arXiv 2021
Jun-Hyun Bae, Inchul Choi, Minho Lee
Kyungpook National University
๐Ÿ“„ Paper

Abstract

Empirical Risk Minimization (ERM) based machine learning algorithms have suffered from weak generalization performance on data obtained from out-of-distribution (OOD). To address this problem, Invariant Risk Minimization (IRM) objective was suggested to find invariant optimal predictor which is less affected by the changes in data distribution. However, even with such progress, IRMv1, the practical formulation of IRM, still shows performance degradation when there are not enough training data, and even fails to generalize to OOD, if the number of spurious correlations is larger than the number of environments. In this paper, to address such problems, we propose a novel meta-learning based approach for IRM. In this method, we do not assume the linearity of classifier for the ease of optimization, and solve ideal bi-level IRM objective with Model-Agnostic Meta-Learning (MAML) framework. Our method is more robust to the data with spurious correlations and can provide an invariant optimal classifier even when data from each distribution are scarce. In experiments, we demonstrate that our algorithm not only has better OOD generalization performance than IRMv1 and all IRM variants, but also addresses the weakness of IRMv1 with improved stability.


Overview

IRMv1์—์„œ๋Š” gradient penalty๊ฐ€ ์ž‘์•„์ ธ๋„ OoD ์„ฑ๋Šฅ์ด ๋‚ฎ์„ ์ˆ˜ ์žˆ๋‹ค. ์ด ๋…ผ๋ฌธ์€ ํ™˜๊ฒฝ๋ณ„ ๋ฐ์ดํ„ฐ๊ฐ€ ์ ๊ฑฐ๋‚˜ ์—ฌ๋Ÿฌ spurious feature๊ฐ€ ์žˆ๋Š” ์กฐ๊ฑด์—์„œ ๊ทธ ๋ฌธ์ œ๋ฅผ ์‚ดํŽด๋ณด๊ณ , IRM์˜ bi-level ๊ตฌ์กฐ๋ฅผ ๋ฉ”ํƒ€๋Ÿฌ๋‹์œผ๋กœ ๊ตฌํ˜„ํ•œ๋‹ค.

meta-IRM์€ ํ•œ training environment์—์„œ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ ์‘์‹œํ‚จ ๋’ค, ๋‹ค๋ฅธ training environment์˜ ์†์‹ค๋กœ ์›๋ž˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์—…๋ฐ์ดํŠธํ•œ๋‹ค. MAML์˜ inner/outer loop ์‚ฌ์ด์—์„œ ํ™˜๊ฒฝ์„ ๋ฐ”๊พธ๋Š” ๊ตฌ์„ฑ์ด๋‹ค. Colored MNIST์—์„œ ๋‘ loop์— ๊ฐ™์€ ํ™˜๊ฒฝ์„ ๋ฐฐ์ •ํ•œ ๋ณ€ํ˜•์˜ OoD ์ •ํ™•๋„๋Š” 13.6%, ์„œ๋กœ ๋‹ค๋ฅธ ํ™˜๊ฒฝ์„ ์“ฐ๋Š” ์ „์ฒด ๋ฐฉ๋ฒ•์€ 70.4%์˜€๋‹ค.

Meta-IRM Diagram

meta-IRM์˜ ํ•™์Šต ์•„์ด๋””์–ด๋ฅผ ๋‚˜ํƒ€๋‚ธ ๊ฐœ๋…๋„. ํ•œ environment์— ์ ์‘ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ($\theta'_1$, $\theta'_2$)๋ฅผ ๋‹ค๋ฅธ environment์—์„œ ํ‰๊ฐ€ํ•˜๊ณ , ๊ทธ gradient๋ฅผ ํ•ฉ์ณ ํ™˜๊ฒฝ์— ๋œ ์˜์กดํ•˜๋Š” ํ•ด๋ฅผ ์ฐพ๋„๋ก ์œ ๋„ํ•œ๋‹ค. ๊ทธ๋ฆผ์˜ $\theta_{optimal}$์€ ๋ชฉํ‘œ๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๋ชจ๋“  ์กฐ๊ฑด์—์„œ ๊ทธ ํ•ด์— ๋„๋‹ฌํ•œ๋‹ค๋Š” ์ˆ˜๋ ด ๋ณด์žฅ์„ ๋œปํ•˜์ง€ ์•Š๋Š”๋‹ค.


