Moe Matsuki

Neural networks have been shown vulnerable to a variety of adversarial algorithms. A crucial step to understanding the rationale for this lack of robustness is to assess the potential of the neural networks' representation to encode the existing features. Here, we propose a method to understand the representation quality of the neural networks using a novel test based on Zero-Shot Learning, entitled Raw Zero-Shot. The principal idea is that, if an algorithm learns rich features, such features should be able to interpret "unknown" classes as an aggregate of previously learned features. This is because unknown classes usually share several regular features with recognised classes, given the features learned are general enough. We further introduce two metrics to assess these learned features to interpret unknown classes. One is based on inter-cluster validation technique (Davies-Bouldin Index), and the other is based on the distance to an approximated ground-truth. Experiments suggest that adversarial defences improve the representation of the classifiers, further suggesting that to improve the robustness of the classifiers, one has to improve the representation quality also. Experiments also reveal a strong association (a high Pearson Correlation and low p-value) between the metrics and adversarial attacks. Interestingly, the results indicate that dynamic routing networks such as CapsNet have better representation while current deeper neural networks are trading off representation quality for accuracy. Code available at http://bit.ly/RepresentationMetrics.

In this paper, we address Zero-shot learning for sensor activity recognition using word embeddings. The goal of Zero-shot learning is to estimate an unknown activity class (i.e., an activity that does not exist in a given training dataset) by learning to recognize components of activities expressed in semantic vectors. The existing zero-shot methods use mainly 2 kinds of representation as semantic vectors, attribute vector and embedding word vector. However, few zero-shot activity recognition methods based on embedding vector have been studied; especially for sensor-based activity recognition, no such studies exist, to the best of our knowledge. In this paper, we compare and thoroughly evaluate the Zero-shot method with different semantic vectors: (1) attribute vector, (2) embedding vector, and (3) expanded embedding vector and analyze their correlation to performance. Our results indicate that the performance of the three spaces is similar but the use of word embedding leads to a more efficient method, since this type of semantic vector can be generated automatically. Moreover, our suggested method achieved higher accuracy than attribute-vector methods, in cases when there exist similar information in both the given sensor data and in the semantic vector; the results of this study help select suitable classes and sensor data to build a training dataset.