Deep Learning models have achieved remarkable success. Training them is often accelerated by building on top of pre-trained models which poses the risk of perpetuating encoded biases. Here, we investigate biases in the representations of commonly used ImageNet classifiers for facial images while considering intersections of sensitive variables age, race and gender. To assess the biases, we use linear classifier probes and visualize activations as topographic maps. We find that representations in ImageNet classifiers particularly allow differentiation between ages. Less strongly pronounced, the models appear to associate certain ethnicities and distinguish genders in middle-aged groups.

Deep vision models often rely on biases learned from spurious correlations in datasets. To identify these biases, methods that interpret high-level, human-understandable concepts are more effective than those relying primarily on low-level features like heatmaps. A major challenge for these concept-based methods is the lack of image annotations indicating potentially bias-inducing concepts, since creating such annotations requires detailed labeling for each dataset and concept, which is highly labor-intensive. We present CUBIC (Concept embeddings for Unsupervised Bias IdentifiCation), a novel method that automatically discovers interpretable concepts that may bias classifier behavior. Unlike existing approaches, CUBIC does not rely on predefined bias candidates or examples of model failures tied to specific biases, as such information is not always available. Instead, it leverages image-text latent space and linear classifier probes to examine how the latent representation of a superclass label$\unicode{x2014}$shared by all instances in the dataset$\unicode{x2014}$is influenced by the presence of a given concept. By measuring these shifts against the normal vector to the classifier's decision boundary, CUBIC identifies concepts that significantly influence model predictions. Our experiments demonstrate that CUBIC effectively uncovers previously unknown biases using Vision-Language Models (VLMs) without requiring the samples in the dataset where the classifier underperforms or prior knowledge of potential biases.

In this paper, we elucidate how representations in deep neural networks (DNNs) evolve during training. We focus on overparameterized learning settings where the training continues much after the trained DNN starts to perfectly fit its training data. We examine the evolution of learned representations along the entire training process, including its perfect fitting regime, and with respect to the epoch-wise double descent phenomenon. We explore the representational similarity of DNN layers, each layer with respect to its own representations throughout the training process. For this, we use two similarity metrics: (1) The centered kernel alignment (CKA) similarity; (2) Similarity of decision regions of linear classifier probes that we train for the DNN layers. Our extensive experiments discover training dynamics patterns that can emerge in layers depending on the relative layer-depth, DNN width, and architecture. We show that representations at the deeper layers evolve much more in the training when an epoch-wise double descent occurs. For Vision Transformer, we show that the perfect fitting threshold creates a transition in the evolution of representations across all the encoder blocks.

Neural network models have a reputation for being black boxes. We propose to monitor the features at every layer of a model and measure how suitable they are for classification. We use linear classifiers, which we refer to as "probes", trained entirely independently of the model itself. This helps us better understand the roles and dynamics of the intermediate layers. We demonstrate how this can be used to develop a better intuition about models and to diagnose potential problems. We apply this technique to the popular models Inception v3 and Resnet-50. Among other things, we observe experimentally that the linear separability of features increase monotonically along the depth of the model.

We propose a confidence scoring mechanism for multi-layer neural networks based on a paradigm of a base model and a meta-model. The confidence score is learned by the meta-model using features derived from the base model -- a deep multi-layer neural network -- considered a whitebox. As features, we investigate linear classifier probes inserted between the various layers of the base model and trained using each layer's intermediate activations. Experiments show that this approach outperforms various baselines in a filtering task, i.e., task of rejecting samples with low confidence. Experimental results are presented using CIFAR-10 and CIFAR-100 dataset with and without added noise exploring various aspects of the method.