We observe that existing model interpretation methods generally ignore the baseline, and such neglect often results in imprecise or even incorrect interpretation. In this paper, we reformulate the task of model interpretation and the interpretation principles for model interpretation results to demonstrate the importance of the baseline. For the first time, we unify gradient-based methods, Integrated Gradients (IG), and Taylor expansion, clarify the relationships among the three, and explicitly identify the corresponding baseline for each method. This may have a significant impact on the further performance improvement of some gradient-based schemes. On this basis, we analyze the flaws and errors in related model interpretation methods (IG, LayerCAM, ODAM, Difference Map). We advocate evaluating the quality of model interpretation results precisely through the attribution error between the attribution result and the attribution target, rather than adopting flawed evaluation methods, such as those based on marginal-effect or the assumption of perfect model performance. We revise IG and develope a model interpretation method with a clear and reasonable baseline, achieving better resu
The non-relativistic quantum theory has been interpreted causally by de Broglie, David Bohm, and others, where a quantum entity is viewed as a particle with a definite position and momentum. This interpretation opposes the Copenhagen orthodox interpretation, which expresses uncertainty through the commutator relationship between position and momentum operators. This is further exacerbated by the de Broglie-Bohm interpretation, which uses no operators corresponding to these observables. We reconcile these opposing viewpoints by introducing mathematical nonlinear operators for the momentum observable. While nonlinear non-Hermitian operators cannot be easily used as matrices in the Hilbert space, we show that they are implicitly embedded in de Broglie-Bohm's interpretation of quantum theory.
Network traffic, as a key media format, is crucial for ensuring security and communications in modern internet infrastructure. While existing methods offer excellent performance, they face two key bottlenecks: (1) They fail to capture multidimensional semantics beyond unimodal sequence patterns. (2) Their black box property, i.e., providing only category labels, lacks an auditable reasoning process. We identify a key factor that existing network traffic datasets are primarily designed for classification and inherently lack rich semantic annotations, failing to generate human-readable evidence report. To address data scarcity, this paper proposes a Byte-Grounded Traffic Description (BGTD) benchmark for the first time, combining raw bytes with structured expert annotations. BGTD provides necessary behavioral features and verifiable chains of evidence for multimodal reasoning towards explainable encrypted traffic interpretation. Built upon BGTD, this paper proposes an end-to-end traffic-language representation framework (mmTraffic), a multimodal reasoning architecture bridging physical traffic encoding and semantic interpretation. In order to alleviate modality interference and genera
Interpretation is not the only way to explain a theory's success, form and features, and nor is it the only way to solve problems we see with a theory. This can also be done by giving a reductive explanation of the theory, by reference to a newer, more accurate, and/or more fundamental theory. We are seeking a theory of quantum gravity, a more fundamental theory than both quantum mechanics and general relativity, yet, while this theory is supposed to explain general relativity, it's not typically been thought to be necessary, or able, to explain quantum mechanics -- a task instead assigned to interpretation. Here, I question why this is. I also present a new way of assessing the various interpretations of quantum mechanics, in terms of their heuristic and unificatory potential in helping us find a more fundamental theory.
To address the issues of stability and fidelity in interpretable learning, a novel interpretable methodology, ensemble interpretation, is presented in this paper which integrates multi-perspective explanation of various interpretation methods. On one hand, we define a unified paradigm to describe the common mechanism of different interpretation methods, and then integrate the multiple interpretation results to achieve more stable explanation. On the other hand, a supervised evaluation method based on prior knowledge is proposed to evaluate the explaining performance of an interpretation method. The experiment results show that the ensemble interpretation is more stable and more consistent with human experience and cognition. As an application, we use the ensemble interpretation for feature selection, and then the generalization performance of the corresponding learning model is significantly improved.
