分享

Zero-shot cross-lingual transfer by fine-tuning multilingual pretrained models shows promise for low-resource languages, but often suffers from misalignment of internal representations between languages. We hypothesize that even when the model cannot generalize across languages effectively in fine-tuning, it still captures cross-lingual correspondence useful for cross-lingual transfer. We explore this hypothesis with Self-Translate-Train, a method that lets large language models (LLMs) to translate training data into the target language and fine-tunes the model on its own generated data. By demonstrating that Self-Translate-Train outperforms zero-shot transfer, we encourage further exploration of better methods to elicit cross-lingual capabilities of LLMs.

End-to-end neural speaker diarization systems are able to address the speaker diarization task while effectively handling speech overlap. This work explores the incorporation of speaker information embeddings into the end-to-end systems to enhance the speaker discriminative capabilities, while maintaining their overlap handling strengths. To achieve this, we propose several methods for incorporating these embeddings along the acoustic features. Furthermore, we delve into an analysis of the correct handling of silence frames, the window length for extracting speaker embeddings and the transformer encoder size. The effectiveness of our proposed approach is thoroughly evaluated on the CallHome dataset for the two-speaker diarization task, with results that demonstrate a significant reduction in diarization error rates achieving a relative improvement of a 10.78% compared to the baseline end-to-end model.

The autonomous driving industry is increasingly adopting end-to-end learning from sensory inputs to minimize human biases in system design. Traditional end-to-end driving models, however, suffer from long-tail events due to rare or unseen inputs within their training distributions. To address this, we propose TOKEN, a novel Multi-Modal Large Language Model (MM-LLM) that tokenizes the world into object-level knowledge, enabling better utilization of LLM's reasoning capabilities to enhance autonomous vehicle planning in long-tail scenarios. TOKEN effectively alleviates data scarcity and inefficient tokenization by leveraging a traditional end-to-end driving model to produce condensed and semantically enriched representations of the scene, which are optimized for LLM planning compatibility through deliberate representation and reasoning alignment training stages. Our results demonstrate that TOKEN excels in grounding, reasoning, and planning capabilities, outperforming existing frameworks with a 27% reduction in trajectory L2 error and a 39% decrease in collision rates in long-tail scenarios. Additionally, our work highlights the importance of representation alignment and structured reasoning in sparking the common-sense reasoning capabilities of MM-LLMs for effective planning.

This Project was my Undergraduate Final Year dissertation, supervised by Dimitrios Kollias This research delves into the realm of affective computing for image analysis, aiming to enhance the efficiency and effectiveness of multi-task learning in the context of emotion recognition. This project investigates two primary approaches: uni-task solutions and a multi-task approach to the same problems. Each approach undergoes testing, exploring various formulations, variations, and initialization strategies to come up with the best configuration. The project utilizes existing a neural network architecture, adapting it for multi-task learning by modifying output layers and loss functions. Tasks encompass 7 basic emotion recognition, action unit detection, and valence-arousal estimation. Comparative analyses involve uni-task models for each individual task, facilitating the assessment of multi-task model performance. Variations within each approach, including, loss functions, and hyperparameter tuning, undergo evaluation. The impact of different initialization strategies and pre-training techniques on model convergence and accuracy is explored. The research aspires to contribute to the burgeoning field of affective computing, with applications spanning healthcare, marketing, and human-computer interaction. By systematically exploring multi-task learning formulations, this research aims to contribute to the development of more accurate and efficient models for recognizing and understanding emotions in images. The findings hold promise for applications in diverse industries, paving the way for advancements in affective computing

Deep learning models have revolutionized medical imaging and diagnostics, yet their opaque nature poses challenges for clinical adoption and trust. Amongst approaches to improve model interpretability, concept-based explanations aim to provide concise and human-understandable explanations of any arbitrary classifier. However, such methods usually require a large amount of manually collected data with concept annotation, which is often scarce in the medical domain. In this paper, we propose Conceptual Counterfactual Explanations for Chest X-ray (CoCoX), which leverages the joint embedding space of an existing vision-language model (VLM) to explain black-box classifier outcomes without the need for annotated datasets. Specifically, we utilize textual concepts derived from chest radiography reports and a pre-trained chest radiography-based VLM to explain three common cardiothoracic pathologies. We demonstrate that the explanations generated by our method are semantically meaningful and faithful to underlying pathologies.

