Authors:

Bharath An

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

The remarkable success of deep neural networks (DNNs) has led to great interest in their adoption in critical applications such as healthcare, autonomous vehicles, and law enforcement. However, DNNs produce significantly more errors under real-world noisy inputs, presenting a major bottleneck to their use in applications where lives, safety, or significant resources are at stake. Prior efforts to address this problem greatly increase training time and produce improvements only on specific types of noise. Our research is motivated by the observation that humans are resilient to a broad range of noisy inputs that challenge DNNs - in fact, humans barely perceive perturbations that cause ANNs to fail spectacularly. We hypothesize that statistically aligning DNNs to human behavior during training can cause them to inherit desirable robustness traits. We propose BrainTrain, a framework to create more robust DNNs through human behavior alignment that consists of (i) a cross-platform mobile application that enables the collection of human behavioral data at scale, (ii) a novel training method that uses a composite loss function to co-optimize accuracy and human behavior alignment during stochastic gradient descent (SGD) based DNN training, and (iii) an evaluation framework to compare BrainTrain-ed DNN models with conventional models. We implemented BrainTrain using open-source software frameworks and applied it to state-of-the-art DNNs. BrainTrain-ed DNNs showed up to 26% higher accuracy under a wide range of noisy inputs and up to 16 times lower calibration error without increasing training time. BrainTrain offers a pathway to enabling the adoption of DNNs in critical applications.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Bharath An