Recent advances in language modeling consist in pretraining highly parameterized neural networks on extremely large web-mined text corpora. Training and inference with such models can be costly in practice, which incentivizes the use of smaller counterparts. However, it has been observed that smaller models can suffer from saturation, characterized as a drop in performance at some advanced point in training followed by a plateau. In this paper, we find that such saturation can be explained by a mismatch between the hidden dimension of smaller models and the high rank of the target contextual probability distribution. This mismatch affects the performance of the linear prediction head used in such models through the well-known softmax bottleneck phenomenon. We measure the effect of the softmax bottleneck in various settings and find that models based on less than 1000 hidden dimensions tend to adopt degenerate latent representations in late pretraining, which leads to reduced evaluation performance.

Continual learning (CL) is crucial for evaluating adaptability in learning solutions to retain knowledge. Our research addresses the challenge of catastrophic forgetting, where models lose proficiency in previously learned tasks as they acquire new ones. While numerous solutions have been proposed, existing experimental setups often rely on idealized class-incremental learning scenarios. We introduce Realistic Continual Learning (RealCL), a novel CL paradigm where class distributions across tasks are random, departing from structured setups. We also present CLARE (Continual Learning Approach with pRE-trained models for RealCL scenarios), a pre-trained model-based solution designed to integrate new knowledge while preserving past learning. Our contributions include pioneering RealCL as a generalization of traditional CL setups, proposing CLARE as an adaptable approach for RealCL tasks, and conducting extensive experiments demonstrating its effectiveness across various RealCL scenarios. Notably, CLARE outperforms existing models on RealCL benchmarks, highlighting its versatility and robustness in unpredictable learning environments.

We present VoiceShop, a novel speech-to-speech framework that can modify multiple attributes of speech, such as age, gender, accent, and speech style, in a single forward pass while preserving the input speaker's timbre. Previous works have been constrained to specialized models that can only edit these attributes individually and suffer from the following pitfalls: the magnitude of the conversion effect is weak, there is no zero-shot capability for out-of-distribution speakers, or the synthesized outputs exhibit undesirable timbre leakage. Our work proposes solutions for each of these issues in a simple modular framework based on a conditional diffusion backbone model with optional normalizing flow-based and sequence-to-sequence speaker attribute-editing modules, whose components can be combined or removed during inference to meet a wide array of tasks without additional model finetuning. Audio samples are available at \url{https://voiceshopai.github.io}.

Goal Recognition is the task by which an observer aims to discern the goals that correspond to plans that comply with the perceived behavior of subject agents given as a sequence of observations. Research on Goal Recognition as Planning encompasses reasoning about the model of a planning task, the observations, and the goals using planning techniques, resulting in very efficient recognition approaches. In this article, we design novel recognition approaches that rely on the Operator-Counting framework, proposing new constraints, and analyze their constraints' properties both theoretically and empirically. The Operator-Counting framework is a technique that efficiently computes heuristic estimates of cost-to-goal using Integer/Linear Programming (IP/LP). In the realm of theory, we prove that the new constraints provide lower bounds on the cost of plans that comply with observations. We also provide an extensive empirical evaluation to assess how the new constraints improve the quality of the solution, and we found that they are especially informed in deciding which goals are unlikely to be part of the solution. Our novel recognition approaches have two pivotal advantages: first, they employ new IP/LP constraints for efficiently recognizing goals; second, we show how the new IP/LP constraints can improve the recognition of goals under both partial and noisy observability.

Practical applications of artificial intelligence increasingly often have to deal with the streaming properties of real data, which, considering the time factor, are subject to phenomena such as periodicity and more or less chaotic degeneration - resulting directly in the concept drifts. The modern concept drift detectors almost always assume immediate access to labels, which due to their cost, limited availability and possible delay has been shown to be unrealistic. This work proposes an unsupervised Parallel Activations Drift Detector, utilizing the outputs of an untrained neural network, presenting its key design elements, intuitions about processing properties, and a pool of computer experiments demonstrating its competitiveness with state-of-the-art methods.

Stance detection is an important task for many applications that analyse or support online political discussions. Common approaches include fine-tuning transformer based models. However, these models require a large amount of labelled data, which might not be available. In this work, we present two different ways to leverage LLM-generated synthetic data to train and improve stance detection agents for online political discussions: first, we show that augmenting a small fine-tuning dataset with synthetic data can improve the performance of the stance detection model. Second, we propose a new active learning method called SQBC based on the "Query-by-Comittee" approach. The key idea is to use LLM-generated synthetic data as an oracle to identify the most informative unlabelled samples, that are selected for manual labelling. Comprehensive experiments show that both ideas can improve the stance detection performance. Curiously, we observed that fine-tuning on actively selected samples can exceed the performance of using the full dataset.

