$$ \newcommand{\floor}[1]{\left\lfloor #1 \right\rfloor} \newcommand{\ceil}[1]{\left\lceil #1 \right\rceil} \newcommand{\N}{\mathbb{N}} \newcommand{\R}{\mathbb{R}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\Q}{\mathbb{Q}} \newcommand{\C}{\mathbb{C}} \renewcommand{\L}{\mathcal{L}} \newcommand{\x}{\times} \newcommand{\contra}{\scalebox{1.5}{$\lightning$}} \newcommand{\inner}[2]{\left\langle #1 , #2 \right\rangle} \newcommand{\st}{\text{ such that }} \newcommand{\for}{\text{ for }} \newcommand{\Setcond}[2]{ \left\{\, #1 \mid #2 \, \right\}} \newcommand{\setcond}[2]{\Setcond{#1}{#2}} \newcommand{\seq}[1]{ \left\langle #1 \right\rangle} \newcommand{\Set}[1]{ \left\{ #1 \right\}} \newcommand{\set}[1]{ \Set{#1} } \newcommand{\sgn}{\text{sign}} \newcommand{\halfline}{\vspace{0.5em}} \newcommand{\diag}{\text{diag}} \newcommand{\legn}[2]{\left(\frac{#1}{#2}\right)} \newcommand{\ord}{\text{ord}} \newcommand{\di}{\mathrel{|}} \newcommand{\gen}[1] \newcommand{\irr}{\mathrm{irr }} \renewcommand{\deg}{\mathrm{deg }} \newcommand{\nsgeq}{\trianglelefteq} \newcommand{\nsg}{\triangleleft} \newcommand{\argmin}{\mathrm{argmin}} \newcommand{\argmax}{\mathrm{argmax}} \newcommand{\minimize}{\mathrm{minimize}} \newcommand{\maximize}{\mathrm{maximize}} \newcommand{\subto}{\mathrm{subject\ to}} \newcommand{\DKL}[2]{D_{\mathrm{KL}}\left(#1 \di\di #2\right)} \newcommand{\ReLU}{\mathrm{ReLU}} \newcommand{\E}{\mathsf{E}} \newcommand{\V}{\mathsf{Var}} \newcommand{\Corr}{\mathsf{Corr}} \newcommand{\Cov}{\mathsf{Cov}} \newcommand{\covariance}[1]{\Cov\left(#1\right)} \newcommand{\variance}[1]{\V\left[#1\right]} \newcommand{\variancewith}[1]{\V\left[#1\right]} \newcommand{\expect}[1]{\E\left[#1\right]} \newcommand{\expectwith}[2]{\E_{#1}\left[#2\right]} \renewcommand{\P}{\mathsf{P}} \newcommand{\uniform}[2]{\mathrm{Uniform}\left(#1 \dots #2\right)} \newcommand{\gdist}[2]{\mathcal{N}\left(#1, #2\right)} \DeclarePairedDelimiter{\norm}{\lVert}{\rVert} $$ \everymath{\displaystyle}

TA-GATES - An Encoding Scheme for Neural Network Architectures | Hyunju Kim

TA-GATES - An Encoding Scheme for Neural Network Architectures

Xuefei Ning et al. / 2022 / NeurIPS

July 25, 2023

[Abstract]

introduce Training-Analogous Graph-based ArchiTecture Encoding Scheme (TA-GATES)

Untitled

SOTA algorithm: graph-based encodings
- same operation → same embedding
⇒ architecture is not a DAG with fixed computations, but trainable operations
contributions
1. conducts iterative process of forward and backward passes on the DAG and uses the forward output at the final time step as the architecture encoding
2. adopt “symmetric-breaking” technique

[3.1 Iterative Encoding in Analogy to Iterative Parameter Tuning]

get embedding emb = GetEmbedding(O, op)

[3.2 Symmetry Breaking in Analogy to Random Parameter Initialization]

Symmetric한 case에 대해 초기값을 다르게 설정

[4. Experiments]

Untitled