Method

IRM์€ ์—ฌ๋Ÿฌ training environment์—์„œ ๊ณตํ†ต์œผ๋กœ ์ตœ์ ์ธ classifier๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ํ‘œํ˜„์„ ์ฐพ๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ๋‹ค. ์›๋ž˜ ๋ชฉํ‘œ์—๋Š” ํ™˜๊ฒฝ๋ณ„ ์ตœ์ ํ™”๊ฐ€ ์ œ์•ฝ์œผ๋กœ ๋“ค์–ด ์žˆ์–ด ํ’€๊ธฐ ์–ด๋ ต๋‹ค. ์‹ค์šฉ์ ์ธ IRMv1์€ classifier๋ฅผ ๊ณ ์ •๋œ ์Šค์นผ๋ผ๋กœ ๋‘๊ณ  ์ด ์ œ์•ฝ์„ gradient penalty๋กœ ๊ทผ์‚ฌํ•œ๋‹ค. ํ•˜์ง€๋งŒ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ํ™˜๊ฒฝ ์ˆ˜์— ๋น„ํ•ด spurious feature๊ฐ€ ๋งŽ์€ ์กฐ๊ฑด์—์„œ๋Š”, penalty๊ฐ€ ์ž‘์•„๋„ OoD ์„ฑ๋Šฅ์ด ๋‚˜์  ์ˆ˜ ์žˆ๋‹ค. ์•„๋ž˜์˜ ๋ฐ์ดํ„ฐ ๊ฐ์†Œ์™€ ๋ณต์ˆ˜ ํŽธํ–ฅ ์‹คํ—˜์ด ์ด๋Ÿฐ ๊ฒฝ์šฐ๋ฅผ ์‚ดํŽด๋ณธ๋‹ค.

meta-IRM์€ IRM์˜ bi-level ๊ตฌ์กฐ๋ฅผ MAML ๊ธฐ๋ฐ˜ ํ•™์Šต ์ ˆ์ฐจ๋กœ ๊ตฌํ˜„ํ•œ๋‹ค. ํ•ต์‹ฌ ์ฐฉ์•ˆ์ ์€ inner loop์™€ outer loop์— ์„œ๋กœ ๋‹ค๋ฅธ training environment๋ฅผ ๋ฐฐ์ •ํ•˜๋Š” ๊ฒƒ์ด๋‹ค. ํ•œ ํ™˜๊ฒฝ์— ์ ์‘ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค๋ฅธ ํ™˜๊ฒฝ์—์„œ ํ‰๊ฐ€ํ•ด, ํ™˜๊ฒฝ์„ ๋„˜์–ด ์œ ํšจํ•œ ํŠน์ง•์„ ํ•™์Šตํ•˜๋„๋ก ์œ ๋„ํ•œ๋‹ค. ์›๋ž˜์˜ ์ œ์•ฝ ์ตœ์ ํ™” ๋ฌธ์ œ๋ฅผ ์ •ํ™•ํžˆ ํ’€๊ฑฐ๋‚˜ ์ „์—ญ ์ตœ์  ํ•ด๋กœ ์ˆ˜๋ ดํ•œ๋‹ค๋Š” ๋ณด์žฅ์„ ์ œ์‹œํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋‹ค.

