本笔记基于清华大学《机器学习》的课程讲义中有关机器学习的此前未提到的部分，基本为笔者在考试前一两天所作的Cheat Sheet。内容较多，并不详细，主要作为复习和记忆的资料。

Robust Machine Learning

Attack: PGD

• ${\max }_{\delta \in \Delta }Loss(f_{\theta }(x+\delta ),y)$: find $\delta$ that maximizes the loss
• How to compute $\delta$?
  • Projected Gradient Descent: GD then project back to $\Delta$
  • Fast Gradient Sign Method: For example, Δ = { δ : ∣ δ ∣ ∞ ≤ ϵ } \Delta =\{\delta:|\delta|_\infty\le \epsilon\} Δ={δ:∣δ∣∞​≤ϵ}. Then as learning rate goes to $\infty$, we always go to corner. Then $\delta =\eta \cdot sign({\nabla }_x{L}_{\theta }(x+{\delta }_t,y))$, we only take the sign.
  • PGD: runs FGSM multiple times.

Defend: Adversarial Training

• Daskin’s Theorem:
  $$\frac{\partial }{\partial \theta }\underset{\delta }{\max }L(f_{\theta }(x+\delta ,y))=\frac{\partial }{\partial \theta }L(f_{\theta }(x+{\delta }^{\ast },y))$$
  Only care about the worst attack samples. For a batch of samples, find ${\delta }^{\ast }$, then do GD on $\theta$ based on ${\delta }^{\ast }$.

• Models become more semantically meaningful by adversarial training

Robust Feature Learning

• For a dog photo $x$, attack it to $x^{\prime }$, so that $f(x^{\prime })=cat$. This $(x^{\prime },cat)$ training set gives model the “non-robust” features.
• For train image $x$, generate from random initialization $x_{\tau }$ such that $g(x)\approx g(x_{\tau })$. $x_{\tau }$ is also a good training sample that gives similar robust feature to the model.

False Sense of Security

• Backward pass differentiable approximation(ignore the gradient that is hard to compute).
• Take multiple samples to solve the randomness.

Provable Robust Certificates

• Classification problem: use histogram, pick the largest color nearby.
• Compare the histogram centered at $x$ and $x+\delta$:
  • Worst case: Greedy filling(*)

Hyperparameter

Bayesian Optimization

• estimate the parameters as Gaussian.
• explore both the great-variance point and the low point.

Gradient descent

• Directly use GD to find hyperparameter: memory storage problem
• SGD with momentum: $v_{t+1}={\gamma }_t{v}_t-(1-{\gamma }_t){\nabla }_{w}L(w_t,\theta ,t)$
  • Store $w_{t+1},v_{t+1}$ and ${\nabla }_{w}L(w_t,\theta ,t)$, we can recover the whole chain
• Need: Continuous

Random Search

• Work better than grid search

Best Arm identification

• Successive Halving(SH) Algorithm(*)

  • Per round use $B/{\log }_2(n)$, ${\log }_2(n)$ rounds totally.
  • Each round, every survivor use the same budget. Only half of them survive in each round.

• Assume $v_1>v_2\ge ...\ge v_n,{\Delta }_i=v_1-v_i$. With probability $1-\delta$, the algorithm finds the best arm with $B=O({\max }_{i>1}\frac{i}{{\Delta }_i^2}\log n\log (\log n/\delta ))$ assuming $v_i\in [0,1]$

  • Proof. Concentration inequaility, probability that $\frac{1}{3}$ of $\frac{3}{4}{N}_r$ smaller ones are greater than best one is bounded, union bound for all round.

• Hyperparameter tuning

  • $B\ge 2{\log }_2(n)\left(n+{\sum }_{i=\mathrm{2...}n}{\bar{\gamma }}^{-1}(\frac{v_i-v_1}{2})\right)$

Neural Architecture Search

• Reinforcement learning
• ProxylessNAS: each architecture has weight ${\alpha }_i$ to be chosen. Choose a binary variable, that is only one architecture exists, and the probability is about ${\alpha }_i$. $\frac{\partial L}{\partial {\alpha }_i}\approx \frac{\partial L}{\partial g_i}\frac{\partial p_i}{\partial {\alpha }_i}$, $g_i=0/1$, $\partial L/\partial g_j$ means the influence of architecture $j$ to $L$.

Machine Learning Augmented

Differencial Privacy

• Randomness is essential.

  • If we have a non-trivial deterministic algorithm, we can change data base $A$ to $B$ one by one until their output is the same. There exists a pair of databases differ by one row, then we know about that row.

• Database: histogram $N^{|X|}$ for discrete set $X$(the categories).