With the development of deep learning, many different network architectures have been explored in speaker verification. However, most network architectures rely on a single deep learning architecture, and hybrid networks combining different architectures have been little studied in ASV tasks. In this paper, we propose the GMM-ResNext model for speaker verification. Conventional GMM does not consider the score distribution of each frame feature over all Gaussian components and ignores the relationship between neighboring speech frames. So, we extract the log Gaussian probability features based on the raw acoustic features and use ResNext-based network as the backbone to extract the speaker embedding. GMM-ResNext combines Generative and Discriminative Models to improve the generalization ability of deep learning models and allows one to more easily specify meaningful priors on model parameters. A two-path GMM-ResNext model based on two gender-related GMMs has also been proposed. The Experimental results show that the proposed GMM-ResNext achieves relative improvements of 48.1\% and 11.3\% in EER compared with ResNet34 and ECAPA-TDNN on VoxCeleb1-O test set.

Electroencephalography (EEG), a medical imaging technique that captures scalp electrical activity of brain structures via electrodes, has been widely used in affective computing. The spatial domain of EEG is rich in affective information. However, few of the existing studies have simultaneously analyzed EEG signals from multiple perspectives of geometric and anatomical structures in spatial domain. In this paper, we propose a multi-view Graph Transformer (MVGT) based on spatial relations, which integrates information from the temporal, frequency and spatial domains, including geometric and anatomical structures, so as to enhance the expressive power of the model comprehensively. We incorporate the spatial information of EEG channels into the model as encoding, thereby improving its ability to perceive the spatial structure of the channels. Meanwhile, experimental results based on publicly available datasets demonstrate that our proposed model outperforms state-of-the-art methods in recent years. In addition, the results also show that the MVGT could extract information from multiple domains and capture inter-channel relationships in EEG emotion recognition tasks effectively.

Emotion recognition from electroencephalogram (EEG) signals is a thriving field, particularly in neuroscience and Human-Computer Interaction (HCI). This study aims to understand and improve the predictive accuracy of emotional state classification through metrics such as valence, arousal, dominance, and likeness by applying a Long Short-Term Memory (LSTM) network to analyze EEG signals. Using a popular dataset of multi-channel EEG recordings known as DEAP, we look towards leveraging LSTM networks' properties to handle temporal dependencies within EEG signal data. This allows for a more comprehensive understanding and classification of emotional parameter states. We obtain accuracies of 89.89%, 90.33%, 90.70%, and 90.54% for arousal, valence, dominance, and likeness, respectively, demonstrating significant improvements in emotion recognition model capabilities. This paper elucidates the methodology and architectural specifics of our LSTM model and provides a benchmark analysis with existing papers.

Retrosynthesis analysis is pivotal yet challenging in drug discovery and organic chemistry. Despite the proliferation of computational tools over the past decade, AI-based systems often fall short in generalizing across diverse reaction types and exploring alternative synthetic pathways. This paper presents BatGPT-Chem, a large language model with 15 billion parameters, tailored for enhanced retrosynthesis prediction. Integrating chemical tasks via a unified framework of natural language and SMILES notation, this approach synthesizes extensive instructional data from an expansive chemical database. Employing both autoregressive and bidirectional training techniques across over one hundred million instances, BatGPT-Chem captures a broad spectrum of chemical knowledge, enabling precise prediction of reaction conditions and exhibiting strong zero-shot capabilities. Superior to existing AI methods, our model demonstrates significant advancements in generating effective strategies for complex molecules, as validated by stringent benchmark tests. BatGPT-Chem not only boosts the efficiency and creativity of retrosynthetic analysis but also establishes a new standard for computational tools in synthetic design. This development empowers chemists to adeptly address the synthesis of novel compounds, potentially expediting the innovation cycle in drug manufacturing and materials science. We release our trial platform at \url{https://www.batgpt.net/dapp/chem}.

