Despite excellent performance on stationary test sets, deep neural networks (DNNs) can fail to generalize to out-of-distribution (OoD) inputs, including natural, nonadversarial ones, which are common in real-world settings. In this paper, we present a framework for discovering DNN failures that harnesses 3D renderers and 3D models. That is, we estimate the parameters of a 3D renderer that cause a target DNN to misbehave in response to the rendered image. Using our framework and a self-assembled dataset of 3D objects, we investigate the vulnerability of DNNs to OoD poses of well-known objects in ImageNet. For objects that are readily recognized by DNNs in their canonical poses, DNNs incorrectly classify 97% of their pose space. In addition, DNNs are highly sensitive to slight pose perturbations. Importantly, adversarial poses transfer across models and datasets. We find that 99.9% and 99.4% of the poses misclassified by Inception-v3 also transfer to the AlexNet and ResNet-50 image classifiers trained on the same ImageNet dataset, respectively, and 75.5% transfer to the YOLOv3 object detector trained on MS COCO.
The picosecond time-resolved infrared spectra of (CN)5FeCNRu(NH3)5 - were collected following optical excitation and reverse electron transfer. The measured reverse electron transfer rates are greater than 3 × 1012 s-1. In both formamide and deuterated water solutions, vibrational excitation in CN stretch modes is found after reverse electron transfer. The transient spectra at both earlier (1−35 ps) and later (10 ns) times give evidence of environment−solute coupling that can be accounted for by solvent heating and ion pair dynamics. A simulation of the spectral dynamics in formamide solution is presented using a kinetic model for vibrational excitation and relaxation. The simulation includes minor excitation in vibrational modes consistent with resonance Raman derived Franck−Condon factors, but it is also found that a nontotally symmetric mode is equally important as an acceptor.
We present time-resolved infrared and visible spectroscopic data that reveal the relative orientations of donor and acceptor molecules in the 〈τET〉 = 4 ps intermolecular electron transfer between dimethylaniline solvent and coumarin 337 solute. Although consideration of the dipolar attraction suggests the two species are most likely to be aligned with their molecular long axes parallel, the electron transfer occurs between species with their long axes perpendicular.
The charge-transfer resonance Raman spectra of (CN) 5 FeCNRu(NH 3 ) 5taken in several solvents show that the Franck-Condon activity of the different CN stretch modes is strongly solvent dependent. These results imply that the choice of solvent can control high-frequency vibrational mode coupling to electron transfer (ET). Specific couplings between solute (donor/acceptor, DA) vibrations and solvent have previously been thought to be important only for the low-frequency modes of the system.A polar solvent often largely controls the relative free energies of DA and D + Ain solution. [1][2][3] Similarly, intramolecular vibrations of DA and D + Aexhibit different equilibrium displacements in DA and D + Aand, thus, also contribute to the relative free energies, which is crucial in the Marcus inverted regime. 4-6 A model of the ET mechanism must account for the reorganization energies and dynamics among these modes, and resonance Raman spectra are particularly useful for extracting specific mode coupling information. [7][8][9][10][11][12][13][14] As shown in Figure 1, we have collected the resonance (1064 nm) Raman spectra of (CN) 5 FeCNRu(NH 3 ) 5in three solvents: water (H 2 O and D 2 O), formamide (FA), and N-methylformamide (NMF). 15 These data show the CN vibrational stretch region. The mode at ca. 2100 cm -1 has been previously assigned to the bridging CN ligand. [7][8][9]14 The mode at ca. 2000 cm -1 in NMF and FA is assigned to the stretch of the CN ligand trans to the bridging ligand. 16 A much weaker band, also assigned to the trans CN ligand, is seen at 2014 cm -1 in D 2 O. The cis CN modes are not Raman observed to any significant extent in any of these three solvents. On the other hand, the IR spectra in this frequency range (data not shown) indicate a strong band at 2050 cm -1 , which may be assigned to the cis-CN stretch. The intensity of the trans-CN stretch band is strongly solvent dependent, as is its frequency shift.We have also collected the Raman excitation profile for (CN) 5 FeCNRu(NH 3 ) 5in D 2 O. The relative Raman intensities of the CN modes at ca. 2000 and 2100 cm -1 do not depend significantly on the excitation frequency (data not shown, see the Supporting Information). These results indicate that the variation of Raman intensities among the three solvents is not a consequence of the wavelength of the exciting radiation relative to the origin of the charge-transfer band.We have simultaneously fit the resonance Raman and CT absorption spectra. 17,19 A variety of approaches 7,[11][12][13][14][20][21][22][23][25][26][27][28][29][30][31][32] have been used by others to account for solvent broadening of the resonance Raman and charge-transfer absorption bands, to † University of Pittsburgh. ‡ Northwestern University. (1) Siders, P.; Marcus, R. A. Britt, B. M.; Lueck, H. B.; McHale, J. L. Chem. Phys. Lett. 1992, 190, 528-532. (15) Solvents were purchased from Aldrich and used without further purification. The concentration of (CN)5FeCNRu(NH3)5is 0.001 M in all three solvents. The measurements ...
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.