James O’ Neill

James O' Neill, Sourav Dutta, Haytham Assem

While various avenues of research have been explored for iterative pruning, little is known what effect pruning has on zero-shot test performance and its potential implications on the choice of pruning criteria. This pruning setup is particularly important for cross-lingual models that implicitly learn alignment between language representations during pretraining, which if distorted via pruning, not only leads to poorer performance on language data used for retraining but also on zero-shot languages that are evaluated. In this work, we show that there is a clear performance discrepancy in magnitude-based pruning when comparing standard supervised learning to the zero-shot setting. From this finding, we propose two weight regularizers that aim to maximize the alignment between units of pruned and unpruned networks to mitigate alignment distortion in pruned cross-lingual models and perform well for both non zero-shot and zero-shot settings. We provide experimental results on cross-lingual tasks for the zero-shot setting using XLM-RoBERTa$_{\mathrm{Base}}$, where we also find that pruning has varying degrees of representational degradation depending on the language corresponding to the zero-shot test set. This is also the first study that focuses on cross-lingual language model compression.

James O' Neill, Sourav Dutta

We investigate the effects of post-training quantization and quantization-aware training on the generalization of Transformer language models. We present a new method called self-distilled quantization (SDQ) that minimizes accumulative quantization errors and outperforms baselines. We apply SDQ to multilingual models XLM-R-Base and InfoXLM-Base and demonstrate that both models can be reduced from 32-bit floating point weights to 8-bit integer weights while maintaining a high level of performance on the XGLUE benchmark. Our results also highlight the challenges of quantizing multilingual models, which must generalize to languages they were not fine-tuned on.

James O' Neill, Sourav Dutta

Fine-tuning pretrained self-supervised language models is widely adopted for transfer learning to downstream tasks. Fine-tuning can be achieved by freezing gradients of the pretrained network and only updating gradients of a newly added classification layer, or by performing gradient updates on all parameters. Gradual unfreezing makes a trade-off between the two by gradually unfreezing gradients of whole layers during training. This has been an effective strategy to trade-off between storage and training speed with generalization performance. However, it is not clear whether gradually unfreezing layers throughout training is optimal, compared to sparse variants of gradual unfreezing which may improve fine-tuning performance. In this paper, we propose to stochastically mask gradients to regularize pretrained language models for improving overall fine-tuned performance. We introduce GradDrop and variants thereof, a class of gradient sparsification methods that mask gradients during the backward pass, acting as gradient noise. GradDrop is sparse and stochastic unlike gradual freezing. Extensive experiments on the multilingual XGLUE benchmark with XLMR-Large show that GradDrop is competitive against methods that use additional translated data for intermediate pretraining and outperforms standard fine-tuning and gradual unfreezing. A post-analysis shows how GradDrop improves performance with languages it was not trained on, such as under-resourced languages.

James O' Neill, Santhosh Subramanian, Eric Lin et al.

The trend towards large language models (LLMs) for guardrailing against undesired behaviors is increasing and has shown promise for censoring user inputs. However, increased latency, memory consumption, hosting expenses and non-structured outputs can make their use prohibitive. In this work, we show that task-specific data generation can lead to fine-tuned classifiers that significantly outperform current state of the art (SoTA) while being orders of magnitude smaller. Secondly, we show that using a single model, \texttt{MultiTaskGuard}, that is pretrained on a large synthetically generated dataset with unique task instructions further improves generalization. Thirdly, our most performant models, \texttt{UniGuard}, are found using our proposed search-based model merging approach that finds an optimal set of parameters to combine single-policy models and multi-policy guardrail models. % On 7 public datasets and 4 guardrail benchmarks we created, our efficient guardrail classifiers improve over the best performing SoTA publicly available LLMs and 3$^{\text{rd}}$ party guardrail APIs in detecting unsafe and safe behaviors by an average F1 score improvement of \textbf{29.92} points over Aegis-LlamaGuard and \textbf{21.62} over \texttt{gpt-4o}, respectively. Lastly, our guardrail synthetic data generation process that uses custom task-specific guardrail poli

James O' Neill, Sourav Dutta, Haytham Assem

Pruning aims to reduce the number of parameters while maintaining performance close to the original network. This work proposes a novel \emph{self-distillation} based pruning strategy, whereby the representational similarity between the pruned and unpruned versions of the same network is maximized. Unlike previous approaches that treat distillation and pruning separately, we use distillation to inform the pruning criteria, without requiring a separate student network as in knowledge distillation. We show that the proposed {\em cross-correlation objective for self-distilled pruning} implicitly encourages sparse solutions, naturally complementing magnitude-based pruning criteria. Experiments on the GLUE and XGLUE benchmarks show that self-distilled pruning increases mono- and cross-lingual language model performance. Self-distilled pruned models also outperform smaller Transformers with an equal number of parameters and are competitive against (6 times) larger distilled networks. We also observe that self-distillation (1) maximizes class separability, (2) increases the signal-to-noise ratio, and (3) converges faster after pruning steps, providing further insights into why self-distilled pruning improves generalization.

James O' Neill

Overparameterized networks trained to convergence have shown impressive performance in domains such as computer vision and natural language processing. Pushing state of the art on salient tasks within these domains corresponds to these models becoming larger and more difficult for machine learning practitioners to use given the increasing memory and storage requirements, not to mention the larger carbon footprint. Thus, in recent years there has been a resurgence in model compression techniques, particularly for deep convolutional neural networks and self-attention based networks such as the Transformer. Hence, this paper provides a timely overview of both old and current compression techniques for deep neural networks, including pruning, quantization, tensor decomposition, knowledge distillation and combinations thereof. We assume a basic familiarity with deep learning architectures\footnote{For an introduction to deep learning, see ~\citet{goodfellow2016deep}}, namely, Recurrent Neural Networks~\citep[(RNNs)][]{rumelhart1985learning,hochreiter1997long}, Convolutional Neural Networks~\citep{fukushima1980neocognitron}~\footnote{For an up to date overview see~\citet{khan2019survey}} and Self-Attention based networks~\citep{vaswani2017attention}\footnote{For a general overview of self-attention networks, see ~\citet{chaudhari2019attentive}.},\footnote{For more detail and their use in natural language processing, see~\citet{hu2019introductory}}. Most of the papers discussed are proposed in the context of at least one of these DNN architectures.

