APD method decomposes neural network parameters into simple, faithful components.
problem Understanding the internal mechanisms learned by neural networks.
method Attribution-based Parameter Decomposition (APD) method.
result Demonstrated effectiveness in recovering features, separating computations, and identifying representations.
This paper tackles interpretability of LLMs in finance.
problem Complexity and lack of transparency of LLMs in finance.
method Mechanistic interpretability to understand LLM behavior.
result Demonstrates practical relevance of mechanistic interpretability in financial use cases.
Hybrid model learns interpretable meal-level glycemic control.
problem Lack of flexible, interpretable meal-level glycemic control methods.
method Hybrid variational autoencoder grounding latent space to mechanistic differential equation.
result Unsupervised representation discovers separation between individuals based on disease severity.
Unified model bridges mechanistic and non-mechanistic epidemic approaches.
problem Understanding the dynamics of epidemics with flexibility and interpretability.
method A mixture-based model representing time series as Gaussian mixtures, derived from a networked SIR framework.
result The model provides interpretable parameters and low prediction error, validating its use in understanding interventions.
SAEs struggle with feature consistency across runs, hindering MI reliability.
problem Inconsistency of learned SAE features across different training runs.
method Propose using the Pairwise Dictionary Mean Correlation Coefficient (PW-MCC) to measure feature consistency.
result High levels of feature consistency (0.80 for TopK SAEs on LLM activations) are achievable with appropriate architectural choices.
Triangulation filters spurious circuits in multilingual models.
problem Unreliable explanations of multilingual models across languages.
method Formalizes reference families and introduces triangulation as a causal acceptance rule.
result Triangulation provides a falsifiable standard for mechanistic claims.
SGNNs use simulations to train neural networks, improving scientific forecasting and interpretability.
problem Combining precise theory and machine learning for robust scientific modeling.
method Pretraining neural networks on diverse mechanistic simulations as training data.
result SGNNs outperform data-driven and physics-constrained models in forecasting and interpretability.
Paper develops efficient Bayesian inference for enzymatic SRNs with LNA metamodel.
problem Bayesian inference for nonlinear SDE-based mechanistic models with partial observations and measurement errors.
method Interpretable Bayesian updating LNA metamodel and efficient posterior sampling.
result Proposed approach demonstrates promising performance in empirical studies.
Breiman's data analysis dichotomy is outdated, offering a third approach: mechanistic models.
problem Data analysis dichotomy between data modelers and algorithmic modelers.
method Interpolating between simple interpretable models and flexible function approximations using mechanistic models.
result Flexible, interpretable, and scientifically-informed hybrids can provide accurate and robust predictions.
Study detects and explains positional bias in financial LLMs.
problem Positional bias in financial decision-making using LLMs.
method Unified framework and benchmark for detecting and quantifying bias in Qwen2.5 models.
result Positional bias is pervasive, scale-sensitive, and resurfaces under nuanced prompt designs.
Gradient descent optimization improved by circuit perspective.
problem Efficient training of large AI systems.
method Utilizing circuit perspective for mechanistic interpretability.
result Designing a curriculum for efficient learning.
Bilinear MLPs offer a new way to interpret deep learning models without complex nonlinearities.
problem Lack of mechanistic understanding in how MLPs compute.
method Introduced bilinear MLPs without element-wise nonlinearities, analyzed their weights using tensor and eigendecomposition.
result Bilinear MLPs provide interpretable weight structures and enable adversarial attacks and overfitting analysis.
New framework evaluates model explanations based on decision task improvement.
problem Evaluation of model explanations often misses practical value.
method Decision-theoretic framework quantifying three key values.
result Provides benchmarks and interprets human-AI decision support.
Machine learning predicts COVID-19 activity in China.
problem Real-time forecasting of COVID-19 activity in Chinese provinces.
method Combines mechanistic disease models with digital traces (internet searches, news alerts). Uses clustering and data augmentation techniques.
result Stable and accurate forecasts 2 days ahead of current time, outperforming baseline models in 27 out of 32 provinces.
Physics-informed machine learning models improve biomolecular system simulations.
problem Modeling unresolved interactions beyond classical force fields.
method Physics-informed neural networks and operator learning.
result Accurate, mechanistic, generalizable models for long-timescale kinetics.
Study forecasts cholera outbreaks in Malawi using dynamic models.
problem Cholera transmission forecasting in developing countries.
method Qualitative dynamics, Monte Carlo Markov Chain, sensitivity analysis, machine learning.
result Enhanced cholera forecasting models improve future trends prediction.
