New method learns robust representations by modeling environment variation.
problem Learning invariant representations across varying environments.
method Explicitly modeling variation across environments and marginalizing it out.
result Proposed method outperforms invariant-learning methods in various settings.
A framework for learning disentangled representations of symmetric environments.
problem Discovering and modelling the underlying structure of environments.
method Group representation theory for disentangled representations of dynamical environments.
result Our method enables accurate long-horizon predictions and correlates with disentanglement quality.
Self-supervised methods, wherein an agent learns representations solely by observing the results of its actions, become crucial in environments which do not provide a dense reward signal or have labels. In most cases, such methods are used for pretraining or auxiliary tasks for "downstream" tasks, such as control, expl…
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.
Method learns attractive areas from agent motions to represent environments.
problem Representing environments based on moving agents' nonlinear motions.
method Switching model of velocity fields, parametric representation of attractive spots.
result Dynamic map of attractive areas for online learning.
Finding a generally accepted formal definition of a disentangled representation in the context of an agent behaving in an environment is an important challenge towards the construction of data-efficient autonomous agents. Higgins et al. recently proposed Symmetry-Based Disentangled Representation Learning, a definition…
Animals need to devise strategies to maximize returns while interacting with their environment based on incoming noisy sensory observations. Task-relevant states, such as the agent's location within an environment or the presence of a predator, are often not directly observable but must be inferred using available sens…
Paper shows regularization improves robustness in domain generalization.
problem Improving robustness in domain generalization.
method Derives novel theoretical analysis to control representation smoothness and proposes a regularization method.
result Regularization improves robustness in domain generalization.
New method recovers causal DAGs from general environments without strict assumptions.
problem Recovering causal DAGs from real-world data with varying distributions.
method Formalizes desiderata for causal representation learning in general environments, leveraging sufficient change conditions up to third-order derivatives.
result Fully recovers latent DAG and identifies latent variables up to minor indeterminacies under nonparametric mixing.
GALA framework learns invariant graph representations via environment augmentation with minimal assumptions.
problem Learning invariant graph representations from different environments without additional assumptions.
method Developed GALA framework with minimal assumptions of variation sufficiency and consistency. Uses an assistant model to differentiate graph environment changes.
result Extracting maximally invariant subgraphs to proxy predictions identifies underlying invariant subgraphs for successful out-of-distribution generalization.
Unsupervised representation learning has succeeded with excellent results in many applications. It is an especially powerful tool to learn a good representation of environments with partial or noisy observations. In partially observable domains it is important for the representation to encode a belief state, a sufficie…
PFedRL-Rep learns shared and personalized policies for heterogeneous environments.
problem Poor performance of single policy in heterogeneous environments.
method Develops PFedRL-Rep framework with shared feature representation and personalized weights.
result Proves linear convergence speedup with respect to the number of agents.
A method for self-supervised representation learning in partially observable environments.
problem Sparse rewards and stochasticity in partially observable environments.
method World model in latent space for estimating missing information.
result Significant improvement in exploration compared to prior work.
Paper introduces automatic learning of internal representations for better generalization.
problem Improving machine learning by biasing hypothesis space.
method Learning internal representation for a learning environment, then using it to bias hypothesis space for future tasks.
result Representation learning can drastically reduce the number of examples needed per task.
In model-based reinforcement learning, generative and temporal models of environments can be leveraged to boost agent performance, either by tuning the agent's representations during training or via use as part of an explicit planning mechanism. However, their application in practice has been limited to simplistic envi…
State representation learning aims at learning compact representations from raw observations in robotics and control applications. Approaches used for this objective are auto-encoders, learning forward models, inverse dynamics or learning using generic priors on the state characteristics. However, the diversity in appl…
New approach for obstacle avoidance in robotics using learned representations.
problem Challenges in sensor-based motion planning for new and dynamic environments.
method Proposes a new obstacle representation using PointNet architecture trained jointly with policies for obstacle avoidance.
result Significant improvements in accuracy and efficiency compared to state of the art.
This thesis argues for learning representations from multiple similar tasks for better generalization.
problem Insufficient information in a single task for good generalization.
method Introduces the concept of an environment of a learner and uses a sample from it to learn a representation.
result Learning a representation from n tasks reduces the number of examples required for each task by a factor of n. 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.
ICIL learns policies invariant to multiple environments, improving generalization.
problem Learning policies from multiple environments leads to spurious correlations.
method ICIL learns invariant feature representations and a matching imitation policy.
result ICIL policies generalize better to unseen environments.
When agents interact with a complex environment, they must form and maintain beliefs about the relevant aspects of that environment. We propose a way to efficiently train expressive generative models in complex environments. We show that a predictive algorithm with an expressive generative model can form stable belief-…
The paper proposes a deep learning technique for structured and composable representations.
problem Learning structured and composable representations from input images and discrete labels.
method End-to-end deep learning to learn representations based on distance estimates between class label and contextual information.
result The representations have a clear structure allowing for class and environment decomposition.
Deep RL approach improves MIS for complex environments.
problem Improving off-policy evaluation for complex environments.
method Uses successor representation from deep RL to decouple reward and dynamics.
result Empirically stable and applicable to high-dimensional domains.
Animals (especially humans) have an amazing ability to learn new tasks quickly, and switch between them flexibly. How brains support this ability is largely unknown, both neuroscientifically and algorithmically. One reasonable supposition is that modules drawing on an underlying general-purpose sensory representation a…
OBSER framework infers sub-environments from objects, outperforming scene-based methods.
problem Zero-shot recognition of environments from object distributions.
method Bayesian framework using metric and self-supervised learning models to estimate object distributions in latent space.
result OBSER framework reliably performs inference in open-world and photorealistic environments, outperforming scene-based methods.