Meta-Learning Framework

MAML์˜ inner/outer ๊ตฌ์กฐ๋ฅผ IRM์˜ bi-level ๊ตฌ์กฐ์— ๋Œ€์‘์‹œํ‚จ๋‹ค. Inner loop๋Š” ํ™˜๊ฒฝ๋ณ„ ์ ์‘์„, outer loop๋Š” ๋‹ค๋ฅธ ํ™˜๊ฒฝ๊นŒ์ง€ ๊ณ ๋ คํ•œ risk ๊ฐ์†Œ๋ฅผ ๋‹ด๋‹นํ•œ๋‹ค. ์‹ค์ œ inner update๋Š” classifier๋งŒ ๊ณ ์ •์ ์œผ๋กœ ๊ณจ๋ผ ๊ฐฑ์‹ ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ \(\theta\)์— ์ ์šฉํ•œ๋‹ค. IRMv1์ฒ˜๋Ÿผ classifier๋ฅผ ๊ณ ์ •๋œ ์„ ํ˜• ํ•จ์ˆ˜๋กœ ์ œํ•œํ•˜์ง€ ์•Š๋Š”๋‹ค.

Inner optimization: ๊ฐ training environment \(e_i\)์— ๋Œ€ํ•ด ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ ์‘์‹œํ‚จ๋‹ค:

\[\theta_i' = \theta - \alpha \nabla_\theta R^{e_i}(f_\theta)\]

Meta-optimization: ์ ์‘๋œ ํŒŒ๋ผ๋ฏธํ„ฐ \(\theta_i'\)๋กœ ๋‹ค๋ฅธ environment \(e_j\)์—์„œ์˜ meta-loss๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ \(\theta\)๋ฅผ ์—…๋ฐ์ดํŠธํ•œ๋‹ค:

\[\theta \leftarrow \theta - \beta \nabla_\theta \left\{ \sum_i \sum_j R^{e_j}(f_{\theta_i'}) + \lambda \sigma \right\}, \quad e_j \sim \mathcal{E}_{tr} \setminus e_i\]

Learning Process

meta-IRM ํ•™์Šต ๊ณผ์ •. ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ $\theta$๊ฐ€ inner optimization์—์„œ ๊ฐ training environment์— ์ ์‘๋˜๊ณ , ์ ์‘๋œ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๋‹ค๋ฅธ environment์—์„œ meta-loss๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค.

Auxiliary Standard Deviation Loss

Meta-loss์˜ ํ‘œ์ค€ํŽธ์ฐจ \(\sigma\)๋ฅผ ๋ณด์กฐ ์†์‹ค๋กœ ์ถ”๊ฐ€ํ•˜์—ฌ, ๋ชจ๋“  environment์—์„œ ๊ท ์ผํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋„๋ก ์œ ๋„ํ•œ๋‹ค. ์ด ์ •๊ทœํ™”๋Š” ํ•™์Šต ์•ˆ์ •์„ฑ์„ ๋†’์ด๊ณ  ์ˆ˜๋ ด ์†๋„๋ฅผ ๊ฐœ์„ ํ•œ๋‹ค. \(\lambda\)๋Š” ์ด ๋ณด์กฐ ์†์‹ค์˜ ๊ฐ€์ค‘์น˜๋‹ค.