• $M$ is $(ϵ,\delta )$-differentially privacy if $\forall x,y\in N^{|X|},|x-y{|}_1\le 1,S\subseteq M(N^{|X|})$
  $$P(M(x)\in S)\le e^{ϵ}P(M(y)\in S)+\delta$$

• Differencial privacy is immune to post-processing. Proof:

  • For deterministic function: Proved easily by inverse function
  • Random function is convex combination of deterministic function, that is, each deterministic function is chosen with some probability.

• With $(ϵ,0)$-differential privacy(DP mechanism), the voting result will not change too much by changing one’s vote.
  $$\begin{array}{rl}{E}_{a\sim f(M(x))}[u_i(a)]& =\sum _{a\in A}{u}_i(a)\underset{f(M(x))}{\mathrm{Pr}}[a]\\ & \le \sum _{a\in A}{u}_i(a)\exp (ϵ)\underset{f(M(y))}{\mathrm{Pr}}[a]\\ & =\exp (ϵ)E_{a\sim f(M(y))}[u_i(a)]\end{array}$$
  $f$ is a map to the feature, and we only consider the expected voting utility $u_i(a)$ about the feature $a$.

• Laplace Mechanism: reach $(ϵ,0)$-DP by just adding a random gaussian noise to $f(x),x\in N^{|X|}$

  • Assume $f:N^{|X|}\to {\mathbb{R}}^k$

  • Definition: $M_L(x,f,ϵ)=f(x)+(Y_1,..,Y_k)$

    • $Y_i$ is iid random variable from $Lap(\frac{\Delta f}{ϵ})$
    • $l_1$ sensitive of $f$ is $\Delta f={\max }_{x,y,|x-y{|}_1\le 1}|f(x)-f(y){|}_1$. How sensitive $f$ is by changing one entry a little.
    • Probability density: $Lap(b)=\frac{1}{2b}\exp (-\frac{|x|}{b})$. Variance ${\sigma }^2=2b^2$.

  • Proof. The probability density of $M_L(x,f,ϵ)$ is $p_x(z)={\prod }_{i=1}^k\exp \left(-\frac{ϵ|f(x{)}_i-z_i|}{\Delta f}\right)$. Easily prove $p_x(z)/p_y(z)\le \exp (ϵ)$

  • Accuracy: for $\delta \in (0,1]$,
    $$\mathrm{Pr}\left[|f(x)-M_L(x,f,ϵ){|}_{\infty }\ge \ln \left(\frac{k}{\delta }\right)\cdot \left(\frac{\Delta f}{ϵ}\right)\right]\le \delta$$

    • Directly prove by $\mathrm{Pr}[|Y|\ge t\cdot b]=\exp (-t)$ for $Y\sim Lap(b)$

Big data

• Idea: data distribution rarely changes.
• Replace B-tree with a model. Know err_min err_max because you only care about the data in your database
  • advantage: less storage, faster lookups, more parallelism, hardware accelators
  • use linear model(fast)
  • Do exponential search since the prediction is good enough
  • Bloom Filter: is the key in my set=>maybe yes/no

Low Rank Apporximation

• $A\in {\mathbb{R}}^{n\times m}$, $[A{]}_k=\arg {\min }_{rank(A^{\prime })\le k}|A-A^{\prime }{|}_F$
• Learn $S\in {\mathbb{R}}^{p\times m}$ and do low rank decomposition for $SA$, then do some recover.

Recified flow

• Given empirical observations about distribution ${\pi }_0$ and ${\pi }_1$. Find a transport map $T$ that $T(Z)\sim {\pi }_1$ for $Z\sim {\pi }_0$.
  • E.g. ${\pi }_0$ is gaussian noise, ${\pi }_1$ are data containing pictures. We want to find a map from noise to picture(diffusion) that has some features.
  • If $T(X_0)=X_1$, then we want $v(t{X}_0+(1-t)X_1,t)=\frac{dX}{dt}=X_1-X_0$. The transformation is linear. So that one cluster will map to another clusters.
  • Minimize the loss $\mathbb{E}[\Vert (X_1-X_0)-v(X_t,t){\Vert }^2]$
• Algorithm:
  • randomly connect ${\pi }_0,{\pi }_1$
  • minimize loss of $\theta $ for $v_{\theta }$
  • connect ${\pi }_0,{\pi }_1$ by $v_{\theta }$ again, and repeat minimization.

Rope Attention

• Embed position to $q_m,k_n$ for attention by rotation in complex field
  • $\beta$-base system: focus on different accuracy level
• use inner product of $q_m,k_n$ to represent their similarity
• Softmax attention $\text{Attention}(Q,K,V)=\text{softmax}(Q{K}^{\top })V$
• Faster computation
  • linear
  • softmax separately
  • design $Q{K}^{\top }$