【摘 要】
:
Artificial neural networks can achieve impressive performances,and even outperform humans in some specific tasks.Nevertheless,unlike biological brains,the artif
【机 构】
:
PMI Laboratory,School of Physics,Sun Yat-sen University,Guangzhou 510275,China
论文部分内容阅读
Artificial neural networks can achieve impressive performances,and even outperform humans in some specific tasks.Nevertheless,unlike biological brains,the artificial neural networks suffer from tiny perturbations in sensory input,under various kinds of adversarial attacks.It is therefore necessary to study the origin of the adversarial vulnerability.Here,we establish a fundamental relationship between geometry of hidden representations(manifold perspective)and the general-ization capability of the deep networks.For this purpose,we choose a deep neural network trained by local errors,and then analyze emergent properties of the trained networks through the manifold dimensionality,manifold smoothness,and the generalization capability.To explore effects of adversarial examples,we consider independent Gaussian noise attacks and fast-gradient-sign-method(FGSM)attacks.Our study reveals that a high generalization accuracy requires a relatively fast power-law decay of the eigen-spectrum of hidden representations.Under Gaussian attacks,the relationship between generalization accuracy and power-law exponent is monotonic,while a non-monotonic behavior is observed for FGSM attacks.Our empirical study provides a route towards a final mechanistic interpretation of adversarial vulnerability under adversarial attacks.
其他文献
The unprecedented wheat yield increases during the Green Revolution were achieved through the introduction of the Reduced height (Rht)-B1b and Rht-D1b semi-dwar
Using the innovative method of the additional Bloch vector,the electron transfer properties of a double quantum dot(DQD)system measured by a quantum point conta
Thanks to the quantum simulation,more and more problems in quantum mechanics which were previously inaccessi-ble are now open to us.Capitalizing on the state-of
Verification in quantum computations is crucial since quantum systems are extremely vulnerable to the environment.However,verifying directly the output of a qua
Optimal creation of photon Fock states is of importance for quantum information processing and state engineering.Here an efficient strategy is presented for spe
Recent advances in quantum technology have led to the development and the manufacturing of programmable quan-tum annealers that promise to solve certain combina
The biological functions of the epitranscriptomic modification N6-methyladenosine (m6A) in plants are not fully understood.CPSF30-L is a predominant isoform of
Nuclear accidents have led to the release to the environment of radioactive isotopes, mainly radioactive cesium (Cs+), contaminating large areas of soil for ver
Identifying the uniqueness of FeP-based superconductors may shed new lights on the mechanism of superconductivity in iron-pnictides.Here,we report nuclear magne
This review deals with restricted Boltzmann machine(RBM)under the light of statistical physics.The RBM is a classical family of machine learning(ML)models which