Multilingual pretraining has been a successful solution to the challenges posed by the lack of resources for languages. These models can transfer knowledge to target languages with minimal or no examples. Recent research suggests that monolingual models also have a similar capability, but the mechanisms behind this transfer remain unclear. Some studies have explored factors like language contamination and syntactic similarity. An emerging line of research suggests that the representations learned by language models contain two components: a language-specific and a language-agnostic component. The latter is responsible for transferring a more universal knowledge. However, there is a lack of comprehensive exploration of these properties across diverse target languages. To investigate this hypothesis, we conducted an experiment inspired by the work on the Scaling Laws for Transfer. We measured the amount of data transferred from a source language to a target language and found that models initialized from diverse languages perform similarly to a target language in a cross-lingual setting. This was surprising because the amount of data transferred to 10 diverse target languages, such as Spanish, Korean, and Finnish, was quite similar. We also found evidence that this transfer is not related to language contamination or language proximity, which strengthens the hypothesis that the model also relies on language-agnostic knowledge. Our experiments have opened up new possibilities for measuring how much data represents the language-agnostic representations learned during pretraining.

Supply chain networks in enterprises are typically composed of complex topological graphs involving various types of nodes and edges, accommodating numerous products with considerable demand and supply variability. However, as supply chain networks expand in size and complexity, traditional supply chain planning methods (e.g., those found in heuristic rule-based and operations research-based systems) tend to become locally optimal or lack computational scalability, resulting in substantial imbalances between supply and demand across nodes in the network. This paper introduces a novel Generative AI technique, which we call Generative Probabilistic Planning (GPP). GPP generates dynamic supply action plans that are globally optimized across all network nodes over the time horizon for changing objectives like maximizing profits or service levels, factoring in time-varying probabilistic demand, lead time, and production conditions. GPP leverages attention-based graph neural networks (GNN), offline deep reinforcement learning (Offline RL), and policy simulations to train generative policy models and create optimal plans through probabilistic simulations, effectively accounting for various uncertainties. Our experiments using historical data from a global consumer goods company with complex supply chain networks demonstrate that GPP accomplishes objective-adaptable, probabilistically resilient, and dynamic planning for supply chain networks, leading to significant improvements in performance and profitability for enterprises. Our work plays a pivotal role in shaping the trajectory of AI adoption within the supply chain domain.

This study examines the effects of de-globalization trends on international trade networks and their role in improving forecasts for economic growth. Using section-level trade data from nearly 200 countries from 2010 to 2022, we identify significant shifts in the network topology driven by rising trade policy uncertainty. Our analysis highlights key global players through centrality rankings, with the United States, China, and Germany maintaining consistent dominance. Using a horse race of supervised regressors, we find that network topology descriptors evaluated from section-specific trade networks substantially enhance the quality of a country's GDP growth forecast. We also find that non-linear models, such as Random Forest, XGBoost, and LightGBM, outperform traditional linear models used in the economics literature. Using SHAP values to interpret these non-linear model's predictions, we find that about half of most important features originate from the network descriptors, underscoring their vital role in refining forecasts. Moreover, this study emphasizes the significance of recent economic performance, population growth, and the primary sector's influence in shaping economic growth predictions, offering novel insights into the intricacies of economic growth forecasting.

In most applications of utilizing neural networks for mathematical optimization, a dedicated model is trained for each specific optimization objective. However, in many scenarios, several distinct yet correlated objectives or tasks often need to be optimized on the same set of problem inputs. Instead of independently training a different neural network for each problem separately, it would be more efficient to exploit the correlations between these objectives and to train multiple neural network models with shared model parameters and feature representations. To achieve this, this paper first establishes the concept of common information: the shared knowledge required for solving the correlated tasks, then proposes a novel approach for model training by adding into the model an additional reconstruction stage associated with a new reconstruction loss. This loss is for reconstructing the common information starting from a selected hidden layer in the model. The proposed approach encourages the learned features to be general and transferable, and therefore can be readily used for efficient transfer learning. For numerical simulations, three applications are studied: transfer learning on classifying MNIST handwritten digits, the device-to-device wireless network power allocation, and the multiple-input-single-output network downlink beamforming and localization. Simulation results suggest that the proposed approach is highly efficient in data and model complexity, is resilient to over-fitting, and has competitive performances.