If Trump can put a gag order on EPA, he probably will try to do the same to BLS in the future. No more unemployment statistics. That would be the easiest way for him to claim credit for creating jobs.

彈性午餐時間:下晝四點食壽司一定冇人同你爭。
Flexible lunch hours: I have the whole sushi restaurant all for myself.

20170125_160853

除了AlphaGo,其實還有很多值得留意的人工智能發展。例如現在正在美國進行中的無上限德州撲克大賽,由Carnegie Mellon大學的人工智能對世界上最頂尖的人類選手進行20天的比賽。到目前為止戰況可說是一面倒向電腦一方。
和圍棋不同,德州撲克是一種有隱蔽資訊的遊戲,參賽者並不知道對手以及未派發的牌為何。因此德州撲克相比起圍棋對解決很多現實中的問題意義更大。
Texas Hold’em is a much more interesting game than Go when it comes to real-world artificial intelligence application, since many important real world issues inherently involves imperfect information. 5 days into the game and the human pros are losing badly.

Claudico battles Donger Kim, Jason Les, Bjorn Li, Doug Polk in Brains vs. AI at Rivers Casion, the first time that a computer program has played Heads-up No-limit Texas Hold’em in competition with top human players.

Source: Brains vs AI

有同學傳來CU Secrets個好grade批鬥大會,說GPA 3.5都怕無second hon云云。少年你們太年輕了,都未上Intermediate Macro,急乜?
Dear students, wait till you went through Intermediate Macro before you say everyone is getting high GPA.

係同一個地方教咗六年書,facebook上面一大部分嘅friends都係學系嘅學生,令到facebook嗰feed亦都好受學生影響。例如今晚個feed成日都係某位女同學嘅個人特寫。
Having taught in the same department for six years, a significant number of my facebook friends are students and alumni. This means that my feed is highly influenced by this group—like tonight, a portrait of a certain female student keeps popping up because too many of my facebook friends “liked” the photo.

研究助理入緊香港車牌拍賣數據,問:「點解 DGS 高 [DBS] 咁多」?
係囉,點解?
While entering Hong Kong license plate auction data, research assistant noted that #DGS fetched a much higher price than #DBS…as if it could have been the other way round.

TensorFlow and CUDA 8.0

TensorFlow up and running with CUDA 8.0. Life-saving tip: always explicitly enter your CUDA and cuDNN versions during configuration.

My first impression is TF is a bit behind Nervana Neon in terms of training speed and ease of use. Too bad the latter now belongs to Intel, which clearly wants to promote its own Pentium-based Xeon Phi over Nvidia GPU.

Screenshot 2016-08-30 02.08.53

全民退保學者與統計學

prediction

有同學問我如何看全民退保學者方案的計算,特別最近引起一番爭議的變數問題。我想從社會科學就統計運用出發,解釋一下問題所在。

在社會科學研究中,統計學的主要應用在於估算某因素對另一因素的影響:例如吸煙對存活率的影響,又或加價對需求的影響等。這樣的估算用的總是已有數據,分別只是數據有多新而已。因為已發生的事實只可能有一個,學者的責任就是盡力準確估算這個未知的事實。

因應上述的要求,傳統的社會科學統計訓練非常重視估算的無偏性(unbiasedness)。無偏性的意思是估算平均來說是準確的,而最具代表性的例子莫過於常見的線性回歸分析(Ordinary Least Square)。讀過計量經濟學的同學或許會記得(惡夢!)OLS的特點「BLUE」—Best Linear Unbiased Estimator。意思就是在無偏性的要求下,沒有其他線性模型比OLS更佳。

問題是,準確估算過往效果和就未來作出推算是兩會事。事實上,無偏性和估算的誤差範圍是有著取捨關係,統計學就此亦有相當的理論和驗證。在較進階的計量經濟學課程中—即係每年只有幾個本科同學會選修嗰啲—會教授一些相關的知識,但在整個經濟學課程中並非太受重視,因為如前所述,學者最需要的技巧是準確分析已發生的數據。因為經濟學以外之社會科學要求的統計學訓練更少,問題相信只會比經濟學更嚴重。

可能有同學會問,像全民退保這樣重要的社會議題,怎麼沒有統計學者去做嚴謹的推算?答案很簡單,因為對仕途有害無益。我過去已經解釋過,香港由教資會以降均看重學者著作的國際影響力,而像全民退保這樣的本土議題是上不了好的國際期刊,搞不好還要被人批評一番。吃力不討好,又何必自討苦吃?

《全民退保學者方案》:
http://www.legco.gov.hk/yr15-16/chinese/panels/ws/ws_rp/papers/ws_rp20151207cb2-398-1-c.pdf