Deep neural networks (DNNs) have demonstrated state-of-the-art results on many pattern recognition tasks, especially vision classification problems. Understanding the inner workings of such computational brains is both fascinating basic science that is interesting in its own right---similar to why we study the human brain---and will enable researchers to further improve DNNs. One path to understanding how a neural network functions internally is to study what each of its neurons has learned to detect. One such method is called activation maximization, which synthesizes an input (e.g. an image) that highly activates a neuron. Here we dramatically improve the qualitative state of the art of activation maximization by harnessing a powerful, learned prior: a deep generator network. The algorithm (1) generates qualitatively state-of-the-art synthetic images that look almost real, (2) reveals the features learned by each neuron in an interpretable way, (3) generalizes well to new datasets and somewhat well to different network architectures without requiring the prior to be relearned, and (4) can be considered as a high-quality generative method (in this case, by generating novel, creative, interesting, recognizable images).

Deep neural networks (DNNs) have demonstrated state-of-the-art results on many pattern recognition tasks, especially vision classification problems. Understanding the inner workings of such computational brains is both fascinating basic science that is interesting in its own right---similar to why we study the human brain---and will enable researchers to further improve DNNs. One path to understanding how a neural network functions internally is to study what each of its neurons has learned to detect. One such method is called activation maximization, which synthesizes an input (e.g. an image) that highly activates a neuron. Here we dramatically improve the qualitative state of the art of activation maximization by harnessing a powerful, learned prior: a deep generator network. The algorithm (1) generates qualitatively state-of-the-art synthetic images that look almost real, (2) reveals the features learned by each neuron in an interpretable way, (3) generalizes well to new datasets and somewhat well to different network architectures without requiring the prior to be relearned, and (4) can be considered as a high-quality generative method (in this case, by generating novel, creative, interesting, recognizable images).

Neural Information Processing Systems http://nips.cc/

In explainable AI, Concept Activation Vectors (CAVs) are typically obtained by training linear classifier probes to detect human-understandable concepts as directions in the activation space of deep neural networks. It is widely assumed that a high probe accuracy indicates a CAV faithfully representing its target concept. However, we show that the probe's classification accuracy alone is an unreliable measure of concept alignment, i.e., the degree to which a CAV captures the intended concept. In fact, we argue that probes are more likely to capture spurious correlations than they are to represent only the intended concept. As part of our analysis, we demonstrate that deliberately misaligned probes constructed to exploit spurious correlations, achieve an accuracy close to that of standard probes. To address this severe problem, we introduce a novel concept localization method based on spatial linear attribution, and provide a comprehensive comparison of it to existing feature visualization techniques for detecting and mitigating concept misalignment. We further propose three classes of metrics for quantitatively assessing concept alignment: hard accuracy, segmentation scores, and augmentation robustness. Our analysis shows that probes with translation invariance and spatial alignment consistently increase concept alignment. These findings highlight the need for alignment-based evaluation metrics rather than probe accuracy, and the importance of tailoring probes to both the model architecture and the nature of the target concept.

Understanding internal feature representations of deep neural networks (DNNs) is a fundamental step toward model interpretability. Inspired by neuroscience methods that probe biological neurons using visual stimuli, recent deep learning studies have employed Activation Maximization (AM) to synthesize inputs that elicit strong responses from artificial neurons. In this work, we propose a unified feature visualization framework applicable to both Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). Unlike prior efforts that predominantly focus on the last output-layer neurons in CNNs, we extend feature visualization to intermediate layers as well, offering deeper insights into the hierarchical structure of learned feature representations. Furthermore, we investigate how activation maximization can be leveraged to generate adversarial examples, revealing potential vulnerabilities and decision boundaries of DNNs. Our experiments demonstrate the effectiveness of our approach in both traditional CNNs and modern ViT, highlighting its generalizability and interpretive value.

This study demonstrates the capabilities of several methods for analyzing the sensitivity of neural networks to perturbations of the input data and interpreting their underlying mechanisms. The investigated approaches include the Sobol global sensitivity analysis, the local sensitivity method for input pixel perturbations and the activation maximization technique. As examples, in this study we consider a small feedforward neural network for analyzing an open tabular dataset of clinical diabetes data, as well as two classical convolutional architectures, VGG-16 and ResNet-18, which are widely used in image processing and classification. Utilization of the global sensitivity analysis allows us to identify the leading input parameters of the chosen tiny neural network and reduce their number without significant loss of the accuracy. As far as global sensitivity analysis is not applicable to larger models we try the local sensitivity analysis and activation maximization method in application to the convolutional neural networks. These methods show interesting patterns for the convolutional models solving the image classification problem. All in all, we compare the results of the activation maximization method with popular Grad-CAM technique in the context of ultrasound data analysis.

