Transformers infer tasks from context via two modes, geometrically shaped task vectors explain their behavior.
problem Understanding how transformers infer tasks from context and the geometric properties of task vectors.
method Synthetic setting to train small transformers, mathematical characterization of task-vector geometry and inference modes.
result Task-vector geometry shapes in-distribution and out-of-distribution behavior of transformers.
COCOA can learn new tasks without forgetting previously learned ones in a distributed setting.
problem Continual learning with tasks arriving sequentially.
method Distributed optimization with COCOA algorithm.
result COCOA can perform continual learning without forgetting.
Multi-task learning aims to learn multiple tasks jointly by exploiting their relatedness to improve the generalization performance for each task. Traditionally, to perform multi-task learning, one needs to centralize data from all the tasks to a single machine. However, in many real-world applications, data of differen…
Transformers learn to generalize out-of-distribution with diverse pretraining tasks.
problem Conditions for pretrained transformers to generalize out-of-distribution.
method Empirical study of task diversity and pretraining distribution.
result As task diversity increases, transformers transition from specialized to generalized solutions.
Gradient-based meta-learners such as MAML are able to learn a meta-prior from similar tasks to adapt to novel tasks from the same distribution with few gradient updates. One important limitation of such frameworks is that they seek a common initialization shared across the entire task distribution, substantially limiti…
A meta-model is trained on a distribution of similar tasks such that it learns an algorithm that can quickly adapt to a novel task with only a handful of labeled examples. Most of current meta-learning methods assume that the meta-training set consists of relevant tasks sampled from a single distribution. In practice, …
DEN learns diverse tasks to generalize to unseen tasks.
problem Generalization from a diverse set of classification tasks with limited data.
method Three-block architecture: covariate transformation, distribution embedding, and classification.
result DEN outperforms existing methods in various synthetic and real tasks.
ST-MAML tackles task ambiguity in meta-learning by encoding tasks with stochastic representations.
problem Handling tasks from multiple distributions is challenging for meta-learning due to task ambiguity.
method ST-MAML uses a stochastic neural network module to encode tasks and propagate task representations to revise input variable encoding.
result ST-MAML matches or outperforms state-of-the-art methods on various tasks.
A distributed algorithm for online multi-task learning reduces communication and runtime costs.
problem Heavy communication and high runtime complexity in online multi-task learning.
method Adaptive primal-dual algorithm that synchronizes data across geographically distributed tasks.
result The proposed algorithm achieves optimal regret and is effective on real-world datasets.
Defines a similarity measure for classification distributions.
problem Measuring similarity between classification distributions.
method Proposes task similarity, a novel measure quantifying performance of source distributions on target distributions.
result Empirical task similarity correlates with transfer efficiency and semantic similarity of source distributions.
New research reveals how the pretraining distribution affects in-context learning in large language models.
problem Understanding how the pretraining distribution influences in-context learning in large language models.
method Developed a theoretical framework to characterize the relationship between pretraining distribution properties and in-context learning performance.
result Characterized a fundamental trade-off between robust task selection and generalization in ICL due to the pretraining distribution's statistical properties.
Gradient-based meta-RL fails with incorrect task distributions, leading to instability and poor performance.
problem Gradient-based meta-RL's sensitivity to task distributions causes instability and poor performance.
method Proposes meta Active Domain Randomization (meta-ADR) to learn task distributions for gradient-based meta-RL.
result Meta-ADR improves stability and generalization of MAML on simulated locomotion and navigation tasks.
Enhances cooperative multi-task SemCom for distributed users.
problem Performance degradation in cooperative multi-tasking due to negative information transfer.
method Federated learning (FL) with semantic-aware task clustering.
result Constructive cooperation across distributed users with semantic-aware task clustering.
New algorithm for online meta-learning with task boundary detection.
problem Adapting to new tasks in a non-stationary environment.
method Two detection mechanisms for task switches and distribution shift; online model updates based on current data.
result Achieves sublinear task-averaged regret under mild conditions.
Model-agnostic meta-learners aim to acquire meta-learned parameters from similar tasks to adapt to novel tasks from the same distribution with few gradient updates. With the flexibility in the choice of models, those frameworks demonstrate appealing performance on a variety of domains such as few-shot image classificat…
ATA optimizes task allocation in distributed machine learning.
problem Greedy task allocation leads to inefficiencies in distributed machine learning.
method Adaptive Task Allocation (ATA) adapts to unknown computation time distributions.
result ATA identifies optimal task allocation without prior knowledge of computation times.