Results

Colored MNIST

์ˆซ์ž 0โ€“4์™€ 5โ€“9๋ฅผ ๋‘ ํด๋ž˜์Šค๋กœ ๋‚˜๋ˆ„๊ณ , ๋ ˆ์ด๋ธ”์„ 25% ํ™•๋ฅ ๋กœ ๋’ค์ง‘์–ด(label noise) ์ˆซ์ž ๋ชจ์–‘๋งŒ์œผ๋กœ๋Š” 75%๊นŒ์ง€๋งŒ ๋งžํž ์ˆ˜ ์žˆ๊ฒŒ ๋งŒ๋“ ๋‹ค. ์ƒ‰์€ ๋ ˆ์ด๋ธ”์„ ํ™•๋ฅ  \(p_e\)๋กœ ๋’ค์ง‘์–ด ์ •ํ•˜๋ฏ€๋กœ, training environment(\(p_e=0.1, 0.2\))์—์„œ๋Š” ์ƒ‰์ด ๋ ˆ์ด๋ธ”๊ณผ ๊ฐ๊ฐ 90%, 80% ์ผ์น˜ํ•˜๊ณ  test environment(\(p_e=0.9\))์—์„œ๋Š” ์ด ๊ด€๊ณ„๊ฐ€ ๋’ค์ง‘ํžŒ๋‹ค. ๋”ฐ๋ผ์„œ ์ƒ‰์— ๊ธฐ๋Œ€์ง€ ์•Š๋Š” ์ด์ƒ์ ์ธ ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋Š” 75%๋‹ค.

Test accuracy (%) on Colored MNIST. $p_e=0.9$๊ฐ€ OOD ํ™˜๊ฒฝ (color-label ์ƒ๊ด€๊ด€๊ณ„ ์—ญ์ „).

Algorithm\(p_e=0.1\)\(p_e=0.2\)\(p_e=0.9\) (OOD)
ERM88.6 ยฑ 0.379.7 ยฑ 0.616.4 ยฑ 0.8
IRMv171.4 ยฑ 0.970.8 ยฑ 1.066.9 ยฑ 2.5
MM-REx70.8 ยฑ 1.570.4 ยฑ 2.066.1 ยฑ 4.9
V-REx71.5 ยฑ 0.871.1 ยฑ 0.968.6 ยฑ 1.2
meta-IRM (Ours)70.9 ยฑ 0.970.8 ยฑ 1.070.4 ยฑ 0.9
Random505050
ERM (grayscale)72.6 ยฑ 0.372.7 ยฑ 0.373.0 ยฑ 0.5
Optimal757575

์ƒ‰์ด ํฌํ•จ๋œ ์ž…๋ ฅ์„ ๋ฐ›๋Š” ๋น„๊ต ๋ฐฉ๋ฒ• ์ค‘ meta-IRM(70.4%)์ด ๊ฐ€์žฅ ๋†’๊ณ , V-REx(68.6%), IRMv1(66.9%), MM-REx(66.1%)๋ณด๋‹ค ์‹คํ–‰ ๊ฐ„ ํ‘œ์ค€ํŽธ์ฐจ๋„ ์ž‘๋‹ค. ์ƒ‰์„ ๋ฏธ๋ฆฌ ์ œ๊ฑฐํ•œ ERM (grayscale)์€ ๋ณ„๋„์˜ ์ฐธ์กฐ๊ฐ’์œผ๋กœ 73.0%๋‹ค. ์•„๋ž˜ ๊ทธ๋ฆผ์—์„œ ์ƒ‰โ€“๋ ˆ์ด๋ธ” ๊ด€๊ณ„๋ฅผ ๋ฐ”๊พธ์–ด๋„ meta-IRM์˜ ์ •ํ™•๋„๊ฐ€ ๊ฑฐ์˜ ์ผ์ •ํ•œ ๊ฒƒ์€ ์ƒ‰์— ๋Œ€ํ•œ ์˜์กด๋„๊ฐ€ ๋‚ฎ๋‹ค๋Š” ํ•ด์„์„ ๋’ท๋ฐ›์นจํ•œ๋‹ค.

Various Test Environments

๋‹ค์–‘ํ•œ test environment ($p_e$ 0.1~0.9)์—์„œ์˜ ์ •ํ™•๋„.

Multi-Class Problem

ํด๋ž˜์Šค ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•˜๋ฉด spurious correlation์˜ ๊ตฌ์กฐ๊ฐ€ ๋ณต์žกํ•ด์ง€๋ฉฐ, meta-IRM์˜ ์ด์ ์ด ๋”์šฑ ๋‘๋“œ๋Ÿฌ์ง„๋‹ค.

