转发: Samsung ML岗 onsite 面经

Samsung onsite interview day1 (10/4/2018)1st round:1 What’s l1 norm and l2 norm? What’s the difference between them? What’s the advantages and disadvantages?2 The input feature size is 248x248x5, the kernel size is 5x5x3, the stride is 2, what’s the total multiplication operations?

2nd round:1 How is batch normalization implemented?2 The difference to implement logistic regression & SVM?3 Coding: implement a 1D convolution

3rd round:1 coding: permutation. Given two lists, to decide if they are permutation of each other.2 What’s difference of SGD, Momentum SGD and Adam?3 If two signals have rectangle shape, what’s the shape for result after doing convolution? And doing convolution on two triangle signals, what’s shape for output?

4t’h round:1 How many points on earth that when you walk 1 mile to south, and then 1 mile to west, further 1 mile to north, and finally you get back to the starting point?2 There are three boxes, one is apples, one is oranges, and one is apples and oranges. They all have wrong labels. Assume you can have only one chance to get one fruit from the box, how to operate and get the right label for all three labels.

5th round:1 Why l1 and group lasso can do sparsity?2 Write the group lasso function.

6th round:1 Floating-point representation2 概率题3 IQ 题

7th round:1 How is RSA doing encryption and decryption?2 How is your optimization on RSA doing?

Samsung onsite interview day2(10/5/2018)1st round:1 What’s the difference between MobileNet and MobieNetV2?2 What’s the difference between MobileNet and ShuffleNet?3 Coding: h-index in citations

2nd round:1 Fixed-point and floating-point2 nips using Gaussion weight distribution to do fixed-point representation

3rd round:Good suggestion: train the network for more times and pick the best quantized value

4th round:1 How to get the reverse matrix of one matrix?2 eigenvalue3 傅里叶变换公式4 signal convolution5 Write MSE error function, cross entropy function6 How is batch normalization implemented in inference stage?

5th round:1 Coding: given a list of numbers, return the longest increasing sequence2 Coding: implement 2D convolution

6th round:1 The probability that meet 1 self-driving car in one hour is 90%. What’s the probability that meet 1 self-driving car in 10min. You can use Possion distribution2 You need one switch to get 1 output from 2 inputs. And three switched to get 1 output from 4 inputs. How many switches should you use get 1 output from 1000 inputs? 999. Because each switch can help you to get rid of one input, so if you have only 1 output, you need 999 switches to get rid of 999 input signals.3 How do you address difficulties when you doing research? 多回答一些，回答个五六条4 What’s the most challenges when you doing your research?5 Why you choose Samsung?

7th round:1 Some examples of supervised learning and unsupervised learning2 Quicksort3 Reversesort4 Epoch, iteration5 Tensorflow code for defining a shape