Despite huge success, deep networks are unable to learn effectively in sequential multitask learning settings as they forget the past learned tasks after learning new tasks. Inspired from complementary learning systems theory, we address this challenge by learning a generative model that couples the current task to the…
AdMRL improves meta-reinforcement learning by minimizing worst-case sub-optimality gap.
problem Meta-reinforcement learning's sensitivity to task distribution shift.
method Model-based adversarial approach with minimax objective and alternating optimization.
result Efficacy in worst-case performance, generalization to out-of-distribution tasks, and sample efficiency.
Paper tackles Byzantine resilience in distributed multi-task learning.
problem Resilience of distributed algorithms in the presence of Byzantine agents.
method Online weight assignment rule based on accumulated loss and filtering.
result Aggregation with proposed weight assignment rule improves expected regret.
New algorithm optimizes privacy and utility in multi-task learning with skewed data.
problem Privacy constraints in multi-task learning with uneven data distribution.
method Adaptive reweighting of privacy budget allocation among tasks.
result Significant improvement in utility with state-of-the-art performance on benchmarks.
We propose methods for distributed graph-based multi-task learning that are based on weighted averaging of messages from other machines. Uniform averaging or diminishing stepsize in these methods would yield consensus (single task) learning. We show how simply skewing the averaging weights or controlling the stepsize a…
The paper proposes a learning algorithm that improves adaptability and generalization.
problem Improving adaptability and generalization in learning models.
method Learning to meta-learn by meta-finetuning on related tasks before adapting to specific tasks.
result Learning to meta-learn improves adaptability and generalization across various tasks.
We study the problem of distributed multi-task learning with shared representation, where each machine aims to learn a separate, but related, task in an unknown shared low-dimensional subspaces, i.e. when the predictor matrix has low rank. We consider a setting where each task is handled by a different machine, with sa…
Algorithm learns new tasks efficiently from past experience.
problem Lack of robustness to distributional shift in meta-reinforcement learning.
method Model Identification and Experience Relabeling (MIER) using dynamics models.
result Efficient extrapolation to out-of-distribution tasks.
Meta-learning improves with explicit modeling of task covariate distributions.
problem Ignoring the relationship between task covariates and conditional distributions limits meta-learning performance.
method Introducing a hierarchical Bayesian model that leverages samples from the marginal task covariates to better infer optimal parameters.
result Our method outperforms initialization-based meta-learning on popular classification benchmarks.
We tackle the Multi-task Batch Reinforcement Learning problem. Given multiple datasets collected from different tasks, we train a multi-task policy to perform well in unseen tasks sampled from the same distribution. The task identities of the unseen tasks are not provided. To perform well, the policy must infer the tas…
Proposes a FoE prior for improving CNN performance in distribution shifts.
problem Improving CNN performance in image analysis tasks with distribution shifts.
method Uses a field-of-experts (FoE) prior to match feature distributions of test and training images.
result Outperforms previous TTA methods in lesion segmentation and most healthy tissue segmentation tasks.
The paper analyzes meta-learning in a Gaussian setting, providing bounds and matching algorithms.
problem Understanding how task distributions influence transfer risk in meta-learning.
method Fixed design linear regression with Gaussian noise, Gaussian task distribution, weighted biased regularized regression method.
result A novel weighted version of biased regularized regression method matches distribution-dependent lower bounds on transfer risk up to a constant factor.
New MAML variant prioritizes hardest tasks, improving robustness.
problem Meta-learning's focus on average performance, ignoring worst-case.
method Reformulate MAML to minimize max loss over observed tasks.
result Task-robust model performs well across various task distributions.
OCEAN infers online task identities from context variables.
problem Online task inference for compositional tasks with context adaptation.
method Variational inference framework OCEAN models global and local context variables in a joint latent space.
result OCEAN provides more effective task inference with sequential context adaptation.
This work adopts the very successful distributional perspective on reinforcement learning and adapts it to the continuous control setting. We combine this within a distributed framework for off-policy learning in order to develop what we call the Distributed Distributional Deep Deterministic Policy Gradient algorithm, …
Shifts dataset evaluates uncertainty in real-world tasks across modalities.
problem Lack of standard datasets for evaluating uncertainty estimation and robustness to distributional shift.
method Proposes Shifts Dataset for evaluation of uncertainty estimates and robustness to distributional shift across tabular, audio, text, and sensor data.
result Baseline results for tabular weather prediction, machine translation, and SDC vehicle motion prediction.
