Exploring the Effect of Baselines Generated from Latent Space in Integrated Gradient

An explanation of a decision is important to believe the decision of a deep neural network. A widely used model-agnostic explanation method is integrated gradient (IG) which interpolates gradient signals from a manually chosen baseline to the input. One of the fundamental problems of IG is the choice of a baseline as the baseline only requires zero prediction signal for the explantion target. There are lots of possible baselines in an input space and they produce very different explanation results. Therefore, it is necessary to explore the image space to understand the role of baselines in IG. However, the input space is broad that evaluation of several baselines is intractable. To mitigate the burden of baseline analysis, we propose baselines generated from 2-dimensional bottleneck which represent the input image. With an extensive evaluation of several baselines in CIFAR10 and MNIST, we experimentally show that baselines could be class-specific and the they should not have the same value with the important regions of an input.

EarChew:  Towards Identifying Chewing Side Preference using Earables

We present EarChew, which utilizes an earable worn on the patient’s ear to be able to identify the chewing side – i.e., left or right. EarChew utilizes the microphone embedded in the earable to capture the minute but inherent vibrations caused by each chewing action to be able to identify the chewing side. We present a preliminary evaluation with a participant chewing on almonds and demonstrate a promising preliminary result of 96% average accuracy.

Towards identifying Robustness of Out-of-Distribution Score on Variational Auto-encoder

This study focuses on the challenge of detecting out-of-distribution (OOD) samples in Variational Auto-encoders (VAEs), extending the Likelihood Regret score. While deep learning models have excelled in various domains, their vulnerability to OOD data remains a significant obstacle in practical applications. Existing OOD detection methods, designed primarily for classifiers, often rely on labeled data, restricting their use in semi- or unsupervised scenarios. Deep probabilistic generative models, such as VAEs, present a promising alternative by utilizing likelihood estimates to distinguish between in-distribution (ID) and OOD samples. The Likelihood Regret score is introduced as a novel OOD detection metric for VAEs. The experiments cover various datasets, including transformed, corrupted, and adversarial examples. The results indicate that the VAE trained on MNIST excels in detecting transformed samples but struggles with differentiating corrupted versions. Both Likelihood Regret and Likelihood scores exhibit proficiency in detecting adversarial samples. Overall, this work contributes to advancing OOD detection methodologies tailored for the unique characteristics of VAEs, enhancing their robustness in real-world applications.

Application of Energy-Based Model to Image Classification

    Langevin dynamics algorithm with energy-based model (EBM) can generate a sample close to the data manifold. We defined the real samples as positive samples and the generated samples as negative samples and utilized them in the classifier training with new loss functions. The objective is to obtain the classifier with high validation accuracy and low decision-making for negative samples, which is an adversarial objective contrastive to the EBM model. 

    It was noticed that our trained EBM model could make negative samples so close to positive samples in terms of energy and visualization. In contrast, our classifiers didn't take advantage of the negative samples and even the validation accuracy decreased. In the test phase, KL divergence between the discrete uniform distribution and the logit distribution, and the entropy distribution are compared to measure the model performance for the negative samples.