为什么李老师比我牛X

此刻距离我高考结束已经七年。曾经的辉煌早就锈迹斑斑,也许除了自己,没人愿意打磨掉尘埃与铜绿去追忆曾经闪耀的自己。往昔的荣耀渐渐变为今日的耻辱——曾经高起点的我(或者是我们)如今渐渐泯然众人,毫无闪光点。每日面对李老师,高考比我低一百分的李老师,语文总是不及格的李老师,渐渐觉得,被那些曾经不如自己的人追赶或者超越,其实也许是一种必然。

因为也许从高考结束的那一刻起,或者是更早,从出生的一刻,我就输给了李老师。

反思自己的平庸之前,我觉得应该先看看我原本应该有的赢面。也就是交大为什么优于成电。如果从一开始就坚定了出国的理想,那么交大对外宣扬的那些优势:超多的重点学科,超多的青千,还有教育部给的超多的钱,其实和我并没有什么关系——我终究也只能选择一个学科,跟一个老板,做一个优质的项目。换句话而讲,如果我没有学一个学科排名比成电高的学科,跟一个比成电老板们水平高的老板,做一个成电做不了的项目,那么其实上交大和上成电,根本没什么区别。

而与此同时,我的输面比赢面多很多。

小城(或小镇,小村)考生面对大城市考生最大的弱点就是英语。大城市出身的考生(能考到还不错的学校的),不论其他成绩如何差,英语总是好的。每次听室友说英语,听李博林博说英语,听大师嫂说英语,听大师和师弟说英语,常常能听得我无地自容。再看看我,看看大师兄,看看标哥。即无优质资源,又无环境熏陶,不仅会造成基础的不夯实,还会影响兴趣。对于大多依靠自学的本科,兴趣是一个太重要的选项,如果没了兴趣,完全是为了出国学英语,成绩好不好很难说,口语一定一定是不好的。我想我这辈子也很难做到像室友那样去自然而然去追新的美剧,也做不到像李老师那样去津津有味的翻韦氏词典,所以从一开始我的英语就输在了起跑线,不仅起跑线输了,发动机更是烂。自然,小城市里也不乏英语优秀的,比如琳琳,我一直敬佩她对于英语的浓厚兴趣,她可以耐下性子来去看英文哈利波特,她会学习那些有趣的俚语,而我对此毫无兴趣,也许是家庭熏陶和性格使然。

再者,李老师比我有更高的数学天赋。我自认为在微积分学习上花的时间绝不少于李老师,而效果很弱。比较灵我自己失望的是,我对数学学习一直抱有不求甚解的想法,从高中时与老田相对比我就发现了这一点。这个不好的习惯并未影响我高中成绩,而对于深入又复杂的大学及其后时代的数学课,这使得我的成绩一直无法提升。而李老师从来都是以最少的努力换取最好的成绩,甚至在某些课上成为了全班及助教的灯塔。。。而他只不过是个每天晚上一到家就开始嘟着嘴打游戏,10点准时上床睡觉,周末也从来不学习的宝宝。

输面很大的情况下,更不幸的是,对于利用交大优质资源这件事情,我没有任何想法,从来都是追寻别人的足迹,看着同学们都去找某老板,我也忙不迭地的接过了橄榄枝。也许我应该提前思考。老王对学生的确不错,对我更是仁至义尽,但不得不承认的是老王也处于事业刚起步阶段……实力甚至还不如李老师的老板。相比而言,项目和文章的质量上我就比李老师差了很多。

再者,对于留学申请各项资料的比重我根本没有衡量。对,我说的就是推荐信。交大拥有的出国访学机会比成电多得多,而我从来没想过去利用这些资源,瞎猫碰死耗子碰到了UCLA的机会,又被自己蹩脚的英语毁尽,这可以说是本科阶段最后悔的事情,没有之一。相比而言,李老师成功漂洋过海去了UBC,早就比我高到不知道哪里去了。

此外,交大成绩评价比成电严格。父母层次差异,提供的资源差异都是外界因素,我无法改变,所以也不想比较。

想来,我这个样子,也能被马里兰录取,和优秀的李老师肩并肩,真是我的荣幸。

下面一个问题,如果可以选择穿越回2011年9月3日大学刚开始的时候,对当时的我指点迷津,我都会说些什么。

首先是学好英语。

如果英语更好,就不会为了考G考托耽误那么长的时间,那么长的时间,应该可以读很多paper,也许就可以想到更好的idea,发出更好的论文。

如果英语更好,也许可以拿住UCLA的机会,那么现在的一切也许会很不一样。

如果英语更好,也许就可以编出更好的推荐信和SOP.

