K-Origins is a neural network layer designed to improve image-based network performances when learning colour, or intensities, is beneficial. Over 250 encoder-decoder convolutional networks are trained and tested on 16-bit synthetic data, demonstrating that K-Origins improves semantic segmentation accuracy in two scenarios: object detection with low signal-to-noise ratios, and segmenting multiple objects that are identical in shape but vary in colour. K-Origins generates output features from the input features, $\textbf{X}$, by the equation $\textbf{Y}_k = \textbf{X}-\textbf{J}\cdot w_k$ for each trainable parameter $w_k$, where $\textbf{J}$ is a matrix of ones. Additionally, networks with varying receptive fields were trained to determine optimal network depths based on the dimensions of target classes, suggesting that receptive field lengths should exceed object sizes. By ensuring a sufficient receptive field length and incorporating K-Origins, we can achieve better semantic network performance.
Dialect adapters that improve the performance of LLMs for NLU tasks on certain sociolects/dialects/national varieties ('dialects' for the sake of brevity) have been reported for encoder models. In this paper, we extend the idea of dialect adapters to decoder models in our architecture called LoRDD. Using MD-3, a publicly available dataset of word game-playing conversations between dialectal speakers, our task is Target Word Prediction (TWP) from a masked conversation. LoRDD combines task adapters and dialect adapters where the latter employ contrastive learning on pseudo-parallel conversations from MD-3. Our experiments on Indian English and Nigerian English conversations with two models (Mistral and Gemma) demonstrate that LoRDD outperforms four baselines on TWP. Additionally, it significantly reduces the performance gap with American English, narrowing it to 12% and 5.8% for word similarity, and 25% and 4.5% for accuracy, respectively. The focused contribution of LoRDD is in its promise for dialect adaptation of decoder models using TWP, a simplified version of the commonly used next-word prediction task.
To accurately understand engineering drawings, it is essential to establish the correspondence between images and their description tables within the drawings. Existing document understanding methods predominantly focus on text as the main modality, which is not suitable for documents containing substantial image information. In the field of visual relation detection, the structure of the task inherently limits its capacity to assess relationships among all entity pairs in the drawings. To address this issue, we propose a vision-based relation detection model, named ViRED, to identify the associations between tables and circuits in electrical engineering drawings. Our model mainly consists of three parts: a vision encoder, an object encoder, and a relation decoder. We implement ViRED using PyTorch to evaluate its performance. To validate the efficacy of ViRED, we conduct a series of experiments. The experimental results indicate that, within the engineering drawing dataset, our approach attained an accuracy of 96\% in the task of relation prediction, marking a substantial improvement over existing methodologies. The results also show that ViRED can inference at a fast speed even when there are numerous objects in a single engineering drawing.
One-shot voice conversion(VC) aims to change the timbre of any source speech to match that of the target speaker with only one speech sample. Existing style transfer-based VC methods relied on speech representation disentanglement and suffered from accurately and independently encoding each speech component and recomposing back to converted speech effectively. To tackle this, we proposed Pureformer-VC, which utilizes Conformer blocks to build a disentangled encoder, and Zipformer blocks to build a style transfer decoder as the generator. In the decoder, we used effective styleformer blocks to integrate speaker characteristics effectively into the generated speech. The models used the generative VAE loss for encoding components and triplet loss for unsupervised discriminative training. We applied the styleformer method to Zipformer's shared weights for style transfer. The experimental results show that the proposed model achieves comparable subjective scores and exhibits improvements in objective metrics compared to existing methods in a one-shot voice conversion scenario.