Multi-class Colored MNIST ($k$=5, 10)์˜ train/test accuracy (%). ERM, IRMv1, ERM (grayscale) ์ˆ˜์น˜๋Š” Choe et al. (2020)์—์„œ ์ธ์šฉ.

Algorithm# ClassesTrainTest (OOD)
ERM595.2 ยฑ 0.241.0 ยฑ 0.6
IRMv1582.2 ยฑ 0.462.0 ยฑ 2.4
meta-IRM (Ours)576.4 ยฑ 1.474.0 ยฑ 3.6
Random52020
ERM (grayscale)573.2 ยฑ 0.271.7 ยฑ 0.4
ERM1092.6 ยฑ 0.239.2 ยฑ 0.9
IRMv11083.4 ยฑ 0.558.6 ยฑ 2.5
meta-IRM (Ours)1079.5 ยฑ 0.673.4 ยฑ 3.2
Random101010
ERM (grayscale)1073.2 ยฑ 0.171.9 ยฑ 0.5
Optimalโ€”7575

10 ํด๋ž˜์Šค ์„ค์ •์—์„œ meta-IRM์˜ OoD ์ •ํ™•๋„๋Š” 73.4%๋กœ IRMv1(58.6%)๋ณด๋‹ค 14.8%p ๋†’๋‹ค. ERM (grayscale) ์ฐธ์กฐ๊ฐ’์€ 71.9%๋‹ค. 5ยท10 ํด๋ž˜์Šค ๋ชจ๋‘์—์„œ meta-IRM์˜ ํ‰๊ท  ์ •ํ™•๋„๊ฐ€ ๋†’์ง€๋งŒ, ์‹คํ–‰ ๊ฐ„ ํ‘œ์ค€ํŽธ์ฐจ๋„ 2-class ์„ค์ •(0.9)๋ณด๋‹ค ์ปค์ง„๋‹ค(3.6, 3.2).

Insufficient Data

Training data๊ฐ€ ์ ˆ๋ฐ˜(12,500)์œผ๋กœ ์ค„์—ˆ์„ ๋•Œ, IRM ๊ณ„์—ด ๋ฐฉ๋ฒ•๋“ค์˜ ์„ฑ๋Šฅ์ด ์–ผ๋งˆ๋‚˜ ์œ ์ง€๋˜๋Š”์ง€ ํ‰๊ฐ€ํ•œ๋‹ค.

Training data ๊ฐ์†Œ ์‹œ test accuracy (%). 25,000 โ†’ 12,500 per environment.

Algorithm\(p_e=0.1\)\(p_e=0.2\)\(p_e=0.9\) (OOD)
IRMv173.2 ยฑ 1.671.7 ยฑ 1.358.5 ยฑ 2.5
V-REx75.0 ยฑ 0.873.4 ยฑ 0.661.3 ยฑ 1.4
meta-IRM (Ours)70.6 ยฑ 1.570.5 ยฑ 1.668.3 ยฑ 2.3
Optimal757575

๋ฐ์ดํ„ฐ๊ฐ€ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์—ˆ์„ ๋•Œ, IRMv1์€ 66.9% โ†’ 58.5%๋กœ 8.4%p ํ•˜๋ฝํ•˜๋Š” ๋ฐ˜๋ฉด, meta-IRM์€ 70.4% โ†’ 68.3%๋กœ 2.1%p ํ•˜๋ฝํ•œ๋‹ค. ์ด ์„ค์ •์—์„œ๋Š” meta-IRM์ด ๋ฐ์ดํ„ฐ ๊ฐ์†Œ์— ๋”ฐ๋ฅธ OoD ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ๋” ์ž˜ ๊ฒฌ๋”˜๋‹ค.

