notesum.ai

Published at November 4

Double Descent Meets Out-of-Distribution Detection: Theoretical Insights and Empirical Analysis on the role of model complexity

stat.ML

cs.AI

cs.CV

cs.LG

math.ST

stat.TH

Released Date: November 4, 2024

Authors: Mouïn Ben Ammar¹, David Brellmann², Arturo Mendoza¹, Antoine Manzanera¹, Gianni Franchi¹

Aff.: ¹U2IS Lab ENSTA Paris, Palaiseau, FRANCE; ²SafranTech, Chateaufort 78117, FRANCE

Arxiv: http://arxiv.org/abs/2411.02184v1

Model	$NC1_{u/o}$	Method	SUN		Places365		CIFAR-100
Model	$NC1_{u/o}$	Method	$\texttt{AUC}_{u}$ $\uparrow$	$\texttt{AUC}_{o}$ $\uparrow$	$\texttt{AUC}_{u}$ $\uparrow$	$\texttt{AUC}_{o}$ $\uparrow$	$\texttt{AUC}_{u}$ $\uparrow$	$\texttt{AUC}_{o}$ $\uparrow$
CNN	0.88	Softmax score	76.09 $\pm$ 0.96	75.08 $\pm$ 0.75	75.95 $\pm$ 0.76	74.59 $\pm$ 0.45	74.33 $\pm$ 0.32	72.68 $\pm$ 0.31
		MaxLogit	72.98 $\pm$ 2.22	60.13 $\pm$ 0.35	73.33 $\pm$ 2.17	61.25 $\pm$ 0.42	73.38 $\pm$ 0.39	70.37 $\pm$ 0.34
		Energy	68.08 $\pm$ 3.66	59.73 $\pm$ 0.36	69.00 $\pm$ 3.56	60.90 $\pm$ 0.43	70.78 $\pm$ 0.73	70.24 $\pm$ 0.35
		Energy+ReAct	59.12 $\pm$ 4.73	47.49 $\pm$ 0.58	60.79 $\pm$ 4.53	49.45 $\pm$ 0.52	66.00 $\pm$ 1.58	63.57 $\pm$ 0.51
		NECO	70.43 $\pm$ 2.53	64.22 $\pm$ 1.36	71.20 $\pm$ 2.59	63.59 $\pm$ 0.98	72.40 $\pm$ 0.95	70.17 $\pm$ 0.76
		ViM	59.94 $\pm$ 2.01	59.77 $\pm$ 0.66	61.88 $\pm$ 1.89	60.93 $\pm$ 0.40	69.23 $\pm$ 0.78	70.25 $\pm$ 0.35
		ASH-P	68.60 $\pm$ 3.59	60.36 $\pm$ 0.37	69.35 $\pm$ 3.50	61.48 $\pm$ 0.43	71.11 $\pm$ 0.53	70.45 $\pm$ 0.41
ResNet	1.96	Softmax score	71.18 $\pm$ 0.93	75.82 $\pm$ 0.89	71.22 $\pm$ 0.93	75.52 $\pm$ 0.88	71.21 $\pm$ 0.48	75.37 $\pm$ 0.42
		MaxLogit	70.64 $\pm$ 1.53	72.51 $\pm$ 1.03	70.69 $\pm$ 1.29	72.64 $\pm$ 0.94	69.76 $\pm$ 0.39	73.65 $\pm$ 0.38
		Energy	69.11 $\pm$ 2.49	72.46 $\pm$ 1.03	69.19 $\pm$ 2.08	72.59 $\pm$ 0.94	67.58 $\pm$ 0.46	73.61 $\pm$ 0.39
		Energy+ReAct	69.57 $\pm$ 2.35	71.83 $\pm$ 0.88	69.63 $\pm$ 1.93	71.97 $\pm$ 0.78	67.25 $\pm$ 0.91	72.63 $\pm$ 0.45