In classification tasks, achieving a harmonious balance between exploration and precision is of paramount importance. To this end, this research introduces two novel deep learning models, SleepNet and DreamNet, to strike this balance. SleepNet seamlessly integrates supervised learning with unsupervised ``sleep" stages using pre-trained encoder models. Dedicated neurons within SleepNet are embedded in these unsupervised features, forming intermittent ``sleep" blocks that facilitate exploratory learning. Building upon the foundation of SleepNet, DreamNet employs full encoder-decoder frameworks to reconstruct the hidden states, mimicking the human "dreaming" process. This reconstruction process enables further exploration and refinement of the learned representations. Moreover, the principle ideas of our SleepNet and DreamNet are generic and can be applied to both computer vision and natural language processing downstream tasks. Through extensive empirical evaluations on diverse image and text datasets, SleepNet and DreanNet have demonstrated superior performance compared to state-of-the-art models, showcasing the strengths of unsupervised exploration and supervised precision afforded by our innovative approaches.
Convolutional layers with 1-D filters are often used as frontend to encode audio signals. Unlike fixed time-frequency representations, they can adapt to the local characteristics of input data. However, 1-D filters on raw audio are hard to train and often suffer from instabilities. In this paper, we address these problems with hybrid solutions, i.e., combining theory-driven and data-driven approaches. First, we preprocess the audio signals via a auditory filterbank, guaranteeing good frequency localization for the learned encoder. Second, we use results from frame theory to define an unsupervised learning objective that encourages energy conservation and perfect reconstruction. Third, we adapt mixed compressed spectral norms as learning objectives to the encoder coefficients. Using these solutions in a low-complexity encoder-mask-decoder model significantly improves the perceptual evaluation of speech quality (PESQ) in speech enhancement.
Visual Question Answering with Natural Language Explanation (VQA-NLE) task is challenging due to its high demand for reasoning-based inference. Recent VQA-NLE studies focus on enhancing model networks to amplify the model's reasoning capability but this approach is resource-consuming and unstable. In this work, we introduce a new VQA-NLE model, ReRe (Retrieval-augmented natural language Reasoning), using leverage retrieval information from the memory to aid in generating accurate answers and persuasive explanations without relying on complex networks and extra datasets. ReRe is an encoder-decoder architecture model using a pre-trained clip vision encoder and a pre-trained GPT-2 language model as a decoder. Cross-attention layers are added in the GPT-2 for processing retrieval features. ReRe outperforms previous methods in VQA accuracy and explanation score and shows improvement in NLE with more persuasive, reliability.
Multimodal variational autoencoders have demonstrated their ability to learn the relationships between different modalities by mapping them into a latent representation. Their design and capacity to perform any-to-any conditional and unconditional generation make them appealing. However, different variants of multimodal VAEs often suffer from generating low-quality output, particularly when complex modalities such as images are involved. In addition to that, they frequently exhibit low coherence among the generated modalities when sampling from the joint distribution. To address these limitations, we propose a new variant of the multimodal VAE ELBO that incorporates a better decoder using a diffusion generative model. The diffusion decoder enables the model to learn complex modalities and generate high-quality outputs. The multimodal model can also seamlessly integrate with a standard feed-forward decoder for different types of modality, facilitating end-to-end training and inference. Furthermore, we introduce an auxiliary score-based model to enhance the unconditional generation capabilities of our proposed approach. This approach addresses the limitations imposed by conventional multimodal VAEs and opens up new possibilities to improve multimodal generation tasks. Our model provides state-of-the-art results compared to other multimodal VAEs in different datasets with higher coherence and superior quality in the generated modalities.
Recognizing overlapping speech from multiple speakers in conversational scenarios is one of the most challenging problem for automatic speech recognition (ASR). Serialized output training (SOT) is a classic method to address multi-talker ASR, with the idea of concatenating transcriptions from multiple speakers according to the emission times of their speech for training. However, SOT-style transcriptions, derived from concatenating multiple related utterances in a conversation, depend significantly on modeling long contexts. Therefore, compared to traditional methods that primarily emphasize encoder performance in attention-based encoder-decoder (AED) architectures, a novel approach utilizing large language models (LLMs) that leverages the capabilities of pre-trained decoders may be better suited for such complex and challenging scenarios. In this paper, we propose an LLM-based SOT approach for multi-talker ASR, leveraging pre-trained speech encoder and LLM, fine-tuning them on multi-talker dataset using appropriate strategies. Experimental results demonstrate that our approach surpasses traditional AED-based methods on the simulated dataset LibriMix and achieves state-of-the-art performance on the evaluation set of the real-world dataset AMI, outperforming the AED model trained with 1000 times more supervised data in previous works.
