Shaping Noise for Robust Attributions in Neural Stochastic Differential Equations

Abstract

Neural SDEs with Brownian motion as noise lead to smoother attributions than traditional ResNets. Various attribution methods such as saliency maps, integrated gradients, DeepSHAP and DeepLIFT have been shown to be more robust for neural SDEs than for ResNets using the recently proposed sensitivity metric. In this paper, we show that neural SDEs with adaptive attribution-driven noise lead to even more robust attributions and smaller sensitivity metrics than traditional neural SDEs with Brownian motion as noise. In particular, attribution-driven shaping of noise leads to 6.7%, 6.9% and 19.4% smaller sensitivity metric for integrated gradients computed on three discrete approximations of neural SDEs with standard Brownian motion noise: stochastic ResNet-50, WideResNet-101 and ResNeXt-101 models respectively. The neural SDE model with adaptive attribution-driven noise leads to 25.7% and 4.8% improvement in the SIC metric over traditional ResNets and Neural SDEs with Brownian motion as noise. To the best of our knowledge, we are the first to propose the use of attributions for shaping the noise injected in neural SDEs, and demonstrate that this process leads to more robust attributions than traditional neural SDEs with standard Brownian motion as noise.

Linear-Scaling Kernels for Protein Sequences and Small Molecules Outperform Deep Learning While Providing Uncertainty Quantitation and Improved Interpretability

“…Adding subtle noise not perceptible to the human eye, for example, could reliably cause deep learning models to misclassify a photo of a cat as guacamole, or a lionfish as eggnog . In biology, protein structure prediction models have been shown to exhibit the same vulnerability to adversarial attacks . Finally, deep learning often entails enormous computational cost.…”

Adversarial Attacks on Protein Language Models

“…Adversarial embeddings are at least as far from the reference as the most stabilizing and destabilizing embeddings in ProTherm (Figure 2b), while BLOSUM distances are comparable (Figure 3b). Moreover, Jha et al [7] observed that larger BLOSUM distances between original and perturbed sequences lead to higher RMSD in predicted structures, therefore we expect adversarial perturbations to produce a significant structural change. We stress that adversarial mutant positions for ProTherm are those that maximize the attention scores and, in most cases, differ from ProTherm mutation sites.…”

“…It is natural to interpret this reasoning through the lens of adversarial attacks [6], where the goal is to craft small perturbations of the inputs while inducing the most significant change in predictions. A similar approach is presented in [7], where the authors already propose a notion of “adversarial mutation” on amino acid sequences and use the perturbations to assess the reliability of structure prediction on the RoseTTAFold [1] model. Specifically, they define a robustness measure for structure prediction based on the computation of the inverse RMSD between original and perturbed structures and use it to produce adversarial sequences.…”

“…BLOSUM matrix contains integer similarity scores between all couples of residues in a sequence. In particular, scores in the popular BLOSUM62 matrix are based on replacement frequencies observed in known alignments with less than 62% sequence similarity [7]. The selection of amino acids at a given position is restricted to residues with non-null BLOSUM scores, i.e.…”

“…Despite the remarkable capabilities of protein language models, they are often unable to predict misfolding caused by single-point mutations [2]. Moreover, it has been observed that biologically small perturbations in amino acid sequences manage to induce radical changes in the 3d structures [7]. This limitation is presumably due to the scarcity of structure disruptive mutations among the available PDBs [10], which makes protein language models unable to catch all the biological features causing structural variations.…”