Automatic data selection and curriculum design for training large language models is challenging, with only a few existing methods showing improvements over standard training. Furthermore, current schemes focus on domain-level selection, overlooking the more fine-grained contributions of each individual training point. It is difficult to apply traditional datapoint selection methods on large language models: most online batch selection methods perform two-times forward or backward passes, which introduces considerable extra costs with large-scale models. To mitigate these obstacles, we propose irreducible curriculum as a curriculum learning algorithm for language model pretraining, which prioritizes samples with higher learnability. Specifically, to avoid prohibitive extra computation overhead, we simulate the sample loss along the main model's training trajectory using a small-scale proxy model. Our experiments on the RedPajama-1B dataset demonstrate a consistent improvement on validation perplexity across all 7 domains compared to random uniform baseline and the anti-curriculum strategy. Our method also reduces the sharpness of the network and illustrates a better 5-shot accuracy on MMLU benchmarks.

Lagrangian relaxation stands among the most efficient approaches for solving a Mixed Integer Linear Programs (MILP) with difficult constraints. Given any duals for these constraints, called Lagrangian Multipliers (LMs), it returns a bound on the optimal value of the MILP, and Lagrangian methods seek the LMs giving the best such bound. But these methods generally rely on iterative algorithms resembling gradient descent to maximize the concave piecewise linear dual function: the computational burden grows quickly with the number of relaxed constraints. We introduce a deep learning approach that bypasses the descent, effectively amortizing the local, per instance, optimization. A probabilistic encoder based on a graph convolutional network computes high-dimensional representations of relaxed constraints in MILP instances. A decoder then turns these representations into LMs. We train the encoder and decoder jointly by directly optimizing the bound obtained from the predicted multipliers. Numerical experiments show that our approach closes up to 85~\% of the gap between the continuous relaxation and the best Lagrangian bound, and provides a high quality warm-start for descent based Lagrangian methods.

Data preprocessing is a crucial part of any machine learning pipeline, and it can have a significant impact on both performance and training efficiency. This is especially evident when using deep neural networks for time series prediction and classification: real-world time series data often exhibit irregularities such as multi-modality, skewness and outliers, and the model performance can degrade rapidly if these characteristics are not adequately addressed. In this work, we propose the EDAIN (Extended Deep Adaptive Input Normalization) layer, a novel adaptive neural layer that learns how to appropriately normalize irregular time series data for a given task in an end-to-end fashion, instead of using a fixed normalization scheme. This is achieved by optimizing its unknown parameters simultaneously with the deep neural network using back-propagation. Our experiments, conducted using synthetic data, a credit default prediction dataset, and a large-scale limit order book benchmark dataset, demonstrate the superior performance of the EDAIN layer when compared to conventional normalization methods and existing adaptive time series preprocessing layers.

We now have rigorous convergence guarantees that, for certain well-behaved posteriors, BBVI achieves a convergence rate of (1), corresponding We prove that black-box variational inference to a computational complexity of (1)(Domke et al., (BBVI) with control variates, particularly 2023a; Kim et al., 2023b). A remaining theoretical question the sticking-the-landing(STL) estimator, is whether BBVI can achieve better rates, in particular converges at a geometric (traditionally called geometric convergence rates, which is traditionally "linear") rate under perfect variational family called "linear" convergence in the optimization literature specification. In particular, we prove a (see the textbook by Nesterov 2004, 1.2.3), correspondingtoacomplexityof(log(1)).

Developing efficient solutions for inference problems in intelligent sensor networks is crucial for the next generation of location, tracking, and mapping services. This paper develops a scalable distributed probabilistic inference algorithm that applies to continuous variables, intractable posteriors and large-scale real-time data in sensor networks. In a centralized setting, variational inference is a fundamental technique for performing approximate Bayesian estimation, in which an intractable posterior density is approximated with a parametric density. Our key contribution lies in the derivation of a separable lower bound on the centralized estimation objective, which enables distributed variational inference with one-hop communication in a sensor network. Our distributed evidence lower bound (DELBO) consists of a weighted sum of observation likelihood and divergence to prior densities, and its gap to the measurement evidence is due to consensus and modeling errors. To solve binary classification and regression problems while handling streaming data, we design an online distributed algorithm that maximizes DELBO, and specialize it to Gaussian variational densities with non-linear likelihoods. The resulting distributed Gaussian variational inference (DGVI) efficiently inverts a $1$-rank correction to the covariance matrix. Finally, we derive a diagonalized version for online distributed inference in high-dimensional models, and apply it to multi-robot probabilistic mapping using indoor LiDAR data.

