Transformer-based models have achieved dominant performance in numerous NLP tasks. Despite their remarkable successes, pre-trained transformers such as BERT suffer from a computationally expensive self-attention mechanism that interacts with all tokens, including the ones unfavorable to classification performance. To overcome these challenges, we propose integrating two strategies: token pruning and token combining. Token pruning eliminates less important tokens in the attention mechanism’s key and value as they pass through the layers. Additionally, we adopt fuzzy logic to handle uncertainty and alleviate potential mispruning risks arising from an imbalanced distribution of each token’s importance. Token combining, on the other hand, condenses input sequences into smaller sizes in order to further compress the model. By integrating these two approaches, we not only improve the model’s performance but also reduce its computational demands. Experiments with various datasets demonstrate superior performance compared to baseline models, especially with the best improvement over the existing BERT model, achieving +5%p in accuracy and +5.6%p in F1 score. Additionally, memory cost is reduced to 0.61x, and a speedup of 1.64x is achieved.

Focus on the Core: Efficient Attention via Pruned Token Compression for Document Classification

Decision tree models form the basis for many machine learning approaches and have several useful properties; however, single-tree applications are limited by the balance between predictive accuracy and generalization. We introduce a new method for increasing the generalized predictive performance of pre-trained decision tree models. Using this method, the structure of the tree is leveraged to derive a set of decision threshold-based weights controlling the relative contribution of prediction values. The final prediction is then the result of an aggregation of these weighted values. This “soft trees” approach is applied to a set of benchmark datasets and is shown to increase r2 performance over standard decision trees. We further explore the introduction of missing or categorical input features, utility in probabilistic classification, and applications to tasks in resource-limited hardware environments.

Improving decision tree generalization using threshold smoothing

The diverse range of naval vessels, ranging from aircraft carriers to unmanned surface vessels, and infrastructure, such as smart naval mines or buoys, presents a challenge for traditional machine learning (ML) techniques in naval defense. Traditional ML, which centralizes network data, struggles with the vast distances and unstable or insecure communication links typical in naval defense scenarios. 

Federated learning (FL) offers a solution by keeping the data and ML model training local to the naval vessels and only performing occasional ML model combinations, or "global aggregations" at a central server, or "mothership" for naval applications. FL does address some of the heterogeneity inherent to the naval setting, for example, it minimizes the communication burden during deployments through sporadic global ML model aggregations. "Statistical heterogeneity" here refers to data-related differences among network elements, such as various dataset sizes. "Systems heterogeneity" relates to physical differences between naval objects, such as varying reliability of communication links. 

However, existing literature in FL typically assumes the presence of fully labeled datasets across all naval elements, an unrealistic scenario in practical naval deployments where a diverse mixture of both \textbf{labeled} and \textbf{unlabeled} data is common. Current FL-based methodologies tend to produce a uniform global ML model for all naval elements, neglecting the information contained with the unlabeled data, potentially leading to suboptimal global ML models for diverse naval resources. 

To address this, we introduce a decentralized federated domain adaptation methodology which leverages both labeled and unlabeled data to identify the underlying statistical divergences among naval vessels. In doing so, we are able to optimize the transfer of ML models from naval elements with high quality labeled data (called sources) to naval resources with low quality or unlabeled data (called targets), thereby personalizing ML models to better fit the real-time needs of different naval vessels in a deployment. Our methodology, called Source-Target Determination and Link Formation (ST-LF), optimizes both (i) classification of naval resources into sources and targets and (ii) source-target link formation, in a manner that considers the trade-off between ML model accuracy and communication energy expenditures. To obtain a concrete objective function, we derive a measurable generalization error bound that accounts for estimates of source-target hypothesis deviations and divergences between data distributions. The resulting optimization problem is a mixed-integer signomial program, a class of NP-hard problems, for which we develop an algorithm based on successive convex approximations to solve it tractably. Subsequent numerical evaluations of ST-LF demonstrate that it improves classification accuracy and energy efficiency over state-of-the-art baselines.

Envisioning Multi-Source to Multi-Target Federated Domain Adaptation for Naval Defense Scenarios

The ever-growing size of deep learning models necessitates more extensive resources, making them unsuitable for tasks requiring lower latency and less computational power. This is particularly true for edge devices. Multiple compression techniques have been developed to bring the capabilities of these larger models to a smaller form factor, one of the most successful of which has been knowledge distillation. Knowledge distillation is employed to transfer knowledge from a large "teacher" model to a smaller "student" model. This technique facilitates learning a compressed representation of data, maintaining the performance benefits of large models while reducing computational and memory demands. It involves training the student model on both hard labels and soft labels provided by the teacher model, leading to better generalization and informed predictions.