Data Decrease

Training data ๊ฐ์†Œ์— ๋”ฐ๋ฅธ ์˜ํ–ฅ (25,000 โ†’ 12,500 โ†’ 6,250). ์œ„๋ถ€ํ„ฐ IRMv1, V-REx, meta-IRM. ๋ฐ์ดํ„ฐ๊ฐ€ ์ค„์ˆ˜๋ก IRMv1๊ณผ V-REx๋Š” ERM์ฒ˜๋Ÿผ $p_e$๊ฐ€ ์ปค์งˆ์ˆ˜๋ก ์ •ํ™•๋„๊ฐ€ ๊ธ‰๋ฝํ•˜์ง€๋งŒ, meta-IRM์€ ๊ฑฐ์˜ ํ‰ํƒ„ํ•˜๊ฒŒ ์œ ์ง€๋œ๋‹ค.

Two Spurious Features

Spurious correlation์ด 2๊ฐœ(color + patch)์ด๊ณ  training environment๋„ 2๊ฐœ์ธ ๊ฒฝ์šฐ๋‹ค. IRMv1์˜ ์ด๋ก ์  ๋ณด์žฅ์€ environment ์ˆ˜๊ฐ€ spurious feature ์ˆ˜๋ณด๋‹ค ๋งŽ์„ ๋•Œ์—๋งŒ ์„ฑ๋ฆฝํ•˜๋ฏ€๋กœ, ๋‘˜์ด ๊ฐ™์€ ์ด ์„ค์ •์—์„œ๋Š” ์„ฑ๋ฆฝํ•˜์ง€ ์•Š๋Š”๋‹ค.

Algorithm\(p_e=0.1\)\(p_e=0.2\)\(p_e=0.9\) (OOD)
IRMv193.5 ยฑ 0.286.4 ยฑ 0.313.4 ยฑ 0.3
V-REx93.6 ยฑ 0.486.3 ยฑ 0.313.5 ยฑ 0.3
meta-IRM (Ours)58.1 ยฑ 3.156.8 ยฑ 2.954.5 ยฑ 4.0
Optimal757575

IRMv1๊ณผ V-REx๋Š” ํ•™์Šต ํ™˜๊ฒฝ๊ณผ ๊ฐ™์€ ์ƒ๊ด€๊ด€๊ณ„์—์„œ๋Š” ์ •ํ™•๋„๊ฐ€ ๋†’์ง€๋งŒ, ๋‘ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์—ญ์ „๋œ OoD ํ™˜๊ฒฝ์—์„œ๋Š” 13.4โ€“13.5%๋กœ ๋–จ์–ด์ง„๋‹ค. meta-IRM์€ 54.5 ยฑ 4.0%๋กœ ์ด ์‹คํŒจ๋ฅผ ์™„ํ™”ํ•œ๋‹ค. ๋‹ค๋งŒ ํ‰๊ท ์ด ๋žœ๋ค ๊ธฐ์ค€(50%)๋ณด๋‹ค ์กฐ๊ธˆ ๋†’์€ ์ˆ˜์ค€์ด๊ณ  ์ด์ƒ์  ์ •ํ™•๋„(75%)์™€๋„ ์ฐจ์ด๊ฐ€ ์ปค, ๋ณต์ˆ˜ ํŽธํ–ฅ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ–ˆ๋‹ค๊ณ  ๋ณด๊ธฐ๋Š” ์–ด๋ ต๋‹ค.

Two Spurious Features

๋‘ ๊ฐ€์ง€ spurious feature (color + patch) ์กด์žฌ ์‹œ ๊ฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ test environment๋ณ„ ์ •ํ™•๋„.

PunctuatedSST-2 (NLP)

Colored MNIST์—์„œ ํ™•์ธํ•œ ํšจ๊ณผ๊ฐ€ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ์—์„œ๋„ ๋‚˜ํƒ€๋‚˜๋Š”์ง€๋ฅผ PunctuatedSST-2๋กœ ๊ฒ€์ฆํ•œ๋‹ค. ๊ฐ์ • ๋ถ„์„ ํƒœ์Šคํฌ์—์„œ ๊ตฌ๋‘์ ("!" vs “.")์ด spurious feature๋กœ ์ž‘์šฉํ•˜๋Š” ์„ค์ •์ด๋‹ค.