Statistical models outperform mechanistic models in short-term COVID-19 incidence forecasts.
problem Comparing accuracy of mechanistic vs statistical models for short-term COVID-19 incidence forecasts.
method Empirical comparison of forecasts from mechanistic and statistical models using daily incidence data from six US states.
result Statistical models are at least as accurate as mechanistic models and better capture volatility.
Unified framework for causal models at different levels of abstraction.
problem Relating causal models at varying levels of abstraction.
method Categorical framework using natural transformations between Markov functors.
result Generalized and unified causal abstractions with categorical proofs.
Unified understanding of neural networks on group operations verified.
problem Understanding and verifying neural networks trained on group operations.
method Investigated one-hidden-layer neural networks trained on binary operation of finite groups, revealing structure and providing a compact proof of model performance.
result Verified explanation applies to a large fraction of networks trained on the symmetric group S5, providing a >=95% accuracy bound for 45% of models.
Hybrid models combine domain knowledge and data-driven learning for Earth observation.
problem Challenges in modelling Earth observation data with either purely mechanistic or data-driven methods.
method Gaussian process convolution models, specifically latent force models (LFMs), integrating physical knowledge into multioutput GP models.
result Model automatically estimates soil moisture persistence and discovers latent forces related to precipitation.
New method sparsifies hybrid neural ODEs for better performance and stability.
problem Excessive latent states and interactions from mechanistic models lead to training inefficiency and over-fitting.
method Automatic state selection and structure optimization combining domain-informed graph modifications with data-driven regularization.
result Improved predictive performance and robustness with desired sparsity.
Bayesian approach for learning spatiotemporal systems from noisy data.
problem Efficiently modeling and learning from spatiotemporal dynamical systems with noisy data.
method Hierarchical state-space models with Gaussian process regression for efficient interpolation and training.
result Efficient modeling and learning of spatiotemporal dynamics using Bayesian methods.
IGSD separates task-specific content channels in transformer components by comparing activation replacement with zero ablation.
problem Mechanistic interpretability of transformer components
method IGSD: paired-intervention framework for comparing activation replacement with zero ablation
result IGSD identifies an early-layer content channel in transformer components that standard importance methods underestimate.
Adjoint SA speeds up bioprocess parameter learning.
problem Challenges in digital twin development for biomanufacturing.
method Adjoint sensitivity analysis on multi-scale enzymatic reaction networks.
result Resilient sensitivities reveal bioprocess regulatory mechanisms.
Develops adaptive algorithms for sustainable fertilizer use in agriculture.
problem Sustaining high yields while reducing environmental impacts of fertilizer use.
method Nonlinear model-based bandit algorithms linking biological processes to decision-making.
result Faster learning and higher profits with interpretable recommendations.
The roundworm C. elegans exhibits robust escape behavior in response to rapidly rising temperature. The behavior lasts for a few seconds, shows history dependence, involves both sensory and motor systems, and is too complicated to model mechanistically using currently available knowledge. Instead we model the process p…
nUDEs use neural networks to model biology without negative values.
problem Unrealistic negative values in hybrid models of biology.
method Developed non-negative UDEs (nUDEs) with regularization techniques.
result nUDEs provide realistic solutions for biological models.
The adoption of deep learning techniques in genomics has been hindered by the difficulty of mechanistically interpreting the models that these techniques produce. In recent years, a variety of post-hoc attribution methods have been proposed for addressing this neural network interpretability problem in the context of g…
Natural language processing (NLP) can be done using either top-down (theory driven) and bottom-up (data driven) approaches, which we call mechanistic and phenomenological respectively. The approaches are frequently considered to stand in opposition to each other. Examining some recent approaches in deep learning we arg…
The paper argues for prioritizing identifying structure over complex models for scientific discovery.
problem Underdetermination of mechanisms in high-dimensional data, leading to unreliable explanations.
method Proposes concrete standards for 'mechanistic ML' to avoid collapsing explanations.
result Large language models (LLMs) can collapse large equivalence classes of explanations, making it hard to distinguish between mechanisms.
New method uses Gaussian processes to improve PDE solver accuracy.
problem Uncertainty in PDE solver parameters and measurements.
method Physics-informed Gaussian process regression.
result Strictly generalizes weighted residual methods.
Unified framework for disentangled representations using mechanistic independence.
problem Identifiability of disentangled latent factors under statistical dependencies.
method Introduces mechanistic independence to characterize latent factors by their actions on observed variables, proposing various independence criteria.
result Establishes conditions for identifiability of latent subspaces without statistical assumptions.
