On the Behavior of Convolutional Nets for Feature Extraction

An Out-of-the-box Full-Network Embedding for Convolutional Neural Networks

“…So far, this choice has been supported by performance comparisons of single-layer embeddings, where high-level layer embeddings have been shown to consistently outperform low-level layer embeddings [7,8]. However, it is also known that all layers within a deep network, including low-level ones, can contribute to the characterization of the data in different ways [16]. This implies that the richest and most versatile representation that can be generated by a feature extraction process must include all layers from the network, i.e., it must define a full-network embedding.…”

“…These differences account for large variations in the corresponding feature activations (e.g., distribution, magnitude, etc.). Since our method considers an heterogeneous set of features, a feature standard-Figure 2: For the mit67 dataset, distribution of average standardized feature values for those features belonging to the sets identified in [16]. Vertical dashed lines mark the f t − and f t + thresholds separating the two pairs of distributions as computed by the Kolmogrov-Smirnov statistic.…”

“…To find consistent thresholds, we consider the work of Garcia-Gasulla et al [16], who use a supervised statistical approach to evaluate the importance of CNN features for characterization. Given a feature f and a class c, this work uses an empirical statistic to measure the difference between activation values of f for instances of c and activation values of f for the rest of classes in the dataset.…”

“…In their work, authors further separate these pairs in three different sets: characteristic by abscence, uncharacteristic and characteristic by presence. We use these three sets to find our thresholds f t − and f t + , by mapping the feature/class relevances of Garcia-Gasulla et al [16] to our corresponding feature/image activations. We do so on the datasets explored in [16]: mit67, flowers102 and cub200, by computing the average values of the features belonging to each of the three sets.…”

“…We use these three sets to find our thresholds f t − and f t + , by mapping the feature/class relevances of Garcia-Gasulla et al [16] to our corresponding feature/image activations. We do so on the datasets explored in [16]: mit67, flowers102 and cub200, by computing the average values of the features belonging to each of the three sets. Figure 2 shows the three resulting distributions of values for the mit67 dataset.…”