Deep neural networks achieve outstanding results in a large variety of tasks, often outperforming human experts. However, a known limitation of current neural architectures is the poor accessibility in understanding and interpreting the network’s response to a given input. This is directly related to the huge number of variables and the associated non-linearities of neural models, which are often used as black boxes. This lack of transparency, particularly in crucial areas like autonomous driving, security, and healthcare, can trigger skepticism and limit trust, despite the networks’ high performance. In this work, we want to advance the interpretability in neural networks. We present Agglomerator++, a framework capable of providing a representation of part-whole hierarchies from visual cues and organizing the input distribution to match the conceptual-semantic hierarchical structure between classes. We evaluate our method on common datasets, such as SmallNORB, MNIST, FashionMNIST, CIFAR-10, and CIFAR-100, showing that our solution delivers a more interpretable model compared to other state-of-the-art approaches. Our code is available at https://mmlab-cv.github.io/Agglomeratorplusplus/ . • We introduce a novel model, called Agglomerator++, mimicking the functioning of the cortical columns in the human brain. • Our solution provides interpretability of relationships in data, namely the hierarchical organization of the feature space. • We introduce positional encoding and input masking during pre-training for self- supervised reconstruction. • This neural representation is more efficient and closely resembles human lexical similarities.

Deep neural networks achieve outstanding results in a large variety of tasks, often outperforming human experts. However, a known limitation of current neural architectures is the poor accessibility to understand and interpret the network response to a given input. This is directly related to the huge number of variables and the associated non-linearities of neural models, which are often used as black boxes. When it comes to critical applications as autonomous driving, security and safety, medicine and health, the lack of interpretability of the network behavior tends to induce skepticism and limited trustworthiness, despite the accurate performance of such systems in the given task. Furthermore, a single metric, such as the classification accuracy, provides a non-exhaustive evaluation of most realworld scenarios. In this paper, we want to make a step forward towards interpretability in neural networks, providing new tools to interpret their behavior. We present Agglomerator, a framework capable of providing a representation of part-whole hierarchies from visual cues and organizing the input distribution matching the conceptual-semantic hierarchical structure between classes. We evaluate our method on common datasets, such as SmallNORB, MNIST, FashionMNIST, CIFAR-10, and CIFAR-100, providing a more interpretable model than other state-of-the-art approaches.

The goal of out-of-distribution (OOD) detection is to handle the situations where the test samples are drawn from a different distribution than the training data. In this paper, we argue that OOD samples can be detected more easily if the training data is embedded into a low-dimensional space, such that the embedded training samples lie on a union of 1-dimensional subspaces. We show that such embedding of the in-distribution (ID) samples provides us with two main advantages. First, due to compact representation in the feature space, OOD samples are less likely to occupy the same region as the known classes. Second, the first singular vector of ID samples belonging to a 1-dimensional subspace can be used as their robust representative. Motivated by these observations, we train a deep neural network such that the ID samples are embedded onto a union of 1-dimensional subspaces. At the test time, employing sampling techniques used for approximate Bayesian inference in deep learning, input samples are detected as OOD if they occupy the region corresponding to the ID samples with probability 0. Spectral components of the ID samples are used as robust representative of this region. Our method does not have any hyperparameter to be tuned using extra information and it can be applied on different modalities with minimal change. The effectiveness of the proposed method is demonstrated on different benchmark datasets, both in the image and video classification domains.