The paper shows that certain learned representations are identifiable in function space.
problem Identifiability of learned representations in deep neural networks.
method Using recent advances in nonlinear ICA, the paper shows that a large family of discriminative models are identifiable in function space, up to a linear indeterminacy.
result Many models for representation learning are identifiable in function space, including text, images, and audio.
Unsupervised learning models can be indistinguishable without identifiability, leading to unreliable representations.
problem Unsupervised learning models may be indistinguishable without identifiability, making it impossible to recover a ground truth generative model.
method Construction based on nonlinear independent component analysis theory to illustrate potential failure cases.
result Counterexamples show that identifiability is crucial for reliable unsupervised representation learning.
Unified framework for representation and causal structure learning using exchangeable data.
problem Identifying latent representations or causal structures in non-i.i.d. data.
method Identifiable Exchangeable Mechanisms (IEM) framework for representation and structure learning.
result New insights and identifiability results for causal structure and representation learning.
We identify action representations from video data, proving their statistical benefits.
problem Identifying latent action policies from video data.
method Entropy-regularized LAPO objective, formalizing desiderata for action representations.
result Entropy-regularized LAPO identifies action representations satisfying desiderata under suitable conditions.
SIGNNAP learns stable and identifiable node representations in GNNs against graph perturbations.
problem Fragility of GNN models to graph perturbations leading to unreliable node representations.
method SIGNNAP proposes a novel model that learns stable and identifiable node representations in an unsupervised manner, formalizing stability and identifiability through a contrastive objective and preserving smoothness with existing GNN backbones.
result SIGNNAP demonstrates effectiveness in learning stable and identifiable node representations in GNNs against graph perturbations on six benchmarks.
Study identifies specialist representations from generalist models without parametric constraints.
problem Identify task-relevant latent representations from generalist models.
method Nonparametric, fully unsupervised approach, proving identifiability of task structure and latent representations.
result Identifiability of task structure and latent representations in a nonparametric setting.
This research improves neural network representation identifiability through task structures.
problem Improving neural network representation identifiability in multi-task settings.
method Analyzing the effects of task distributions and causal structures on latent factors, leading to simpler optimization.
result A straightforward optimization procedure enables better representation recovery in both synthetic and real-world data.
Unpaired multi-domain causal representation learning is possible with sufficient conditions.
problem Learning shared causal representation from unpaired data across domains.
method Identify sufficient conditions for joint distribution and shared causal graph recovery.
result Practical method to recover shared latent causal graph from marginal distributions.
Paper tackles intervention extrapolation using identifiable representations.
problem Predicting effects of unseen interventions on outcomes.
method Combines identifiable representation learning with autoencoders to enforce linear invariance.
result Identifiable representations enable non-linear extrapolation of interventions.
Identifiability, or recovery of the true latent representations from which the observed data originates, is de facto a fundamental goal of representation learning. Yet, most deep generative models do not address the question of identifiability, and thus fail to deliver on the promise of the recovery of the true latent …
New approach identifies latent properties from mechanisms, not just data.
problem Identifying latent properties from data generating processes.
method Equivariance perspective on identifiable representation learning.
result Identification of latent properties is possible up to shared equivariances in known mechanisms.
Study on robustness of unsupervised representation learning in slightly misspecified settings.
problem Identify nonlinear representation learning in slightly misspecified settings.
method Formalize and investigate Independent Component Analysis (ICA) with observations generated by a mixing function close to a local isometry.
result Approximate identifiability of nonlinear ICA with almost isometric mixing functions.
Defines metrics to compare neural network representations.
problem Comparing neural network representations across different architectures and tasks.
method Developed a family of metric spaces and modified existing measures to quantify representational dissimilarity.
result Identified relationships between neural representations and anatomical features.
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.
Identifier names convey useful information about the intended semantics of code. Name-based program analyses use this information, e.g., to detect bugs, to predict types, and to improve the readability of code. At the core of name-based analyses are semantic representations of identifiers, e.g., in the form of learned …
Method identifies latent variables from high-dimensional data with piecewise affine mixing.
problem Identifying latent variables from high-dimensional observations with dependencies and piecewise affine transformations.
method Proposes a two-stage method with sparsity and Gaussianity regularization.
result Effectively recovers ground-truth latent variables from synthetic and image data.