Recent years have witnessed an increasing number of interpretation methods being developed for improving transparency of NLP models. Meanwhile, researchers also try to answer the question that whether the obtained interpretation is faithful in explaining mechanisms behind model prediction? Specifically, (Jain and Wallace, 2019) proposes that "attention is not explanation" by comparing attention interpretation with gradient alternatives. However, it raises a new question that can we safely pick one interpretation method as the ground-truth? If not, on what basis can we compare different interpretation methods? In this work, we propose that it is crucial to have a concrete definition of interpretation before we could evaluate faithfulness of an interpretation. The definition will affect both the algorithm to obtain interpretation and, more importantly, the metric used in evaluation. Through both theoretical and experimental analysis, we find that although interpretation methods perform differently under a certain evaluation metric, such a difference may not result from interpretation quality or faithfulness, but rather the inherent bias of the evaluation metric.
The integration of AI-assisted biomedical image analysis into clinical practice demands AI-generated findings that are not only accurate but also interpretable to clinicians. However, existing biomedical AI models generally lack the ability to simultaneously generate diagnostic findings and localize corresponding biomedical objects. This limitation makes it challenging for clinicians to correlate AI-generated findings with visual evidence (e.g., tiny lesions) in images and interpret the results of AI models. To address this challenge, we introduce UniBiomed, the first universal foundation model for grounded biomedical image interpretation, which is capable of generating accurate diagnostic findings and simultaneously segmenting the corresponding biomedical targets. UniBiomed is based on a novel integration of Multi-modal Large Language Model and Segment Anything Model, which can effectively unify diverse biomedical tasks in universal training for advancing grounded interpretation. To develop UniBiomed, we curate a large-scale dataset comprising over 27 million triplets of images, region annotations, and text descriptions across ten biomedical imaging modalities. Extensive validatio
Material heritage typically has a whole set of associated immaterial heritage, which is essential to pass on to the visitor as a cultural mission of the destinations and those who manage them. In this sense, the interpretation of material heritage is a complex process that is not a fully efficient process with the mere observation of physical artifacts. In this context, it emerges as fundamental to provide visitors with a set of tools that allow them to correctly interpret the artifacts that come to fully understand the cultural dimension of the destinations and their heritage. Accordingly, the role of virtual reality can leverage the creation of innovative and immersive solutions that allow the visitor to understand and feel part of their own heritage and its ancestral component that defines the sociocultural roots of destinations and their civilizational traditions. This article, after dissecting and substantiating the role of virtual reality in the interpretation of heritage, presents a conceptual model, based on the use of virtual reality, which was, in part, prototyped in the scenario of the Portuguese Museum in the city of Miranda do Douro. This proposal is an ongoing contrib
We investigate proper infall times in the Schwarzschild and Kerr spacetimes from a covariant perspective, focusing on the role of black--hole rotation in the focusing properties of timelike geodesic congruences.To perform a geometrically consistent comparison between rotating and non--rotating black holes, we analyse infall trajectories between surfaces of equal circumferential radius in the equatorial plane. Using equatorial timelike geodesics in the test--particle limit, we compute and compare the corresponding proper infall times for different values of the specific energy $E$, specific angular momentum $L$, and black--hole spin parameter $a$. Within the equal circumferential-radius prescription adopted here, we show that Kerr angular momentum $a$ can produce longer or shorter integrated proper infall times relative to the Schwarzschild case, depending on the orbital configuration and energy regime considered. We then interpret these results within the covariant $1+3$ formalism of general relativity, in terms of the expansion, shear, and Raychaudhuri evolution of timelike congruences. Our analysis shows that the Kerr--Schwarzschild differences in proper infall times are encoded
We present a geometric model of the differential sensitivity of the fidelity error for state transfer in a spintronic network based on the relationship between a set of matrix operators. We show an explicit dependence of the differential sensitivity on the fidelity (error), and we further demonstrate that this dependence does not require a trade-off between the fidelity and sensitivity. Rather, we prove that for closed systems, ideal performance in the sense of perfect state transfer is both necessary and sufficient for optimal robustness in terms of vanishing sensitivity. We demonstrate the utility of this geometric interpretation of the sensitivity by applying the model to explain the sensitivity versus fidelity error data in two examples.