Our work increases the effectiveness of knowledge distillation by augmenting the labels of both the teacher and student models with a vector quantized vector. Vector quantization is a mapping of high-dimensional data to a lower-dimensional codebook, effectively reducing data dimensionality. This technique has traditionally been applied to speech and image compression particularly in Variational Autoencoders (VQ-VAEs), and has proven effective at markedly increasing the effectiveness of student models to learn from teacher model representations.

The work presented here seeks to enhance the classification capabilities of the student model, carefully maintaining its compact size to balance efficiency and compactness. This approach improves retained knowledge distillation performance, while conferring model interpretability benefits. We show that while also increasing student model classification performance, the inclusion of the VQ layer also provides a measure for detecting classification quality and detect data which is out of distribution. This approach promises to increase the capabilities of edge deployed models both with respect to their mission performance as well as out of distribution detection.


Improving Knowledge Distillation with Discrete Feature Vector Representation

SysML (Systems Modeling Language) is a widely-used formalism and international standard for designing and communicating complex system designs, and is especially relied upon in the development of safety-critical autonomous systems. SysML v2 is a major new version of SysML currently undergoing finalization, scheduled to be released mid-2024. Among many key enhancements, SysML v2 introduces a formal semantics for specifying the precise meaning of SysML models, which makes it possible to apply formal verification and automated reasoning to analyze safety and correctness properties of SysML designs automatically.

At Imandra, we are applying our Imandra Automated Reasoning System to the design and analysis of SysML v2 models. Key to this work is a formal semantics for SysML v2 models expressed in Imandra's logic, and corresponding verification automation for answering deep questions about the possible behaviors of SysML models. In addition to verification, we can then leverage Imandra's "proof-carrying prose" integration with LLMs to obtain conversational assistants (chatbots) which can rigorously answer complex questions about SysML v2 designs, backing every answer up with an independently verifiable logical audit trail produced by Imandra reasoning about the SysML v2 model.

During this talk, we will demonstrate how to use Automated Reasoning with SysML v2 models to:
- Mathematically prove properties of system correctness.
- Add a hallucination-free natural-language interface (chatbot) to empower more stakeholders to gain insights and leverage the SysML v2 model.

Imandra + SysML: Automated Reasoning for SysML (Systems Modeling Language) v2

Retrieval-enhanced methods have become a primary approach in fact verification (FV); it requires reasoning over multiple retrieved pieces of evidence to verify the integrity of a claim. To retrieve evidence, existing work often employs off-the-shelf retrieval models whose design is based on the probability ranking principle. We argue that, rather than relevance, for FV we need to focus on the utility that a claim verifier derives from the retrieved evidence. We introduce the $\textbf{feedback-based evidence retriever} (FER)$ that optimizes the evidence retrieval process by incorporating feedback from the claim verifier. As a feedback signal we use the divergence in utility between how effectively the verifier utilizes the retrieved evidence and the ground-truth evidence to produce the final claim label. Empirical studies demonstrate the superiority of FER over prevailing baselines.

From Relevance to Utility: Evidence Retrieval with Feedback for Fact Verification

End-to-end (E2E) automatic speech recognition (ASR) models are becoming increasingly popular in commercial applications, such as virtual assistants, closed captioning, and dictation systems. The accuracy of the ASR is crucial to their success. However, E2E models still struggle to recognize out-of-domain words such as proper nouns and domain-specific terms. In this paper we introduce AdaBERT-CTC, a domain adaptation technique that relies solely on textual data. Our method allows for text-only adaptation by fine-tuning a pre-trained self-supervised text encoder model. Additionally, we show that our method can be made parameter-efficient by adding bottleneck adapters to the pre-trained model. This allows for adaptation with less than a 5% increase in parameters and minimal computational overhead during inference. We demonstrate that our approach outperforms the base BERT-CTC model by up to 14\% relative word error rate improvement on several out-of-domain, publicly available datasets.

AdaBERT-CTC: Leveraging BERT-CTC for Text-Only Domain Adaptation in ASR

Zero-shot retrieval tasks such as the BEIR benchmark reveal out-of-domain generalization as a key weakness of high-performance dense retrievers. As a solution, domain adaptation for dense retrievers has been actively studied. A notable approach is synthesizing domain-specific data, by generating pseudo queries (PQ), for fine-tuning with domain-specific relevance between PQ and documents. Our contribution is showing that key biases can cause sampled PQ to be irrelevant, negatively contributing to generalization. We propose to preempt their generation, by dividing the generation into simpler subtasks, of generating relevance explanations and guiding the generation to avoid negative generalization. Experiment results show that our proposed approach is more robust to domain shifts, validated on challenging BEIR zero-shot retrieval tasks.