如果英语更好,也许我和Michael今天讨论OKR可以更加丝滑一点。。。

当然,也许以上的也许都只是也许,但只要能做到其中任何一点,都会使我获益良多。

其次,我会花时间了解电院的每一个分支,考虑一下自己真的喜欢什么。如果做了这一步,那么我会找一个做CV或者graphics的老板,从大二开始进实验室。能够找到大牛老板或标哥这样的小牛老板最好(这才算真正利用了交大的优质资源),找不到大牛老板,也可以提前做喜欢的事情。

再次,我会选择让自己搬出去住。当然,我猜父母绝对不会同意,但我会用尽所有的力气劝服他们。我愿意把我现在所有的钱都给当时的自己,去租一个好一点的房子,这样就不会活得那么委屈。

然而塞翁失马,焉知非福。能够拿到马里兰的AD,继而申请到TA和scholarship,遇到愿意指导我大师兄,遇到对我包容的老板,成功转系,现在看来都是极大的侥幸。

我有怨恨过交大吸纳了那么多优质的学生,但却未能为每个学生都配备优质的资源,导致交大的学生不如成电的学生。但反过来想,优质资源总是有限的,要靠自己争取,不能等着学校一口一口喂。反之,我也可以通过自己的努力去超越那些清北的本科生,我也可以依靠今日在科研上的上进,去超越那些学校排名比UMD更高的,事实是我也做到了。但同时,我也要淡定的面对一些背景更加平庸,甚至二本出身的人同样在科研上做出优秀的成就。

所以更应该做到的一件事,是心态的平稳。上至清北,下至复交浙科,天南华武,西交西工北邮成电,大家毕业后,还是都要一起做码农,老了以后一起跳广场舞的。表象声色皆是皮下白骨,从来就没有什么高低贵贱,今日的辉煌与黯淡,聪慧与愚蠢,远见与短浅,富贵与贫穷,ML灌水文和我今日的碎碎念,不过都是百年之后的电子垃圾,和可降解的无机盐。

一路走到今天,生命中实在是有太多的侥幸。而来到美帝以后最大的侥幸,还是遇到了好战友李老师。作为班主任,自己刷题很厉害又愿意带小伙伴们一起刷题的李老师,一路带领我刷过狗家&脸家, 无愧于五道杠班主任的称号。作为热心的邻居,在预约看病买保险买车险上悉心指导,无愧于活的百科全书这一称号。作为心思活络的同学,短短几句话就为傻乎乎的我省下了几千刀学费,不枉我这几年尊称其为伟大的老师。

《古剑》中有太子长琴,性格温和沉静,醉心山水,亲近自然,乃翩然神仙之典范。李老师含蓄深邃,睿智韬晦,又热爱大好河山,也热爱小兔子, 小松鼠,自然也是翩然神仙之典范。

行文至此,向李老师致敬,李老师就是牛X。

 

 

 

Structured prediction

baseNP: doesn’t contain any recursive parts.

chunking: build the tree for the sentence

Level of representation:

* Brown Corpus (level1: pos)

* Penn Trecbank (level2: sys)

* PropBank (level3: sen)

* Framenet (level4: )

All of these need lots of human labor.

 

h(x) = argmin(y in Y) E_(y~p(Y|X))[l(y,x,Y)]

l (y*,x,y) = 1 – delta(y,y*)

H(x) = argmax_(y in Y) Pr(y|x)

min_(h in H) E_{p}[loss(X;Y;h)] + model complexity(h)

Empirical risk = 1/N SUM_{I = 1}^{N}loss(x,y*,h)

 

generalized viterbi

recognize speech

wreak a nice beach

an ice beach

 

conditional random fields

Talk: Learning efficiency of outcome in games

  • By Eva Tardos
  • Repeated games
    • player’s value/cost additive over periods, while playing
    • players try to learn what is the best from past data
    • what can we say about the outcome? how long do they have to stay to ensure OK social welfare?
  • Result: routing, limit for very small users
    • Theorem:
      • In any network with continuous, non-decreasing cost functions and small users
      • cost of Nash with rates ri for all i <= cost of opt with rate 2ri for all i
    • Nash equilibrium: stable solution where no player had incentive to deviate.
    • Price of Anarchy = cost of worst Nash equilibrium / social optimum cost;
  • Examples of price of anarchy bounds
  • Price of anarchy in auctions
    • First price is auction
    • All pay auction…
    • Other applications include:
      • public goods
      • fair sharing
      • Walrasian Mechanism
  • Repeated game that is slowly changing
    • Dynamic population model
      • at each step t each player I is replaced with an arbitrary new player with probability p
      • in a population of N players, each step, Np players replaced in expectation
      • population changes all the time: need to adjust
      • players stay long enough…
  • Learning in repeated game
    • what is learning?
    • Does learning lead to finding Nash equilibrium?
    • fictitious play = best respond to past history of other players goal: “pre-play” as a way to learn to play Nash
  • Find a better idea when the game is playing?
    • Change of focus: outcome of learning in playing
  • Nash equilibrium of the one shot game?
    • Nash equilibrium of the one-shot game: stable actions a with no regret for any alternate strategy x.
    • cost_i(x, a_-i) >= cost_i(a)
  • Behavior is far from stable
  • no regret without stability: learning
    • no regret: for any fixed action x (cost \in [0,1]):
      • sum_t(cost_i(a^t)) <= sum_t(cost_i(x, a_-i^t)) + error
      • error <= √T (if o(T) called no-regret)
  • Outcome of no-regret learning in a fixed game
    • limit distribution sigma of play (action vectors a = (a1, a2,…,an))
  • No regret leanring as a behavior model:
    •  Pro:
      • no need for common prior or rationality assumption on opponents
      • behavioral assumption: if there is a consistently good strategy: please notice!
      • algorithm: many simple rules ensure regret approx.
      • Behavior model ….
  • Distribution of smallest rationalizable multiplicative regret
    • strictly positive regret: learning phase maybe better than no-regret
  • Today (with d options):
    • sum_t(cost_i(a^t)) <= sum_t(cost_i(x, a_-i^t)) + √Tlogd
    • sum_t(cost_i(a^t)) <= (1 + epsilon)sum_t(cost_i(x, a_-i^t)) + log(d)/epsilon
  • Quality of learning outcome
  • Proof technique: Smoothness
  • Learning and price of anarchy
  • Learning in dynamic games
    • Dynamic population model
      • at each step t each player I is replaced with an arbitrary new player with probability p
    • how should they learn from data?
  • Need for adaptive learning
  • Adapting result to dynamic populations
    • inequality we wish to have
  • Change in optimum solution
  • Use differential privacy -> stable solution