A generative recurrent neural network is quickly trained in an unsupervised manner to model popular reinforcement learning environments through compressed spatio-temporal representations. The world model's extracted features are fed into compact and simple policies trained by evolution, achieving state of the art resul…
Agent learns diverse hierarchical structures in unknown environments.
problem Autonomous discovery and learning of diverse structures in unknown changing environments.
method Progressive construction of a Hierarchy of Observation Latent Models for Exploration Stratification (HOLMES).
result Agent can learn and reuse representations to progressively explore and discover diverse structures.
New algorithms improve multi-task learning across different environments.
problem Learning across multiple tasks and adapting to unseen environments.
method Learning common representations, dynamic feature weighting via attention mechanism.
result Significant performance improvement on new environments, 1.5x faster.
Transfer learning benefits vary in DRL, affecting representation and policy learning.
problem Understanding how much of DRL's sample complexity is due to representation learning.
method Transfer learning experiments comparing performance on related games.
result Benefits of transfer learning are highly variable and non-symmetric across tasks.
Intrinsically motivated goal exploration processes enable agents to autonomously sample goals to explore efficiently complex environments with high-dimensional continuous actions. They have been applied successfully to real world robots to discover repertoires of policies producing a wide diversity of effects. Often th…
Self-supervised reward prediction improves RL in sparse reward settings.
problem Data efficiency and sparse reward signals in reinforcement learning.
method Learning a state representation for reward prediction and using it to shape rewards.
result Self-supervised reward prediction enhances RL algorithms in single-goal environments.
We consider the problem of building a state representation model in a continual fashion. As the environment changes, the aim is to efficiently compress the sensory state's information without losing past knowledge. The learned features are then fed to a Reinforcement Learning algorithm to learn a policy. We propose to …
GIB improves neural network generalization by dynamically selecting task-relevant features across different sequential environments.
problem Poor generalization of deep neural networks to unseen environments.
method Proposes a gated information bottleneck (GIB) approach that dynamically drops spurious correlations and selects task-relevant features.
result GIB outperforms other IB approaches in adversarial robustness and OOD detection.
AdaRL adapts quickly to new environments with minimal data.
problem Quickly adapting to new environments in reinforcement learning.
method AdaRL uses a parsimonious graphical representation to encode changes across domains.
result AdaRL can efficiently adapt policies to target domains with few samples.
New RL environments help AI learn causal relationships from visual data.
problem Learning causal relationships from visual data for AI agents.
method Designing benchmark RL environments and evaluating representation learning algorithms.
result Explicitly incorporating structure and modularity improves causal induction in model-based RL.
C-SWMs learn structured world models from raw data.
problem Learning structured world models from raw sensory data.
method Contrastive learning with graph neural networks.
result C-SWMs outperform typical models in structured environments.
Learning to cooperate is crucially important in multi-agent environments. The key is to understand the mutual interplay between agents. However, multi-agent environments are highly dynamic, where agents keep moving and their neighbors change quickly. This makes it hard to learn abstract representations of mutual interp…
Tree-based regularization improves latent variable inference from related datasets.
problem Inferring latent variables from multiple related datasets in causal systems.
method Tree-Based Regularization (TBR) for sparse changes across environments.
result TBR identifies true latent variables up to simple transformations under sparse changes.
Proto-value networks improve deep reinforcement learning representations using auxiliary tasks.
problem Improving deep reinforcement learning representations with auxiliary tasks.
method Derived a new family of auxiliary tasks based on the successor measure, combined with off-policy learning rule.
result Proto-value networks produce rich features comparable to established algorithms using only linear approximation and a small number of interactions.
New method identifies causal variables from multi-node interventions, expanding on previous single-node approaches.
problem Inferring high-level causal variables from low-level observations under multiple interventions.
method Exploits variance trace of ground truth causal variables and regularizes for sparsity.
result First identifiability result for causal representation learning with multiple node interventions.
BYOL-Explore learns to explore visually-rich environments by predicting world dynamics.
problem Exploration in visually complex environments.
method Optimizes a single prediction loss in latent space to learn world representation, dynamics, and exploration policy.
result Achieves superhuman performance on Atari games with simpler design.
Representation learning algorithms are designed to learn abstract features that characterize data. State representation learning (SRL) focuses on a particular kind of representation learning where learned features are in low dimension, evolve through time, and are influenced by actions of an agent. The representation i…
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.
Paper uses deep reinforcement learning for optimal stock portfolio management.
problem Optimizing stock portfolio choices in complex market environments.
method Direct deep reinforcement learning to learn factor representations and make optimal decisions.
result Deep learning outperforms average market performance in portfolio allocation.
Bayesian approach learns causal concepts from diverse social surveys.
problem Inferring causal concepts from heterogeneous data with sparse changes.
method Hierarchical Bayesian model with sequential Monte Carlo sampling.
result Model infers meaningful causal concepts and plausible relations.
Proposes a new method to learn representations directly optimized for a task.
problem Learning representations optimized for unrelated tasks.
method Jointly learns representation and prediction function for a specific task.
result Learned representations outperform pre-trained ones and are more sample-efficient.
Intelligent agents can cope with sensory-rich environments by learning task-agnostic state abstractions. In this paper, we propose an algorithm to approximate causal states, which are the coarsest partition of the joint history of actions and observations in partially-observable Markov decision processes (POMDP). Our m…
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.