		NECO	70.39 $\pm$ 2.30	75.60 $\pm$ 1.56	70.46 $\pm$ 1.85	75.20 $\pm$ 1.42	69.92 $\pm$ 0.36	75.28 $\pm$ 0.50
		ViM	66.54 $\pm$ 1.99	74.44 $\pm$ 0.65	65.38 $\pm$ 1.99	73.42 $\pm$ 0.65	64.61 $\pm$ 0.47	71.54 $\pm$ 0.44
		ASH-P	69.11 $\pm$ 2.49	71.73 $\pm$ 1.09	69.19 $\pm$ 2.08	71.85 $\pm$ 0.98	67.58 $\pm$ 0.46	72.89 $\pm$ 0.35
Swin	1.70	Softmax score	58.82 $\pm$ 2.98	67.91 $\pm$ 0.69	59.01 $\pm$ 2.90	67.66 $\pm$ 0.59	61.89 $\pm$ 1.42	65.44 $\pm$ 0.62
		MaxLogit	59.91 $\pm$ 3.11	70.75 $\pm$ 0.45	59.84 $\pm$ 3.40	70.46 $\pm$ 0.46	61.79 $\pm$ 1.73	66.95 $\pm$ 0.53
		Energy	60.12 $\pm$ 3.68	70.79 $\pm$ 0.42	58.77 $\pm$ 3.98	70.50 $\pm$ 0.44	55.52 $\pm$ 2.10	66.92 $\pm$ 0.51
		Energy+ReAct	60.16 $\pm$ 3.79	71.15 $\pm$ 0.44	58.85 $\pm$ 4.03	70.83 $\pm$ 0.48	55.53 $\pm$ 2.10	67.27 $\pm$ 0.52
		NECO	64.26 $\pm$ 3.20	73.29 $\pm$ 0.75	63.88 $\pm$ 3.37	72.38 $\pm$ 0.66	62.62 $\pm$ 2.10	68.13 $\pm$ 0.76
		ViM	60.68 $\pm$ 2.61	71.69 $\pm$ 0.19	58.34 $\pm$ 2.60	71.39 $\pm$ 0.15	55.95 $\pm$ 0.77	68.89 $\pm$ 0.51
		ASH-P	59.49 $\pm$ 4.21	70.74 $\pm$ 0.44	58.15 $\pm$ 4.42	70.42 $\pm$ 0.40	55.10 $\pm$ 2.28	66.89 $\pm$ 0.55
ViT	2.32	Softmax score	66.28 $\pm$ 0.19	64.87 $\pm$ 0.27	66.26 $\pm$ 0.36	64.61 $\pm$ 0.26	65.18 $\pm$ 0.38	62.96 $\pm$ 0.33
		MaxLogit	66.09 $\pm$ 1.48	70.30 $\pm$ 0.46	66.13 $\pm$ 1.50	69.79 $\pm$ 0.26	64.60 $\pm$ 0.35	66.69 $\pm$ 0.39
		Energy	64.79 $\pm$ 2.81	70.50 $\pm$ 0.48	64.86 $\pm$ 2.65	69.98 $\pm$ 0.26	63.08 $\pm$ 0.44	66.79 $\pm$ 038
		Energy+ReAct	64.51 $\pm$ 2.93	70.51 $\pm$ 0.49	64.65 $\pm$ 2.75	69.97 $\pm$ 0.26	62.86 $\pm$ 0.58	66.78 $\pm$ 039
		NECO	67.61 $\pm$ 1.61	75.89 $\pm$ 0.47	67.47 $\pm$ 1.68	74.29 $\pm$ 0.29	66.28 $\pm$ 0.54	67.40 $\pm$ 0.27
		ViM	63.14 $\pm$ 3.54	72.25 $\pm$ 0.37	63.30 $\pm$ 3.36	71.41 $\pm$ 0.15	64.81 $\pm$ 0.65	66.34 $\pm$ 0.30
		ASH-P	64.79 $\pm$ 2.81	70.27 $\pm$ 0.50	64.86 $\pm$ 2.65	69.79 $\pm$ 0.25	63.08 $\pm$ 0.44	66.61 $\pm$ 0.36