Sunday, June 16, 2019

再Pivoting

决定再减法到sequential on paper and algo
mm说的对 感觉不太可能两个parallel
接下来专心pru and paper

Toastmasters 的星期六

早上解决了ocr和anomaly detection 的test case
anomaly detection
a. one class svm是用origin 作为唯一positive example
b. clustering, k-means, k-nearest neighbour 都是常用方法 (还有吗?)
c. statistics 也是常用方法

ocr 的话
a. 基本是字的识别,然后字的辨识

Toastmasters 的时候看了recommender system 的survey 和 youtube 的architecture
recommender system
a. recommender system上各种deep learning大家都几乎做了一遍
b. 基本的idea是 collaborative filtering 是 linear 的, deep learning 是non-linear 的,还能combine other features
c. sequantial collaborative filtering 不一定好,有时会overfit, 比如user search 了query后,点击里面的内容 sequential 的话会provide 没点击的内容 (需要看清楚些?)

Toastmasters主要收获是
1. 了解了HR 眼中的interview
a. 需要和hirer 有相似兴趣
b. 需要bring value to hirer 眼中的problem
c. quantitatively say about your contribution

2. 有个talk的机会

====
早上一直弄到一点多
然后11点30才回到
====

Pru - yes
Paper - yes
Algo - no


Algo 需要加油
Pytorch, Tensorflow 搞起来。




Saturday, June 15, 2019

相邻点的互动

一个澳洲lecturer, 被老板push, 升上了senior lecturer
香港的富豪,因为法令改变,财产签到了新加坡

两件事情毫无关联,unexpected, 但是有个特征:
领近点因为外力连接了起来。

相邻的点,会有意想不到的互动,这个是新体会。

第二次engagement

今天是第二次engagement

Harp on Pru:
1. Business Value driven - 每做一样东西想的是十倍回报
2. Better outcome - Use raw data rather than masked data
3. Better resource utilize - Do more complex project rather than simple one (where their data science can do)
4. Ask for later deadline

Sam pitching
1. Good discount - only for the customer
200k for DSE, appliance, ICPD.
Only 5 discount, quickly chop for ...
Do not move the end of June deadline
2. Praise the customer
a. Yes, that's the phase approach

Me
1. Need to say your name after shake hand, give name cards
2. Let Harp sit first, then sit
====
Pru - Understand contract
Paper - 0
Algo - 0

Too much 心里扰动after talking to Adni. Ultimately, the best state is talk, and forget. And continue to do the work.
====
追根究底,没有持续进行因为没把paper和 algo 看的太重要。
明天
1. OCR + Anomaly detection survey
2. Paper Survey
3. Algo (1 hr)


Thursday, June 13, 2019

第400篇

今天的效率还算可以,就是前面太紧,后面就要松。
其实专心到最后,应该要有boring, 就像和vikesh, LY 的partner coding 那样,不紧,就专心致志。我觉得要把专心变成一种自然,这样才是sustainable。
基本对于recommendation有一个比较大体的了解了。

bottleneck:
- 对于pytorch 和 tensorflow的熟练
- 对于recommendation 整体趋势的了解
- algorithm的训练

paper - 整体来说,需要改变的是要松弛,要fun
algorithm - 必须要放一小时进去
pru - 把data science level 要放入进程



Wednesday, June 12, 2019

杂事+休闲

1. 我觉得这几天主要经验是 杂事不能入,因为他会expand, 会繁衍
2. 休闲必须有- 最少1.5 小时 就是为了杂事。

Tuesday, June 11, 2019

Engagement Training

今天老板给了个三小时的training, 了解了不少当中该注意的东西。
思考为什么ibm可以survive过一百年,
1. 了解企业,
2. 推销 dse team, medium, linkedin, blog, talk, (这部分要算kpi)
3. 不能work就卖掉 rational, lotus
4. 买公司 ilog, spss, redhat
5. 让员工不断学习
6. 其实有很多东西都是ibm这里发明的像des

也了解了各种不同team 之间的dynamic.

===
Pru - 看了些note。缺乏重点学习。应该把重点聊下来在explore.
Paper - 0
Algo - 0
====

Algo - 8:30pm
  a. Crucial thing is to summarize in notes before moving to new algo
Paper- 5:30pm
  a. Crucial thing is to setup the code base for the state-of-the-art, and start exploring
Pru-8:00-5:30pm
  a. Crucial thing is to setup the code base for models + node for feature engineering