The deep learning recipe of casting real-world problems as mathematical optimisation and tackling the optimisation by training deep neural networks using gradient-based optimisation has undoubtedly proven to be a fruitful one. The understanding behind why deep learning works, however, has lagged behind its practical significance. We aim to make steps towards an improved understanding of deep learning with a focus on optimisation and model regularisation. We start by investigating gradient descent (GD), a discrete-time algorithm at the basis of most popular deep learning optimisation algorithms. Understanding the dynamics of GD has been hindered by the presence of discretisation drift, the numerical integration error between GD and its often studied continuous-time counterpart, the negative gradient flow (NGF). To add to the toolkit available to study GD, we derive novel continuous-time flows that account for discretisation drift. Unlike the NGF, these new flows can be used to describe learning rate specific behaviours of GD, such as training instabilities observed in supervised learning and two-player games. We then translate insights from continuous time into mitigation strategies for unstable GD dynamics, by constructing novel learning rate schedules and regularisers that do not require additional hyperparameters. Like optimisation, smoothness regularisation is another pillar of deep learning's success with wide use in supervised learning and generative modelling. Despite their individual significance, the interactions between smoothness regularisation and optimisation have yet to be explored. We find that smoothness regularisation affects optimisation across multiple deep learning domains, and that incorporating smoothness regularisation in reinforcement learning leads to a performance boost that can be recovered using adaptions to optimisation methods.

This paper tackles the problem of the worstclass error rate, instead of the standard error rate averaged over all classes. For example, a three-class classification task with class-wise error rates of 10%, 10%, and 40% has a worst-class error rate of 40%, whereas the average is 20% under the class-balanced condition. The worst-class error is important in many applications. For example, in a medical image classification task, it would not be acceptable for the malignant tumor class to have a 40% error rate, while the benign and healthy classes have 10% error rates. We propose a boosting algorithm that guarantees an upper bound of the worst-class training error Figure 1: A toy example showing the average error and derive its generalization bound. Experimental minimization results in the case of a high worst-class results show that the algorithm lowers error. Note that all five classes have the same number worst-class test error rates while avoiding of instances, and thus, there is no class imbalance.

In the field of natural language processing, open-domain chatbots have emerged as an important research topic. However, a major limitation of existing open-domain chatbot research is its singular focus on short single-session dialogue, neglecting the potential need for understanding contextual information in multiple consecutive sessions that precede an ongoing dialogue. Among the elements that compose the context in multi-session conversation settings, the time intervals between sessions and the relationships between speakers would be particularly important. Despite their importance, current research efforts have not sufficiently addressed these dialogical components. In this paper, we introduce a new 1M multi-session dialogue dataset, called Conversation Chronicles, for implementing a long-term conversation setup in which time intervals and fine-grained speaker relationships are incorporated. Following recent works, we exploit a large language model to produce the data. The extensive human evaluation shows that dialogue episodes in Conversation Chronicles reflect those properties while maintaining coherent and consistent interactions across all the sessions. We also propose a dialogue model, called ReBot, which consists of chronological summarization and dialogue generation modules using only around 630M parameters. When trained on Conversation Chronicles, ReBot demonstrates long-term context understanding with a high human engagement score.

Imitation learning has been applied to a range of robotic tasks, but can struggle when robots encounter edge cases that are not represented in the training data (i.e., distribution shift). Interactive fleet learning (IFL) mitigates distribution shift by allowing robots to access remote human supervisors during task execution and learn from them over time, but different supervisors may demonstrate the task in different ways. Recent work proposes Implicit Behavior Cloning (IBC), which is able to represent multimodal demonstrations using energy-based models (EBMs). In this work, we propose Implicit Interactive Fleet Learning (IIFL), an algorithm that builds on IBC for interactive imitation learning from multiple heterogeneous human supervisors. A key insight in IIFL is a novel approach for uncertainty quantification in EBMs using Jeffreys divergence. While IIFL is more computationally expensive than explicit methods, results suggest that IIFL achieves a 2.8x higher success rate in simulation experiments and a 4.5x higher return on human effort in a physical block pushing task over (Explicit) IFL, IBC, and other baselines.

Weakly-supervised text classification trains a classifier using the label name of each target class as the only supervision, which largely reduces human annotation efforts. Most existing methods first use the label names as static keyword-based features to generate pseudo labels, which are then used for final classifier training. While reasonable, such a commonly adopted framework suffers from two limitations: (1) keywords can have different meanings in different contexts and some text may not have any keyword, so keyword matching can induce noisy and inadequate pseudo labels; (2) the errors made in the pseudo label generation stage will directly propagate to the classifier training stage without a chance of being corrected. In this paper, we propose a new method, PIEClass, consisting of two modules: (1) a pseudo label acquisition module that uses zero-shot prompting of pre-trained language models (PLM) to get pseudo labels based on contextualized text understanding beyond static keyword matching, and (2) a noise-robust iterative ensemble training module that iteratively trains classifiers and updates pseudo labels by utilizing two PLM fine-tuning methods that regularize each other. Extensive experiments show that PIEClass achieves overall better performance than existing strong baselines on seven benchmark datasets and even achieves similar performance to fully-supervised classifiers on sentiment classification tasks.