We find that the cross-entropy loss curves of neural language models empirically adhere to a scaling law with learning rate (LR) annealing over training steps: $$L(s) = L_0 + A\cdot S_1^{-\alpha} - C\cdot S_2,$$ where $L(s)$ is the validation loss at step $s$, $S_1$ is the area under the LR curve, $S_2$ is the LR annealing area, and $L_0$, $A$, $C$, $\alpha$ are constant parameters. This formulation takes into account two factors: (1) power-law scaling over data size, and (2) the additional loss reduction during LR annealing. Therefore, this formulation can describe the full loss curve at each step, rather than the single loss point at the end of training. Applying the scaling law with LR annealing and fitting only one or two training curves, we can accurately predict the loss at any given step across any learning rate scheduler (LRS). This approach significantly reduces computational cost in formulating scaling laws while providing more accuracy and expressiveness for training dynamics. Extensive experiments demonstrate that our findings hold across a range of hyper-parameters and model architectures, and our equation can extend to scaling effect of model sizes. Moreover, our formulation provides accurate theoretical verification and explanation for empirical results observed in numerous previous studies, particularly those focusing on LR schedule and annealing. We believe that this work is promising to enhance the understanding of LLM training dynamics while greatly democratizing scaling laws, and it can guide researchers in refining training strategies (e.g. critical LRS) for further LLMs.

Although instruction tuning is widely used to adjust behavior in Large Language Models (LLMs), extensive empirical evidence and research indicates that it is primarily a process where the model fits to specific task formats, rather than acquiring new knowledge or capabilities. We propose that this limitation stems from biased features learned during instruction tuning, which differ from ideal task-specfic features, leading to learn less underlying semantics in downstream tasks. However, ideal features are unknown and incalculable, constraining past work to rely on prior knowledge to assist reasoning or training, which limits LLMs' capabilities to the developers' abilities, rather than data-driven scalable learning. In our paper, through our novel data synthesis method, DELIA (Diversity-Enhanced Learning for Instruction Adaptation), we leverage the buffering effect of extensive diverse data in LLMs training to transform biased features in instruction tuning into approximations of ideal features, without explicit prior ideal features. Experiments show DELIA's better performance compared to common instruction tuning and other baselines. It outperforms common instruction tuning by 17.07%-33.41% on Icelandic-English translation bleurt score (WMT-21 dataset, gemma-7b-it) and improves accuracy by 36.1% on formatted text generation (Llama2-7b-chat). Notably, among knowledge injection methods we've known, DELIA uniquely align the internal representations of new special tokens with their prior semantics.

Gloss-free Sign Language Translation (SLT) converts sign videos directly into spoken language sentences without relying on glosses. Recently, Large Language Models (LLMs) have shown remarkable translation performance in gloss-free methods by harnessing their powerful natural language generation capabilities. However, these methods often rely on domain-specific fine-tuning of visual encoders to achieve optimal results. By contrast, this paper emphasizes the importance of capturing the spatial configurations and motion dynamics inherent in sign language. With this in mind, we introduce Spatial and Motion-based Sign Language Translation (SpaMo), a novel LLM-based SLT framework. The core idea of SpaMo is simple yet effective. We first extract spatial and motion features using off-the-shelf visual encoders and then input these features into an LLM with a language prompt. Additionally, we employ a visual-text alignment process as a warm-up before the SLT supervision. Our experiments demonstrate that SpaMo achieves state-of-the-art performance on two popular datasets, PHOENIX14T and How2Sign.