This paper presents a mathematical interpretation of self-attention by connecting it to distributional semantics principles. We show that self-attention emerges from projecting corpus-level co-occurrence statistics into sequence context. Starting from the co-occurrence matrix underlying GloVe embeddings, we demonstrate how the projection naturally captures contextual influence, with the query-key-value mechanism arising as the natural asymmetric extension for modeling directional relationships. Positional encodings and multi-head attention then follow as structured refinements of this same projection principle. Our analysis demonstrates that the Transformer architecture's particular algebraic form follows from these projection principles rather than being an arbitrary design choice.
The Minkowski vacuum is often presented in textbooks and reviews as a thermofield double (TFD) state, an entangled state of field modes in the left and right Rindler wedges. This picture is widely used to explain the Unruh effect, motivate entanglement entropy calculations, and connect quantum field theory to black hole thermodynamics and AdS/CFT. However, we show that this interpretation, while elegant, is not exact. We explicitly compute two-point functions and their derivatives for a massless scalar field in two-dimensional Minkowski space, comparing results obtained from canonical quantization with those obtained by assuming a TFD form of the vacuum. Mixed-derivative correlators agree perfectly, but higher-derivative correlators show systematic mismatches that persist even for points well away from horizons and are not removed by infrared regularization. To further test this picture, we construct an alternate coordinate system that divides Minkowski spacetime into two disconnected regions, apply the same derivation that leads to the standard TFD expression, and obtain a new "entangled-state" representation of the vacuum that is not thermal. This demonstrates that the appearance
We propose a model theoretic interpretation of the theorems about the equivalence between mixed characteristic perfectoid spaces and their tilts.
Deep neural networks have been well-known for their superb handling of various machine learning and artificial intelligence tasks. However, due to their over-parameterized black-box nature, it is often difficult to understand the prediction results of deep models. In recent years, many interpretation tools have been proposed to explain or reveal how deep models make decisions. In this paper, we review this line of research and try to make a comprehensive survey. Specifically, we first introduce and clarify two basic concepts -- interpretations and interpretability -- that people usually get confused about. To address the research efforts in interpretations, we elaborate the designs of a number of interpretation algorithms, from different perspectives, by proposing a new taxonomy. Then, to understand the interpretation results, we also survey the performance metrics for evaluating interpretation algorithms. Further, we summarize the current works in evaluating models' interpretability using "trustworthy" interpretation algorithms. Finally, we review and discuss the connections between deep models' interpretations and other factors, such as adversarial robustness and learning from in
Most physicists do not have patience for reading long and obscure interpretation arguments and disputes. Hence, to attract attention of a wider physics community, in this paper various old and new aspects of quantum interpretations are explained in a concise and simple (almost trivial) form. About the "Copenhagen" interpretation, we note that there are several different versions of it and explain how to make sense of "local non-reality" interpretation. About the many-world interpretation, we explain that it is neither local nor non-local, that it cannot explain the Born rule, that it suffers from the preferred basis problem, and that quantum suicide cannot be used to test it. About the Bohmian interpretation, we explain that it is analogous to dark matter, use it to explain that there is no big difference between non-local correlation and non-local causation, and use some condensed-matter ideas to outline how non-relativistic Bohmian theory could be a theory of everything. We also explain how different interpretations can be used to demystify the delayed choice experiment, to resolve the problem of time in quantum gravity, and to provide alternatives to quantum non-locality. Finall