Reweighting training data to better represent new tasks.
problem Deploying machine learning models to new tasks is challenging due to training data distribution.
method Formulate an exponential tilt distribution shift model and learn train data importance weights to minimize KL divergence.
result The learned train data weights improve target performance evaluation, fine-tuning, and model selection.
We consider the problem of distributed multi-task learning, where each machine learns a separate, but related, task. Specifically, each machine learns a linear predictor in high-dimensional space,where all tasks share the same small support. We present a communication-efficient estimator based on the debiased lasso and…
New estimator improves mutual information estimation.
problem Estimating mutual information in data science and machine learning.
method Proposes a new estimator that uses a preliminary estimate of the data distribution.
result A preliminary estimate helps in estimating mutual information more accurately.
Transformers learn to adapt to different task difficulties and resist distribution shifts.
problem Understanding and optimizing a Transformer's performance across various task difficulties and distribution shifts.
method Analyzing a pretrained Transformer on a mixture distribution of tasks, proving optimal convergence rates.
result Transformers achieve optimal convergence rates on tasks of specific difficulty levels, robust to distribution shifts.
Enhances DIM to match learned representations to a specific distribution.
problem Learning representations conforming to a specific distribution.
method Injecting noise into normalized outputs of the encoder while keeping the InfoMax training objective.
result Learning uniformly and normally distributed representations, as well as representations of other absolutely continuous distributions.
This paper tackles online reinforcement learning for unseen tasks with unknown boundaries.
problem Real-world tasks violate assumptions of task distributions, independence, and clear task delineations.
method A mixture of Gaussian Processes models different dynamics, and a transition prior handles temporal dependencies.
result The approach reliably handles task distribution shifts and outperforms alternatives in non-stationary tasks.
Meta-learning reformulated as Bayesian risk minimization.
problem Learning models to quickly adapt to new tasks from small datasets.
method Formalized meta-learning as Bayesian risk minimization, using a probabilistic framework to compute predictive distributions from posterior distributions of latent variables conditioned on contextual datasets.
result A novel Gaussian approximation for the posterior distribution that converges to maximum likelihood estimates and outperforms Neural Process on benchmark datasets.
TA-VAAL improves active learning by better utilizing task structures and overall data distribution.
problem High labeling cost limits deep learning applications; active learning selects informative samples.
method Task-aware variational adversarial active learning (TA-VAAL) modifies VAAL by relaxing task loss prediction and using ranking loss information.
result TA-VAAL outperforms state-of-the-arts on various datasets, including balanced and imbalanced labels.
Unified approach to non-standard classification tasks.
problem Non-standard classification tasks like semi-supervised, positive-unlabelled, multi-positive-unlabelled and noisy-label learning.
method Probabilistic, unified approach training a classifier to predict label-distributions, then inferring class-distributions.
result Unified model for various non-standard classification tasks.
We examine the influence of input data representations on learning complexity. For learning, we posit that each model implicitly uses a candidate model distribution for unexplained variations in the data, its noise model. If the model distribution is not well aligned to the true distribution, then even relevant variati…
Meta learns low-rank covariance factors for better uncertainty estimation.
problem Sub-optimal covariance matrices in multi-task settings.
method Meta learns diagonal or diagonal plus low-rank factors using an attentive set encoder.
result Efficiently constructed task-specific covariance matrices improve uncertainty estimation.
Existing generalization theories analyze the generalization performance mainly based on the model complexity and training process. The ignorance of the task properties, which results from the widely used IID assumption, makes these theories fail to interpret many generalization phenomena or guide practical learning tas…
Proves subgaussian distributions are SoS-certifiably subgaussian, enabling efficient algorithms for various statistical tasks.
problem Efficiently learning from subgaussian distributions in high dimensions.
method Universal constant C and polynomial sum of squares (SoS) approach. result Proves subgaussian distributions are SoS-certifiably subgaussian.
Meta-learning improves OoD detection with minimal in-distribution data.
problem Efficient OoD detection with limited in-distribution data.
method Meta-learning in latent space with Gaussian mixture models.
result Meta-learning enhances OoD detection performance.
Deep networks, composed of multiple layers of hierarchical distributed representations, tend to learn low-level features in initial layers and transition to high-level features towards final layers. Paradigms such as transfer learning, multi-task learning, and continual learning leverage this notion of generic hierarch…
Regression Prior Networks improve ensemble performance on regression tasks.
problem Improving ensemble performance on regression tasks.
method Extending Prior Networks and Ensemble Distribution Distillation (EnD2) to regression tasks using the Normal-Wishart distribution. result Regression Prior Networks yield performance competitive with ensemble approaches on regression tasks.