PunctuatedSST-2 test accuracy (%). $\eta_e$๋Š” label noise ๋น„์œจ.

Algorithm\(\eta_e\)Test (OOD)
ERM0.2530.7 ยฑ 1.5
IRMv10.2562.0 ยฑ 1.9
meta-IRM (Ours)0.2562.2 ยฑ 1.8
ERM (vanilla)0.2562.3 ยฑ 0.5
Optimal0.2575
ERM056.2 ยฑ 2.9
IRMv1067.4 ยฑ 1.4
meta-IRM (Ours)073.0 ยฑ 0.7
ERM (vanilla)076.7 ยฑ 2.7
Optimal0100

Label noise๊ฐ€ ์—†๋Š” ์„ค์ •(\(\eta_e=0\))์—์„œ meta-IRM(73.0%)์€ IRMv1(67.4%)์„ 5.6%p ์ƒํšŒํ•˜๋ฉฐ, ๊ตฌ๋‘์  spurious feature๊ฐ€ ์—†๋Š” vanilla SST-2์—์„œ ํ•™์Šตํ•œ ERM (vanilla) ์ฐธ์กฐ๊ฐ’(76.7%)์— ๊ทผ์ ‘ํ•œ๋‹ค. Label noise๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ(\(\eta_e=0.25\))์—๋Š” meta-IRM(62.2%)๊ณผ IRMv1(62.0%) ๋ชจ๋‘ ๊ตฌ๋‘์  ์—†๋Š” SST-2๋กœ ํ•™์Šตํ•œ ERM (vanilla, 62.3%) ์ˆ˜์ค€์— ๋„๋‹ฌํ•ด, ์ด ์„ค์ •์—์„œ๋Š” ๋‘ ๋ฐฉ๋ฒ•์˜ ์ฐจ์ด๊ฐ€ ๊ฑฐ์˜ ์—†๋‹ค.

Ablation Study

Variant\(p_e=0.1\)\(p_e=0.9\) (OOD)
w/o std. loss73.0 ยฑ 0.864.2 ยฑ 3.2
First-order approx.62.6 ยฑ 5.559.1 ยฑ 6.3
Same env. (inner = meta)89.3 ยฑ 1.613.6 ยฑ 6.6
meta-IRM (Full)70.9 ยฑ 0.970.4 ยฑ 0.9
Optimal7575

Same environment(inner = meta) ๋ณ€ํ˜•์€ 13.6%๋กœ ERM ์ˆ˜์ค€์˜ ์‹คํŒจ๋ฅผ ๋ณด์ธ๋‹ค. Inner loop์™€ outer loop์— ์„œ๋กœ ๋‹ค๋ฅธ environment๋ฅผ ๋ฐฐ์ •ํ•˜๋Š” ๊ฒƒ์ด meta-IRM์˜ ํ•ต์‹ฌ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ž„์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒฐ๊ณผ๋‹ค. First-order approximation์€ second-order gradient ์ •๋ณด์˜ ์†์‹ค๋กœ ์ธํ•ด ์„ฑ๋Šฅ๊ณผ ์•ˆ์ •์„ฑ์ด ๋ชจ๋‘ ์ €ํ•˜๋œ๋‹ค(59.1 ยฑ 6.3%). ํ‘œ์ค€ํŽธ์ฐจ loss๋ฅผ ์ œ๊ฑฐํ•˜๋ฉด OOD ์„ฑ๋Šฅ์ด 6.2%p ๋–จ์–ด์ง€๊ณ  ์‹คํ–‰ ๊ฐ„ ํ‘œ์ค€ํŽธ์ฐจ๋„ ์ปค์ง„๋‹ค(3.2 vs 0.9).