Received 17 February 2025, accepted 5 March 2025, date of publication 13 March 2025, date of current version 21 March 2025. Continual Learning with Quasi-Newton Methods STEVEN VANDER EECKT and HUGO VAN HAMME (Senior, IEEE) Department Electrical Engineering ESAT-PSI, KU Leuven, B-3001 Leuven, Belgium Corresponding author: Steven Vander Eeeckt (e-mail: steven.vandereeckt@esat.kuleuven.be).ABSTRACT Catastrophic forgetting remains a major challenge when neural networks learn tasks sequentially. Elastic Weight Consolidation (EWC) attempts to address this problem by introducing a Bayesian-inspired regularization loss to preserve knowledge of previously learned tasks. However, EWC relies on a Laplace approximation where the Hessian is simplified to the diagonal of the Fisher information matrix, assuming uncorrelated model parameters. This overly simplistic assumption often leads to poor Hessian estimates, limiting its effectiveness. To overcome this limitation, we introduce Continual Learning with Sampled Quasi-Newton (CSQN), which leverages Quasi-Newton methods to compute more accurate Hessian approximations. Experimental results across four benchmarks demonstrate that CSQN consistently outperforms EWC and other state-of-the-art baselines, including rehearsal-based methods. CSQN reduces EWC's forgetting by 50% and improves its performance by 8% on average. Notably, CSQN achieves superior results on three out of four benchmarks, including the most challenging scenarios, highlighting its potential as a robust solution for continual learning.INDEX TERMS artificial neural networks, catastrophic forgetting, continual learning, quasi-Newton methods I. INTRODUCTION Since the 2010s, Artificial Neural Networks (ANNs) have been able to match or even surpass human performance on a wide variety of tasks. However, when presented with a set of tasks to be learned sequentially--a setting referred to as Continual Learning (CL)--ANNs suffer from catastrophic forgetting [1]. Unlike humans, ANNs struggle to retain previously learned knowledge when extending their knowledge. Naively adapting an ANN to a new task generally leads to a deterioration in the network's performance on previous tasks. Many CL methods have been proposed to alleviate catastrophic forgetting. One of the most well-known is Elastic Weight Consolidation (EWC) [2], which approaches CL from a Bayesian perspective. After training on a task, EWC uses Laplace approximation [3] to estimate a posterior distribution over the model parameters for that task. When training on the next task, this posterior is used via a regularization loss to prevent the model from catastrophically forgetting the previous task. To estimate the Hessian, which is needed in the Laplace approximation to measure the (un)certainty of the model parameters, EWC uses the Fisher Information Matrix (FIM). Furthermore, to simplify the computation, EWC assumes that the FIM is approximately diagonal.

While extensively explored in text-based tasks, Named Entity Recognition (NER) remains largely neglected in spoken language understanding. Existing resources are limited to a single, English-only dataset. This paper addresses this gap by introducing MSNER, a freely available, multilingual speech corpus annotated with named entities. It provides annotations to the VoxPopuli dataset in four languages (Dutch, French, German, and Spanish). We have also releasing an efficient annotation tool that leverages automatic pre-annotations for faster manual refinement. This results in 590 and 15 hours of silver-annotated speech for training and validation, alongside a 17-hour, manually-annotated evaluation set. We further provide an analysis comparing silver and gold annotations. Finally, we present baseline NER models to stimulate further research on this newly available dataset.

Most spoken language understanding systems use a pipeline approach composed of an automatic speech recognition interface and a natural language understanding module. This approach forces hard decisions when converting continuous inputs into discrete language symbols. Instead, we propose a representation model to encode speech in rich bidirectional encodings that can be used for downstream tasks such as intent prediction. The approach uses a masked language modelling objective to learn the representations, and thus benefits from both the left and right contexts. We show that the performance of the resulting encodings before fine-tuning is better than comparable models on multiple datasets, and that fine-tuning the top layers of the representation model improves the current state of the art on the Fluent Speech Command dataset, also in a low-data regime, when a limited amount of labelled data is used for training. Furthermore, we propose class attention as a spoken language understanding module, efficient both in terms of speed and number of parameters. Class attention can be used to visually explain the predictions of our model, which goes a long way in understanding how the model makes predictions. We perform experiments in English and in Dutch.

Adapting a trained Automatic Speech Recognition (ASR) model to new tasks results in catastrophic forgetting of old tasks, limiting the model's ability to learn continually and to be extended to new speakers, dialects, languages, etc. Focusing on End-to-End ASR, in this paper, we propose a simple yet effective method to overcome catastrophic forgetting: weight averaging. By simply taking the average of the previous and the adapted model, our method achieves high performance on both the old and new tasks. It can be further improved by introducing a knowledge distillation loss during the adaptation. We illustrate the effectiveness of our method on both monolingual and multilingual ASR. In both cases, our method strongly outperforms all baselines, even in its simplest form.

We explore the benefits that multitask learning offer to speech processing as we train models on dual objectives with automatic speech recognition and intent classification or sentiment classification. Our models, although being of modest size, show improvements over models trained end-to-end on intent classification. We compare different settings to find the optimal disposition of each task module compared to one another. Finally, we study the performance of the models in low-resource scenario by training the models with as few as one example per class. We show that multitask learning in these scenarios compete with a baseline model trained on text features and performs considerably better than a pipeline model. On sentiment classification, we match the performance of an end-to-end model with ten times as many parameters. We consider 4 tasks and 4 datasets in Dutch and English.

Adapting Automatic Speech Recognition (ASR) models to new domains leads to a deterioration of performance on the original domain(s), a phenomenon called Catastrophic Forgetting (CF). Even monolingual ASR models cannot be extended to new accents, dialects, topics, etc. without suffering from CF, making them unable to be continually enhanced without storing all past data. Fortunately, Continual Learning (CL) methods, which aim to enable continual adaptation while overcoming CF, can be used. In this paper, we implement an extensive number of CL methods for End-to-End ASR and test and compare their ability to extend a monolingual Hybrid CTC-Transformer model across four new tasks. We find that the best performing CL method closes the gap between the fine-tuned model (lower bound) and the model trained jointly on all tasks (upper bound) by more than 40%, while requiring access to only 0.6% of the original data.

With deep learning approaches becoming state-of-the-art in many speech (as well as non-speech) related machine learning tasks, efforts are being taken to delve into the neural networks which are often considered as a black box. In this paper it is analyzed how recurrent neural network (RNNs) cope with temporal dependencies by determining the relevant memory time span in a long short-term memory (LSTM) cell. This is done by leaking the state variable with a controlled lifetime and evaluating the task performance. This technique can be used for any task to estimate the time span the LSTM exploits in that specific scenario. The focus in this paper is on the task of separating speakers from overlapping speech. We discern two effects: A long term effect, probably due to speaker characterization and a short term effect, probably exploiting phone-size formant tracks.

Research in deep learning for multi-speaker source separation has received a boost in the last years. However, most studies are restricted to mixtures of a specific number of speakers, called a specific scenario. While some works included experiments for different scenarios, research towards combining data of different scenarios or creating a single model for multiple scenarios have been very rare. In this work it is shown that data of a specific scenario is relevant for solving another scenario. Furthermore, it is concluded that a single model, trained on different scenarios is capable of matching performance of scenario specific models.