How to write MP4 with OpenCV3

When trying to write MP4 file (H264), I tired code of

And I got error saying:

This problem is solved by changing the fourcc to the ASCII number directly to cv2.VideoWriter(), i.e.

reference:

https://devtalk.nvidia.com/default/topic/1029451/jetson-tx2/-python-what-is-the-four-characters-fourcc-code-for-mp4-encoding-on-tx2/

Paper Reading: View Direction and Bandwidth Adaptive 360 Degree Video Streaming using a Two-Tier System

Each segment is coded as a base-tier (BT) chunk, and multiple enhancement-tier (ET) chunks.

BT chunks:

represent the entire 360 view at a low bit rate and are pre-fetched in a long display buffer to smooth the network jitters effectively and guarantee that any desired FOV can be rendered with minimum stalls.

ET chunks:

Facebook 360 video:
https://code.facebook.com/posts/1638767863078802
Assessment:
https://code.facebook.com/posts/2058037817807164

PointNet, PointNet++, and PU-Net

point cloud -> deep network -> classification / segmentation / super-resolution

traditional classification / segmentation:

projection onto 2D plane and use 2D classification / segmentation

unordered set

point(Vec3) -> feature vector (Vec5) -> normalize (end with the bound of the pointcloud)

N points:

segmentation:

feature from N points ->NxK classes of each point (each point will have a class)

classification:

feature from N points -> K x 1 vector (K classes)

 

Lecture 10: Neural Network

  1. Deep learning
  2. Representation learning
  3. Rule-based
    1. high explainability
  4. Linguistic supervision
  5. Semi-supervision
    1. have small set of data with label
    2. has large set of data without label
  6. Recurrent-level supervision
  7. Language structure

description lengths DL= size(lexicon) + size( encoding)

  1. lex1
    1. do
    2. the kitty
    3. you
    4. like
    5. see
  2. Lex2
    1. do
    2. you
    3. like
    4. see
    5. the
    6. kitty
  3. How to evaluate the two lexicons?
    1. lex 1 have 5 words, lex 2 has 6 words
    2. Potential sequence
      1. lex1: 1 3 5 2, 5 2, 1 3 4 2
      2. lex2: 1 3 5 2 6, 5 2 6, 1 3 4 2 6
  1. MDL: minimum description lengths
    1. unsupervised
    2. prosodic bootstrapping

Boltzmenn machine

Lexical space

relatedness vs. similarity

  • use near neighbors: similarity
  • use far neighbors: relatedness

ws-353 has similarity & relatedness

loss function:

 

project:

Part1: potential methods

  • LDA
  • readability
  • syntactic analysis

 

 

Questions about “Foveated 3D Graphics (Microsoft)” User Study

  1. Problem1: They did test for only one scene.
    1. The first problem is that foveation level is highly depentent on scene. They may get totally different parameters if they change to another scene. Of course, this is the problem of all the user studies. Till now, only NVIDIA mentiond the multliple factors affecting vision. However, they don’t have good ways to deal with this.
    2. The second problem is about data analysis. They avoid the problem of one parameter ->multiple result by testing only one scene.
  2. Problem2: I don’t believe that their result is monotone.
    1.  They just said:
      1. Ramp Test: For the ramp test, we identified this threshold as the lowest quality index for which each subject incorrectly labeled the ramp direction or reported that quality did not change over the ramp.
      2. Pair Test: for the pair test, we identified a foveation quality threshold for each subject as the lowest variable index j
        he or she reported as equal to or better in quality than the non-foveated reference.
    2. Suppose their quality level is 11,12,13,14,15. What if they get result of 1,1,1,0,1 ? Is their final quality level 13 or 15?
      1. I don’t believe this situation did happen in their user study.
      2. If it happens, what should we do? Of course we should test for multiple scenes for many participants, and get the average. So we go back to problem 1.