Study relaxes identification assumptions for natural direct effects in non-randomized settings.
problem Identifying causal direct effects under unmeasured confounding.
method Developed relaxed conditions for identifying natural direct effects in non-randomized settings.
result Identified natural direct effect under unmeasured confounding conditions.
New methods for parameter estimation in mechanistic models using data-consistent inversion.
problem Parameter estimation bias in Bayesian analysis for mechanistic models.
method Data-consistent inversion methods based on rejection sampling, MCMC, GANs, and constrained optimization.
result Improved parameter estimation without bias from uninformative priors.
Optimizes biomanufacturing processes with a new digital twin calibration method.
problem Lack of interpretability and sample efficiency in traditional DoE methods.
method Developed a computational approach to calibrate Bio-SoS digital twin model.
result Guides sample-efficient and interpretable DoEs by quantifying sub-model parameter estimation errors.
New method extracts biological concepts from cell microscopy images.
problem Extracting meaningful concepts from vision foundation models trained on cell microscopy images.
method Sparse dictionary learning (DL) combined with PCA whitening pre-processing.
result Successfully retrieved biologically meaningful concepts like cell types and genetic perturbations.
This paper addresses anisotropy in Transformer models, providing geometric insights and empirical support.
problem Anisotropy phenomenon in Transformer models, challenging their geometric interpretation.
method Derive geometric arguments and use concept-based mechanistic interpretability during training.
result Activation-derived directions capture large gradient energy and a larger share of gradient anisotropy than normal controls.
Healthcare companies must submit pharmaceutical drugs or medical devices to regulatory bodies before marketing new technology. Regulatory bodies frequently require transparent and interpretable computational modelling to justify a new healthcare technology, but researchers may have several competing models for a biolog…
New insights into ML models' accuracy and generalization for scientific problems.
problem Quantifying accuracy and generalization of ML models in scientific applications.
method Rigorous numerical analysis and theoretical bounds for linear differential equations.
result Different ML models can have opposing generalization behaviors, contrary to intuition.
Framework for imputing missing heart data to simulate brain-heart interactions.
problem Lack of multi-modal patient data representing heart and brain processes.
method Probabilistic framework for joint cardiac data imputation and mechanistic model personalization.
result Accurate imputation of missing cardiac features in incomplete datasets.
ICAM creates interpretable feature attribution maps for brain images.
problem Challenges in predicting class relevance from brain images due to heterogeneity and background variation.
method A VAE-GAN framework for disentangling class relevance from background features.
result FA maps generated by ICAM outperform baseline methods and support phenotype variation exploration.
Superposition in autoencoders leads to loss in simple models.
problem Mechanistic interpretability of neural networks
method Analyzing mathematical basis for superposition and providing bounds for reconstruction loss
result Upper and lower bounds for L2 reconstruction loss in the sparse regime
A theory of deep learning is emerging, focusing on training dynamics and statistics.
problem Develop a scientific theory to understand deep learning.
method Synthesize research into five areas: idealized settings, tractable limits, mathematical laws, hyperparameters, and universal behaviors.
result The emerging theory is a mechanics of the learning process, named learning mechanics.
To understand how rich dynamics emerge in neural populations, we require models exhibiting a wide range of activity patterns while remaining interpretable in terms of connectivity and single-neuron dynamics. However, it has been challenging to fit such mechanistic spiking networks at the single neuron scale to empirica…
Temporal Functional Circuits explain KAN forecasts with interpretable edge functions.
problem Lack of mechanistic explanations in KAN forecasting.
method Transform KAN edge functions into faithful, temporally grounded explanations using a gated residual KAN.
result Gated KAN achieves lower MSE than linear-only models on regime-switching signals.
This research formalizes uncertainty quantification for Universal Differential Equations models.
problem Quantifying uncertainties in Universal Differential Equations models.
method Formalized uncertainty quantification methods for UDEs, including frequentist and Bayesian approaches.
result Evaluation of ensemble, variational inference, and MCMC sampling methods for UDEs.
Operationalizes engagement as human behavior, linking theory and data.
problem Fuzziness of engagement concept.
method Formal framework, Melchoir Model, model comparison, theory-driven hypothesis.
result Engagement can be shaped and interpreted using data-driven methods.
Mechanistic models of single-neuron dynamics have been extensively studied in computational neuroscience. However, identifying which models can quantitatively reproduce empirically measured data has been challenging. We propose to overcome this limitation by using likelihood-free inference approaches (also known as App…