ν˜„λŒ€ 신경망은 μ‹œκ° 및 μ–Έμ–΄ μž‘μ—… λͺ¨λ‘μ—μ„œ 인상적인 μ„±λŠ₯을 λ‹¬μ„±ν–ˆμ§€λ§Œ, 이듀이 κ΅¬ν˜„ν•˜λŠ” κΈ°λŠ₯에 λŒ€ν•΄μ„œλŠ” μ•Œλ €μ§„ λ°”κ°€ 거의 μ—†μŠ΅λ‹ˆλ‹€. ν•œ 가지 κ°€λŠ₯성은 신경망이 μ•”λ¬΅μ μœΌλ‘œ λ³΅μž‘ν•œ μž‘μ—…μ„ μ„œλΈŒλ£¨ν‹΄μœΌλ‘œ λ‚˜λˆ„κ³ , μ΄λŸ¬ν•œ μ„œλΈŒλ£¨ν‹΄μ— λŒ€ν•œ λͺ¨λ“ˆμ‹ μ†”λ£¨μ…˜μ„ κ΅¬ν˜„ν•˜κ³ , 이λ₯Ό μž‘μ—…μ— λŒ€ν•œ 전체 μ†”λ£¨μ…˜μœΌλ‘œ κ΅¬μ„±ν•˜λŠ” κ²ƒμž…λ‹ˆλ‹€. μ΄λŸ¬ν•œ νŠΉμ„±μ„ ꡬ쑰적 ꡬ성성이라고 ν•©λ‹ˆλ‹€. 또 λ‹€λ₯Έ κ°€λŠ₯성은 μƒˆλ‘œμš΄ μž…λ ₯을 ν•™μŠ΅λœ ν…œν”Œλ¦Ώκ³Ό μΌμΉ˜μ‹œν‚€λŠ” 방법을 ν•™μŠ΅ν•˜μ—¬ μž‘μ—… λΆ„ν•΄λ₯Ό μ™„μ „νžˆ ν”Όν•  수 μžˆλ‹€λŠ” κ²ƒμž…λ‹ˆλ‹€. μ—¬κΈ°μ—μ„œλŠ” λͺ¨λΈ κ°€μ§€μΉ˜κΈ° κΈ°μˆ μ„ ν™œμš©ν•˜μ—¬ λ‹€μ–‘ν•œ μ•„ν‚€ν…μ²˜, μž‘μ—… 및 사전 ν›ˆλ ¨ 방식에 걸쳐 λΉ„μ „κ³Ό μ–Έμ–΄ λͺ¨λ‘μ—μ„œ 이 μ§ˆλ¬Έμ„ μ‘°μ‚¬ν•©λ‹ˆλ‹€. 우리의 κ²°κ³ΌλŠ” λͺ¨λΈμ΄ λ‹€λ₯Έ μ„œλΈŒλ„€νŠΈμ›Œν¬μ˜ κΈ°λŠ₯을 μœ μ§€ν•˜λ©΄μ„œ 제거될 수 μžˆλŠ” λͺ¨λ“ˆμ‹ μ„œλΈŒλ„€νŠΈμ›Œν¬λ₯Ό 톡해 μ„œλΈŒλ£¨ν‹΄μ— λŒ€ν•œ μ†”λ£¨μ…˜μ„ κ΅¬ν˜„ν•˜λŠ” κ²½μš°κ°€ λ§Žλ‹€λŠ” 것을 λ³΄μ—¬μ€λ‹ˆλ‹€.Β μ΄λŠ” 신경망이 ꡬ성성을 ν•™μŠ΅ν•˜μ—¬ νŠΉμˆ˜ν•œ 상징적 λ©”μ»€λ‹ˆμ¦˜μ˜ ν•„μš”μ„±μ„ μ œκ±°ν•  수 μžˆμŒμ„ μ‹œμ‚¬ν•©λ‹ˆλ‹€.



https://arxiv.org/abs/2301.10884

Break It Down: Evidence for Structural Compositionality in Neural NetworksThough modern neural networks have achieved impressive performance in both vision and language tasks, we know little about the functions that they implement. One possibility is that neural networks implicitly break down complex tasks into subroutines, implement modular solutions to these subroutines, and compose them into an overall solution to a task - a property we term structural compositionality. Another possibility is that they may simply learn to match new inputs to learned templates, eliding task decomposition entirely. Here, we leverage model pruning techniques to investigate this question in both vision and language across a variety of architectures, tasks, and pretraining regimens. Our results demonstrate that models often implement solutions to subroutines via modular subnetworks, which can be ablated while maintaining the functionality of other subnetworks. This suggests that neural networks may be able to learn compositionality, obviating the need for specialized symbolic mechanisms.arxiv.org