James O' Neill, Danushka Bollegala

The task of multi-step ahead prediction in language models is challenging considering the discrepancy between training and testing. At test time, a language model is required to make predictions given past predictions as input, instead of the past targets that are provided during training. This difference, known as exposure bias, can lead to the compounding of errors along a generated sequence at test time. In order to improve generalization in neural language models and address compounding errors, we propose a curriculum learning based method that gradually changes an initially deterministic teacher policy to a gradually more stochastic policy, which we refer to as \textit{Nearest-Neighbor Replacement Sampling}. A chosen input at a given timestep is replaced with a sampled nearest neighbor of the past target with a truncated probability proportional to the cosine similarity between the original word and its top $k$ most similar words. This allows the teacher to explore alternatives when the teacher provides a sub-optimal policy or when the initial policy is difficult for the learner to model. The proposed strategy is straightforward, online and requires little additional memory requirements. We report our main findings on two language modelling benchmarks and find that the proposed approach performs particularly well when used in conjunction with scheduled sampling, that too attempts to mitigate compounding errors in language models.

James O' Neill, Danushka Bollegala

Word embeddings have been shown to benefit from ensambling several word embedding sources, often carried out using straightforward mathematical operations over the set of word vectors. More recently, self-supervised learning has been used to find a lower-dimensional representation, similar in size to the individual word embeddings within the ensemble. However, these methods do not use the available manually labeled datasets that are often used solely for the purpose of evaluation. We propose to reconstruct an ensemble of word embeddings as an auxiliary task that regularises a main task while both tasks share the learned meta-embedding layer. We carry out intrinsic evaluation (6 word similarity datasets and 3 analogy datasets) and extrinsic evaluation (4 downstream tasks). For intrinsic task evaluation, supervision comes from various labeled word similarity datasets. Our experimental results show that the performance is improved for all word similarity datasets when compared to self-supervised learning methods with a mean increase of $11.33$ in Spearman correlation. Specifically, the proposed method shows the best performance in 4 out of 6 of word similarity datasets when using a cosine reconstruction loss and Brier's word similarity loss. Moreover, improvements are also made when performing word meta-embedding reconstruction in sequence tagging and sentence meta-embedding for sentence classification.

James O' Neill, Danushka Bollegala

Ensembling word embeddings to improve distributed word representations has shown good success for natural language processing tasks in recent years. These approaches either carry out straightforward mathematical operations over a set of vectors or use unsupervised learning to find a lower-dimensional representation. This work compares meta-embeddings trained for different losses, namely loss functions that account for angular distance between the reconstructed embedding and the target and those that account normalized distances based on the vector length. We argue that meta-embeddings are better to treat the ensemble set equally in unsupervised learning as the respective quality of each embedding is unknown for upstream tasks prior to meta-embedding. We show that normalization methods that account for this such as cosine and KL-divergence objectives outperform meta-embedding trained on standard $\ell_1$ and $\ell_2$ loss on \textit{defacto} word similarity and relatedness datasets and find it outperforms existing meta-learning strategies.

James O' Neill

Capsule Networks have shown encouraging results on \textit{defacto} benchmark computer vision datasets such as MNIST, CIFAR and smallNORB. Although, they are yet to be tested on tasks where (1) the entities detected inherently have more complex internal representations and (2) there are very few instances per class to learn from and (3) where point-wise classification is not suitable. Hence, this paper carries out experiments on face verification in both controlled and uncontrolled settings that together address these points. In doing so we introduce \textit{Siamese Capsule Networks}, a new variant that can be used for pairwise learning tasks. The model is trained using contrastive loss with $\ell_2$-normalized capsule encoded pose features. We find that \textit{Siamese Capsule Networks} perform well against strong baselines on both pairwise learning datasets, yielding best results in the few-shot learning setting where image pairs in the test set contain unseen subjects.

James O' Neill, Danushka Bollegala

Many tasks in natural language understanding require learning relationships between two sequences for various tasks such as natural language inference, paraphrasing and entailment. These aforementioned tasks are similar in nature, yet they are often modeled individually. Knowledge transfer can be effective for closely related tasks. However, transferring all knowledge, some of which irrelevant for a target task, can lead to sub-optimal results due to \textit{negative} transfer. Hence, this paper focuses on the transferability of both instances and parameters across natural language understanding tasks by proposing an ensemble-based transfer learning method. \newline The primary contribution of this paper is the combination of both \textit{Dropout} and \textit{Bagging} for improved transferability in neural networks, referred to as \textit{Dropping} herein. We present a straightforward yet novel approach for incorporating source \textit{Dropping} Networks to a target task for few-shot learning that mitigates \textit{negative} transfer. This is achieved by using a decaying parameter chosen according to the slope changes of a smoothed spline error curve at sub-intervals during training. We compare the proposed approach against hard parameter sharing and soft parameter sharing transfer methods in the few-shot learning case. We also compare against models that are fully trained on the target task in the standard supervised learning setup. The aforementioned adjustment leads to improved transfer learning performance and comparable results to the current state of the art only using a fraction of the data from the target task.