Superposition or Neuron Polysemanticity are important concepts in the field of interpretability and one might say they are these most intricately beautiful blockers in our path of decoding the Machine Learning black-box. The idea behind this paper is to examine whether it is possible to decode Superposition using Active Learning methods. While it seems that Superposition is an attempt to arrange more features in smaller space to better utilize the limited resources, it might be worth inspecting if Superposition is dependent on any other factors. This paper uses CIFAR-10 and Tiny ImageNet image datasets and the ResNet18 model and compares Baseline and Active Learning models and the presence of Superposition in them is inspected across multiple criteria, including t-SNE visualizations, cosine similarity histograms, Silhouette Scores, and Davies-Bouldin Indexes. Contrary to our expectations, the active learning model did not significantly outperform the baseline in terms of feature separation and overall accuracy. This suggests that non-informative sample selection and potential overfitting to uncertain samples may have hindered the active learning model's ability to generalize better suggesting more sophisticated approaches might be needed to decode superposition and potentially reduce it.

Understanding how models process and interpret time series data remains a significant challenge in deep learning to enable applicability in safety-critical areas such as healthcare. In this paper, we introduce Sequence Dreaming, a technique that adapts Activation Maximization to analyze sequential information, aiming to enhance the interpretability of neural networks operating on univariate time series. By leveraging this method, we visualize the temporal dynamics and patterns most influential in model decision-making processes. To counteract the generation of unrealistic or excessively noisy sequences, we enhance Sequence Dreaming with a range of regularization techniques, including exponential smoothing. This approach ensures the production of sequences that more accurately reflect the critical features identified by the neural network. Our approach is tested on a time series classification dataset encompassing applications in predictive maintenance. The results show that our proposed Sequence Dreaming approach demonstrates targeted activation maximization for different use cases so that either centered class or border activation maximization can be generated. The results underscore the versatility of Sequence Dreaming in uncovering salient temporal features learned by neural networks, thereby advancing model transparency and trustworthiness in decision-critical domains.

Deep neural networks (DNNs) have demonstrated state-of-the-art results on many pattern recognition tasks, especially vision classification problems. Understanding the inner workings of such computational brains is both fascinating basic science that is interesting in its own right--similar to why we study the human brain--and will enable researchers to further improve DNNs. One path to understanding how a neural network functions internally is to study what each of its neurons has learned to detect. One such method is called activation maximization (AM), which synthesizes an input (e.g. an image) that highly activates a neuron. Here we dramatically improve the qualitative state of the art of activation maximization by harnessing a powerful, learned prior: a deep generator network (DGN). The algorithm (1) generates qualitatively state-of-the-art synthetic images that look almost real, (2) reveals the features learned by each neuron in an interpretable way, (3) generalizes well to new datasets and somewhat well to different network architectures without requiring the prior to be relearned, and (4) can be considered as a high-quality generative method (in this case, by generating novel, creative, interesting, recognizable images).

Neural networks (NNs), both living and artificial, work due to being complex systems of neurons, each having its own specialization. Revealing these specializations is important for understanding NNs inner working mechanisms. The only way to do this for a living system, the neural response of which to a stimulus is not a known (let alone differentiable) function is to build a feedback loop of exposing it to stimuli, the properties of which can be iteratively varied aiming in the direction of maximal response. To test such a loop on a living network, one should first learn how to run it quickly and efficiently, reaching most effective stimuli (ones that maximize certain neurons activation) in least possible number of iterations. We present a framework with an effective design of such a loop, successfully testing it on an artificial spiking neural network (SNN, a model that mimics the behaviour of NNs in living brains). Our optimization method used for activation maximization (AM) was based on low-rank tensor decomposition (Tensor Train, TT) of the activation function's discretization over its domain the latent parameter space of stimuli (CIFAR10-size color images, generated by either VQ-VAE or SN-GAN from their latent description vectors, fed to the SNN). To our knowledge, the present work is the first attempt to perform effective AM for SNNs. The source code of our framework, MANGO (for Maximization of neural Activation via Non-Gradient Optimization) is available on GitHub.