportunities as possible and choose after I have room to choose.