Relevance-assisted Generation for Robust Zero-shot Retrieval

Large Language Model (LLM) has gained popularity and achieved remarkable results in open-domain tasks, but its performance in real industrial domain-specific scenarios is average due to its lack of specific domain knowledge. This issue has attracted widespread attention, but there are few relevant benchmarks available. In this paper, we provide a benchmark Question Answering (QA) dataset named MSQA, centered around Microsoft products and IT technical problems encountered by customers. This dataset contains industry cloud-specific QA knowledge, an area not extensively covered in general LLMs, making it well-suited for evaluating methods aiming to enhance LLMs' domain-specific capabilities. In addition, we propose a new model interaction paradigm that can empower LLM to achieve better performance on domain-specific tasks where it is not proficient. Extensive experiments demonstrate that the approach following our method outperforms the commonly used LLM with retrieval methods. We make our source code and sample data available at: https://aka.ms/Microsoft_QA.

Empower Large Language Model to Perform Better on Industrial Domain-Specific Question Answering

Prompt-based learning has been an effective paradigm for large pretrained language models (LLM), enabling few-shot or even zero-shot learning. Black-box prompt search has received growing interest recently for its distinctive properties of gradient-free optimization, proven particularly useful and powerful for model-as-a-service usage. However, the discrete nature and the complexity of combinatorial optimization hinder the efficiency of modern black-box approaches. Despite extensive research on search algorithms, the crucial aspect of search space design and optimization has been largely overlooked. In this paper, we first conduct a sensitivity analysis by prompting LLM, revealing that only a small number of tokens exert a disproportionate amount of influence on LLM predictions. Leveraging this insight, we propose the Clustering and Pruning for Efficient Black-box Prompt Search (ClaPS), a simple black-box search method that first clusters and prunes the search space to focus exclusively on influential prompt tokens. By employing even simple search methods within the pruned search space, ClaPS achieves state-of-the-art performance across various tasks and LLMs, surpassing the performance of complex approaches while significantly reducing search costs. Our findings underscore the critical role of search space design and optimization in enhancing both the usefulness and the efficiency of black-box prompt-based learning.

Survival of the Most Influential Prompts: Efficient Black-Box Prompt Search via Clustering and Pruning

This paper reports on the experiments aimed to improve our understanding of the role of the amount of data required for training attention-based transformer language models. Specifically, we investigate the impact of reducing the immense amounts of required pre-training data through sampling strategies that identify and reduce high-frequency tokens as different studies have indicated that the existence of very high-frequency tokens in pre-training data might bias learning, causing undesired effects. In this light, we describe our sampling algorithm that iteratively assesses token frequencies and removes sentences that contain still high-frequency tokens, eventually delivering a balanced, linguistically correct dataset. We evaluate the results in terms of model perplexity and fine-tuning linguistic probing tasks, NLP downstream tasks as well as more semantic SuperGlue tasks. The results show that pre-training with the resulting balanced dataset allows reducing up to three times the pre-training data.

Frequency Balanced Datasets Lead to Better Language Models

Scaling dense PCFGs to thousands of nonterminals via low-rank parameterizations of the rule probability tensor has been shown to be beneficial for unsupervised parsing. However, PCFGs scaled this way still perform poorly as a language model, and even underperform similarly-sized HMMs. This work introduces $\emph{SimplePCFG}$, a simple PCFG formalism with independent left and right productions. Despite imposing a stronger independence assumption than the low-rank approach, we find that this formalism scales more effectively both as a language model and as an unsupervised parser. We further introduce $\emph{FlashInside}$, a hardware IO-aware implementation of the inside algorithm for efficiently scaling simple PCFGs. Through extensive experiments on multiple grammar induction benchmarks, we validate the effectiveness of simple PCFGs over low-rank baselines.

Underline digital video library

Focus on the Core: Efficient Attention via Pruned Token Compression for Document Classification

Improving decision tree generalization using threshold smoothing

Envisioning Multi-Source to Multi-Target Federated Domain Adaptation for Naval Defense Scenarios

Improving Knowledge Distillation with Discrete Feature Vector Representation

Imandra + SysML: Automated Reasoning for SysML (Systems Modeling Language) v2

From Relevance to Utility: Evidence Retrieval with Feedback for Fact Verification

AdaBERT-CTC: Leveraging BERT-CTC for Text-Only Domain Adaptation in ASR

Relevance-assisted Generation for Robust Zero-shot Retrieval

Empower Large Language Model to Perform Better on Industrial Domain-Specific Question Answering

Survival of the Most Influential Prompts: Efficient Black-Box Prompt Search via Clustering and Pruning

Frequency Balanced Datasets Lead to Better Language Models

Simple Hardware-Efficient PCFGs with Independent Left and Right Productions

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES