丹麦语 DU 3.1 Med bussen, Mad og indkøb, Sundhed og livsstil
Med Bussen venter 等待 til sidst 终于 lukker 关闭 bevægelse og bevægelsesretning går ned 向下走 går op 向上走 går hen 向前走 går ud 向后走 går ind 向里走 Han rejser sig (op). 他起立。 Han sætter sig (ned). 他坐下。 går (hen) til 过去 går (væk) fra 离开 går igennem 通过 går forbi 经过 står på 上车 = går ind i bussen står af 下车 = går ud af bussen Retnings- og stedsadverbier 方向和地点副词,有长形式和短形式之分。当动作产生位移的时,用短形式表方向 Retning;当动作没有产生位移时,用长形式表地点 Sted “静长动短” Kort form (retning) Lang form (sted) hjem 回家 hjemme 在家 hen 向前 henne 在前 ind ...
计算机视觉 W41 Training ConvNet (2)
Training pipeline set up architecture i.e. computational graph train using mini-batch stochastic gradient descent (SGD) loop: sample a batch of images forward propagation it through network, get loss back propagation to calculate the gradient update parameters using the gradient Be careful activation function ReLU, careful with learning rates - leaky ReLU try tanh don’t use sigmoid data regularization always use zero-center data normalize to control variance apply the same pre-processing ...
算法博弈论 W41 Exercise, examples, Q&A
Exercise Exercise 3.4 sponsored search auctions nnn bidders, nnn slots. each bidder iii has a value of viv_ivi. the j-th slot has CTR aja_jaj, with a1≥a2≥⋯≥ana_1\ge a_2\ge\cdots\ge a_na1≥a2≥⋯≥an maximize social welfare denoting by δ(i)\delta(i)δ(i) the slot bidder iii occupies, this means that the quantity ∑iaδ(i)vi\sum_ia_{\delta(i)}v_i∑iaδ(i)vi is as high as possible Q: design a DSIC mechanism that maximizes social welfare allocation rule should be monotone. focus on bidder iii and c ...
计算几何 W40 Linear Programming
阅读材料 课堂笔记 Convex Hull and Sorting convex hull algorithms: O(nlogn)O(n\log n)O(nlogn). can we do better? convex hull of points is equivalent to sorted numbers. O(nlogn)O(n\log n)O(nlogn) worst-case is optimal. fewer than n points on the hull O(nlogh)O(n\log h)O(nlogh)? Marriage Before Conquest MbC(P:l,r): find Upper Hull of P from x-coordinates l to r find median x-coordinate xmx_mxm. partition into Pl,PrP_l,P_rPl,Pr. O(n)O(n)O(n) times find the bridge, uv, over xmx_mxm. remove the ...
丹麦语 DU 3.1 将来时/过去时
I klassen 2 一些日常用语 Tak. 谢谢。 Selv tak. 不用谢。 Ja selvfølgelig. 当然。 Undskyld. 对不起。 Det er i orden. 没关系。 Modalverber kan 能够 skal 必须,应该 vil 将要,将会 må 可以 情态动词后接动词不定式 infinitiv Ordstilling: S - V1 - V2 Han kan tale tysk. De vil tag bussen. Spørgsmål: V1 - S - V2 Kan han tale tysk? Vil de tag bussen? 动词的命令式 imperativ,不定式 infinitiv,现在时 nutid Imperativ - Infinitiv -e Nutid -er tal at tale taler kom at komme kommer tag at tage tager bo at bo bor 命令式就是动词的词根 Planer/ønsker med ‘skal’ og ‘ ...
计算机视觉 W40 Training ConvNet (1)
Review cnn: sparse interations and parameter sharing computational graphs: automatic differentiation Training pipeline set up network architecture = computational graph train using mini-batch stochastic gradient descent loop sample a batch of images forward propagation through the network and get loss back propagation to calculate the gradient update parameters using the gradient Activation Function Sigmoid σ(x)=11+e−x\sigma(x)=\frac{1}{1+e^{-x}}σ(x)=1+e−x1 nice interpretation as a sat ...
算法博弈论 W40 Simple Near-Optimal Auctions
Revenue-maximizing auction Disadvantages somehow complicated require detailed information about the probability distributions of the valuations much different than the format used in practice. Review allocation rule: x(v)=argmaxX∑iφi(vi)⋅xi(v)\mathbf x(\mathbf v)=\arg\max_X\sum_i\varphi_i(v_i)\cdot x_i(\mathbf v)x(v)=argmaxX∑iφi(vi)⋅xi(v). for profile v\mathbf vv, φi(vi)=vi−1−Fi(vi)fi(vi)\varphi_i(v_i)=v_i-\frac{1-F_i(v_i)}{f_i(v_i)}φi(vi)=vi−fi(vi)1−Fi(vi) is the virtual val ...
计算几何 W39 Introduction to Computational Geometry
阅读材料 Classroom examples of robustness problems in geometric computations by Lutz Kettner et al. substituting floating point arithmetic for the assumed real arithmetic may cause implementation to fail. focus on floating point - real numbers issue IEEE 754 课堂笔记 Computational Geometry algorithms that deal with geometric data points lines circles hyperplanes applications graphic databases geographical information system robotics one example: a file contains n lines where the i-th line contai ...
丹麦语 DU 3.1 Familie, Klokken, Hverdag of weekend
Familie En familie 直系亲属 en kone 妻子 en mand 丈夫 en mor 母亲 en far 父亲 forældre 父母 børn 孩子 en datter - to døtre 女儿 en søn - to sønner 儿子 兄弟姐妹 en søster - to søstre 姐妹 en bror - to brødre 兄弟 en lillebror - to små brødre 弟弟 en storebror - to storebrødre 哥哥 en lillesøster - to små søster 妹妹 en storesøster - to storesøstre 姐姐 祖辈 en mormor 外祖母 en morfar 外祖父 en farmor 祖母 en farfar 祖父 et barnebarn - to børnebørn 孙辈 en oldemor 曾(外)祖母 en oldefar 曾(外)祖父 en oldebarn 曾(外)孙辈 其他亲戚关系 en moster 姨 = mors søster ...
计算机视觉 W39 Backpropagation
Back propagation with manual derivation Let’s consider a neural network LLL is total number of layers sls_lsl is number of units (except bais) in layer lll. (W,b)=(Wij(l),bi(l))l=1nL(W,b)=(W_{ij}^{(l)},b_i^{(l)})_{l=1}^{n_L}(W,b)=(Wij(l),bi(l))l=1nL is model parameters, where we write Wij(l)W_{ij}^{(l)}Wij(l) to denote the weight associated with the connection between unit jjj in layer lll, unit iii in layer l+1l+1l+1. bi(l)b_i^{(l)}bi(l) is the bias associated with unit iii in layer l+1 ...