New findings show invariance alone isn't enough to identify latent causal variables.
problem Lack of theoretical insights for identifying latent causal variables when variables are latent.
method Assessed the connection between invariance and causal representation learning using impossibility results.
result Invariance alone is insufficient to identify latent causal variables.
A common workflow in data exploration is to learn a low-dimensional representation of the data, identify groups of points in that representation, and examine the differences between the groups to determine what they represent. We treat this workflow as an interpretable machine learning problem by leveraging the model t…
We explore the benefits of augmenting state-of-the-art model-free deep reinforcement algorithms with simple object representations. Following the Frostbite challenge posited by Lake et al. (2017), we identify object representations as a critical cognitive capacity lacking from current reinforcement learning agents. We …
Our work proves CSF can recover ground-truth features in RL, improving understanding of feature learning.
problem Understanding the role of representation and mutual information in reinforcement learning.
method Investigates Contrastive Successor Features (CSF) method for identifiable representation learning in reinforcement learning.
result Proves CSF can recover ground-truth features up to a linear transformation.
This paper tackles causal representation learning with linear and general transformations.
problem Identify and recover latent causal variables and graphs under unknown transformations.
method Score-based algorithms that use gradients of log-density functions for identifiability and achievability.
result Two stochastic hard interventions per node are sufficient for identifiability of general transformations.
New method identifies latent causal variables from observed data, overcoming indeterminacies.
problem Identifying latent causal variables from observed data, especially when latent variables are weight-variant.
method Introduces a novel identifiability condition for latent causal models, proposing SuaVE method.
result Identifies latent causal variables up to trivial permutation and scaling, demonstrating consistency and efficacy.
Identifies Anosov representations of hyperbolic triangle groups in SL(3,R).
problem Classifying Anosov representations of hyperbolic triangle groups into SL(3,R).
method Proving representations are Anosov if they lie in the Hitchin component or the Barbot component, with specific conditions for eigenvalues.
result Anosov representations in SL(3,R) have non-convex boundary maps.
Paper establishes identifiability and achievability for causal representation learning.
problem Identifying and recovering latent causal models and variables from observational and interventional data.
method Establishes identifiability and achievability using uncoupled interventions and a recovery algorithm.
result Guaranteed perfect recovery of latent causal model and variables under uncoupled interventions.
New findings clarify the link between distributional closeness and representational similarity.
problem When and why do different neural network representations become similar?
method Identifiability theory, focusing on model families including autoregressive language models.
result Small Kullback-Leibler divergence does not guarantee similar representations.
This paper tackles sequential distribution shifts in representation learning.
problem Learning meaningful representations in a sequence of distribution shifts.
method Nonlinear Independent Component Analysis (ICA) framework for continual causal representation learning.
result The method achieves performance comparable to joint training on multiple offline distributions and shows no benefit from the incoming new distribution on all latent variables.
New models learn stable latent clusters without side info.
problem Stability of non-linear ICA representations without side information.
method Deep generative models with latent clusterings, compared to standard VAEs and auxiliary labeled models.
result Deep generative models with latent clusterings are as stable as models with side information.
This work shows how disentangled and sparse representations improve multi-task learning.
problem Improving generalization in multi-task learning with disentangled and sparse representations.
method Proved a new identifiability result and proposed a practical approach using sparsity-promoting bi-level optimization.
result Maximally sparse base-predictors yield disentangled representations under certain conditions.
Proposes an extended disentanglement framework with new metrics.
problem Improving disentangled representations in representation learning.
method Connects DCI framework to identifiability, introduces new measures.
result Establishes a formal link between disentanglement and independent component analysis.
Identifies LA-groups via VB-group structure and complementary actions.
problem Understanding the structure and integrability of LA-groups.
method Identifies LA-groups via VB-group structure and complementary actions up to homotopy.
result Establishes an equivalence between LA-groups and LA-matched pairs.
Study identifies latent variables and causal relationships from multiple environments.
problem Identify latent variables and causal relationships from multiple environments.
method Proposes algorithm LiNGCReL for identifying causal graph up to surrounded-node ambiguity.
result Identifies latent variables up to surrounded-node ambiguity (SNA) in linear causal models.
A new VAE model improves disentanglement with identifiable latent factors.
problem Learning disentangled representations without inductive biases.
method A novel VAE with a conditional prior over latent variables.
result Superior performance in disentanglement metrics.