Moderhai Milgrom's Modified Newtonian Dynamics (MOND) correction to Newtonian gravity or inertia is shown to be equivalent to a more fundamental formulation considering a non-inertial local reference frame and the fixed background of the observable universe, in the spirit of Mach's principle. Both Newton's gravitational constant $G\sim c^2/(M_u/R_u)$ and Milgrom's MOND acceleration scale constant $a_0\sim GM_u/R_u^2$ are replaced by two varying, measurable, and cosmological quantities determined by the causally connected mass and size of the universe. They arise from an inverse and an inverse squared distance scalar fields of matter density, respectively. This Machian interpretation of MOND is invariant under global rescalings of mass, length, and time across all regimes and is free from fundamental constants and free parameters, except for the speed of light. Machian MOND satisfies the fundamental consequences of Mach's principle not featured in Newton's and Einstein's theories: the decrease of inertia of a body when masses are removed from its neighborhood, and in the absence of a cosmic background, rotational motion is undefined up to the speed of light. Consequently, Machian MO
Network interpretation as an effort to reveal the features learned by a network remains largely visualization-based. In this paper, our goal is to tackle semantic network interpretation at both filter and decision level. For filter-level interpretation, we represent the concepts a filter encodes with a probability distribution of visual attributes. The decision-level interpretation is achieved by textual summarization that generates an explanatory sentence containing clues behind a network's decision. A Bayesian inference algorithm is proposed to automatically associate filters and network decisions with visual attributes. Human study confirms that the semantic interpretation is a beneficial alternative or complement to visualization methods. We demonstrate the crucial role that semantic network interpretation can play in understanding a network's failure patterns. More importantly, semantic network interpretation enables a better understanding of the correlation between a model's performance and its distribution metrics like filter selectivity and concept sparseness.
Understanding in-context learning (ICL) capability that enables large language models (LLMs) to excel in proficiency through demonstration examples is of utmost importance. This importance stems not only from the better utilization of this capability across various tasks, but also from the proactive identification and mitigation of potential risks, including concerns regarding truthfulness, bias, and toxicity, that may arise alongside the capability. In this paper, we present a thorough survey on the interpretation and analysis of in-context learning. First, we provide a concise introduction to the background and definition of in-context learning. Then, we give an overview of advancements from two perspectives: 1) a theoretical perspective, emphasizing studies on mechanistic interpretability and delving into the mathematical foundations behind ICL; and 2) an empirical perspective, concerning studies that empirically analyze factors associated with ICL. We conclude by highlighting the challenges encountered and suggesting potential avenues for future research. We believe that our work establishes the basis for further exploration into the interpretation of in-context learning. Addit
The Critical Assessment of Genome Interpretation (CAGI) aims to advance the state of the art for computational prediction of genetic variant impact, particularly those relevant to disease. The five complete editions of the CAGI community experiment comprised 50 challenges, in which participants made blind predictions of phenotypes from genetic data, and these were evaluated by independent assessors. Overall, results show that while current methods are imperfect, they have major utility for research and clinical applications. Missense variant interpretation methods are able to estimate biochemical effects with increasing accuracy. Performance is particularly strong for clinical pathogenic variants, including some difficult-to-diagnose cases, and extends to interpretation of cancer-related variants. Assessment of methods for regulatory variants and complex trait disease risk is less definitive, and indicates performance potentially suitable for auxiliary use in the clinic. Emerging methods and increasingly large, robust datasets for training and assessment promise further progress ahead.
The Aharonov-Bohm (A-B) effect showed that the phase of electron wave pattern could be changed by the excluded electromagnetic field, the region where electromagnetic field is zero. This apparent non-local effect has been explained by mainly two salient interpretations called "the interpretation of electromagnetic potentials" and "the interpretation of interaction energy". In this paper the author reviews and investigates both the interpretations and analysis their differences. Then the author clarifies in details and extend the approach of interaction energy interpretation and argue that as quantum mechanics involved energy of quantum systems and electromagnetic energy contained in electromagnetic field, the natural and more physically acceptable interpretation of the effect should be "the interpretation of interaction energy".