Karthik Raghunathan

Karthik Charan Raghunathan, Christian Metzner, Laura Kriener et al.

In a continual learning setting, we require a model to be plastic enough to learn a new task and stable enough to not disturb previously learned capabilities. We argue that this dilemma has an architectural root. A finite network has limited representational and plastic resources, yet the required capacity depends on properties of the future task stream that are unknown: how many tasks will be encountered, and how much they overlap in feature space. Regularization-based methods preserve past knowledge within fixed-capacity architectures and therefore implicitly rely on an oracle architecture sized for this unknown future. When tasks are only weakly related, fixed architectures progressively run out of plastic resources; when tasks are few or strongly overlapping, models are often over-provisioned. Inspired by neurogenesis in biology, we propose NORACL to address the stability-plasticity dilemma by tackling the oracle architecture problem through neuronal growth. Starting from a compact network, NORACL grows only when needed by monitoring two complementary signals for representational and plasticity saturation. We evaluate NORACL against oracle-sized static baselines across varying task counts and geometries. Across all settings, NORACL achieves final average accuracies that are better than or on par with oracle-provisioned static baselines while using fewer parameters. Additionally, NORACL yields architectures with interpretable growth, i.e. dissimilar tasks predominantly expand feature-extraction layers, whereas tasks which rely on common features shift growth toward later feature-combination layers. Our analysis further explains why fixed-capacity networks lose plasticity as tasks accumulate, whereas NORACL creates fresh capacity for new tasks through growth. Together, these results show that adaptive neurogenesis pushes the stability-plasticity Pareto frontier of continual learning.

Sai Muralidhar Jayanthi, Varsha Embar, Karthik Raghunathan

The wide applicability of pretrained transformer models (PTMs) for natural language tasks is well demonstrated, but their ability to comprehend short phrases of text is less explored. To this end, we evaluate different PTMs from the lens of unsupervised Entity Linking in task-oriented dialog across 5 characteristics -- syntactic, semantic, short-forms, numeric and phonetic. Our results demonstrate that several of the PTMs produce sub-par results when compared to traditional techniques, albeit competitive to other neural baselines. We find that some of their shortcomings can be addressed by using PTMs fine-tuned for text-similarity tasks, which illustrate an improved ability in comprehending semantic and syntactic correspondences, as well as some improvements for short-forms, numeric and phonetic variations in entity mentions. We perform qualitative analysis to understand nuances in their predictions and discuss scope for further improvements. Code can be found at https://github.com/murali1996/el_tod

Konstantin Kolokolov, Pavel Pekichev, Karthik Raghunathan

We introduce a novel neural network module that adeptly handles recursive data flow in neural network architectures. At its core, this module employs a self-consistent approach where a set of recursive equations is solved iteratively, halting when the difference between two consecutive iterations falls below a defined threshold. Leveraging this mechanism, we construct a new neural network architecture, an extension of the conformer transducer, which enriches automatic speech recognition systems with a stream of contextual information. Our method notably improves the accuracy of recognizing rare words without adversely affecting the word error rate for common vocabulary. We investigate the improvement in accuracy for these uncommon words using our novel model, both independently and in conjunction with shallow fusion with a context language model. Our findings reveal that the combination of both approaches can improve the accuracy of detecting rare words by as much as 4.5 times. Our proposed self-consistent recursive methodology is versatile and adaptable, compatible with many recently developed encoders, and has the potential to drive model improvements in speech recognition and beyond.