A method identifies domain-general features using causal graph constraints and regularization.
problem Identifying domain-general features without prior knowledge of spurious features.
method Proposes a novel regularization framework based on causal graph constraints.
result Demonstrates effectiveness in both synthetic and real-world data, outperforming state-of-the-art methods.
New method identifies causal relationships without strong assumptions.
problem Causal Representation Learning (CRL) is ill-posed due to representation and causal discovery issues.
method Identifiability based on grouping of observational variables, self-supervised estimation framework.
result Practical identifiability conditions without temporal structure, interventions, or weak supervision.
EBM reduces dimensionality for estimating heterogeneous CATEs.
problem Estimating CATEs requires many confounding variables, increasing sample complexity.
method Proposes an EBM that learns a low-dimensional representation of variables.
result EBM representations keep CATE estimates consistent and perform better than other methods.
iCITRIS learns causal variables from interactive systems with instantaneous effects.
problem Identifying causal variables from temporal sequences with instantaneous effects.
method iCITRIS method for causal representation learning that handles instantaneous effects in intervened temporal sequences.
result iCITRIS accurately identifies causal variables and their causal graph from three interactive system datasets.
This paper tackles CRL for multi-node interventions, achieving identifiability guarantees.
problem CRL under unknown multi-node interventions, focusing on single-node assumptions.
method Establishes identifiability results for general latent causal models under stochastic interventions.
result Identifiability up to ancestors using soft interventions, perfect identifiability using hard interventions.
New framework TDRL identifies latent causal variables from sequential data.
problem Identify latent causal variables from sequential data.
method Proposes TDRL framework to recover time-delayed latent causal variables and identify their relations from measured sequential data.
result Identifies latent causal variables reliably from sequential data.
Optimal feature transfer identified through bias-variance analysis.
problem Optimizing feature transfer in transfer learning.
method Simple linear model with fine-grained bias-variance decomposition.
result Optimal pretrained feature transform is naturally sparse.
COCA refactors training data to identify and erase unsafe concepts in LLMs.
problem Identifying and erasing unsafe concepts in Large Language Models (LLMs) for safety alignment.
method Concept Concentration (COCA) refactors training data with an explicit reasoning process to identify and erase unsafe concepts.
result COCA significantly reduces both in-distribution and out-of-distribution jailbreak success rates while maintaining strong performance on regular tasks.
Study identifies causal relationships without direct supervision from unknown interventions.
problem Identify causal relationships from unknown interventions without direct supervision.
method General nonparametric setting with multiple datasets from unknown interventions.
result Identify ground truth latents and causal graph up to ambiguities.
The paper shows how to learn causal representations with few environments and finite samples.
problem Learning causal representations from limited data and environments.
method Explicit, finite-sample guarantees with a logarithmic number of interventions.
result Consistent recovery of latent causal graph, mixing matrix, and unknown intervention targets.
Study identifies latent variables and models from spacecraft data.
problem Learning reliable models from spacecraft data with complex relationships.
method Inductive bias inspired by controllable canonical forms for sparse, input-dependent latent variables.
result Identifies latent variables up to scaling and determines dynamic models up to transformations for linear and affine systems.
Logit distance bounds representational similarity of models.
problem Approximating linear similarity when distributions are close.
method Defined a logit distance and proved its relationship to representational dissimilarity.
result Logit distance bounds representational similarity, providing nontrivial control in practice.
New method identifies causal variables from partially observed data.
problem Learning from unpaired observations with instance-dependent partial observability.
method Proposes two methods enforcing sparsity in the inferred representation.
result Establishes two identifiability results for linear and piecewise linear mixing functions.
New method identifies latent relationships in deep models without additional constraints.
problem Latent representations in deep latent variable models are not statistically identifiable.
method Identifies relationships between latent variables (distances, angles, volumes) under mild model conditions.
result Empirically demonstrates more reliable latent distances without additional labeled data.
Geometric approach connects Burau representation to sphere metrics, identifying kernels.
problem Faithfulness of the Burau representation for the 4-strand case.
method Geometric and orbifold theory.
result Identifies the kernel of the Burau representation for some cases.
A new method identifies similar mutual funds using graph learning.
problem Identifying similar mutual funds with nuanced portfolio similarities.
method Node2Vec machine learning method applied to a weighted bipartite network of funds and assets.
result Identifies structural similarity among mutual funds' portfolios.