Federated learning is a machine learning paradigm that enables decentralized clients to collaboratively learn a shared model while keeping all the training data local. While considerable research has focused on federated image generation, particularly Generative Adversarial Networks, Variational Autoencoders have received less attention. In this paper, we address the challenges of non-IID (independently and identically distributed) data environments featuring multiple groups of images of different types. Non-IID data distributions can lead to difficulties in maintaining a consistent latent space and can also result in local generators with disparate texture features being blended during aggregation. We thereby introduce FissionVAE that decouples the latent space and constructs decoder branches tailored to individual client groups. This method allows for customized learning that aligns with the unique data distributions of each group. Additionally, we incorporate hierarchical VAEs and demonstrate the use of heterogeneous decoder architectures within FissionVAE. We also explore strategies for setting the latent prior distributions to enhance the decoupling process. To evaluate our approach, we assemble two composite datasets: the first combines MNIST and FashionMNIST; the second comprises RGB datasets of cartoon and human faces, wild animals, marine vessels, and remote sensing images. Our experiments demonstrate that FissionVAE greatly improves generation quality on these datasets compared to baseline federated VAE models.
Recent advancements in Large Language Models (LLMs) have yielded remarkable success across diverse fields. However, handling long contexts remains a significant challenge for LLMs due to the quadratic time and space complexity of attention mechanisms and the growing memory consumption of the key-value cache during generation. This work introduces MemLong: Memory-Augmented Retrieval for Long Text Generation, a method designed to enhance the capabilities of long-context language modeling by utilizing an external retriever for historical information retrieval. MemLong combines a non-differentiable ``ret-mem'' module with a partially trainable decoder-only language model and introduces a fine-grained, controllable retrieval attention mechanism that leverages semantic-level relevant chunks. Comprehensive evaluations on multiple long-context language modeling benchmarks demonstrate that MemLong consistently outperforms other state-of-the-art LLMs. More importantly, MemLong can extend the context length on a single 3090 GPU from 4k up to 80k. Our code is available at https://github.com/Bui1dMySea/MemLong
Intracranial electroencephalography (iEEG) is increasingly used for clinical and brain-computer interface applications due to its high spatial and temporal resolution. However, inter-subject variability in electrode implantation poses a challenge for developing generalized neural decoders. To address this, we introduce a novel decoder model that is robust to inter-subject electrode implantation variability. We call this model RISE-iEEG, which stands for Robust to Inter-Subject Electrode Implantation Variability iEEG Classifier. RISE-iEEG employs a deep neural network structure preceded by a participant-specific projection network. The projection network maps the neural data of individual participants onto a common low-dimensional space, compensating for the implantation variability. In other words, we developed an iEEG decoder model that can be applied across multiple participants' data without requiring the coordinates of electrode for each participant. The performance of RISE-iEEG across multiple datasets, including the Music Reconstruction dataset, and AJILE12 dataset, surpasses that of advanced iEEG decoder models such as HTNet and EEGNet. Our analysis shows that the performance of RISE-iEEG is about 7\% higher than that of HTNet and EEGNet in terms of F1 score, with an average F1 score of 0.83, which is the highest result among the evaluation methods defined. Furthermore, Our analysis of the projection network weights reveals that the Superior Temporal and Postcentral lobes are key encoding nodes for the Music Reconstruction and AJILE12 datasets, which aligns with the primary physiological principles governing these regions. This model improves decoding accuracy while maintaining interpretability and generalization.
This paper presents an innovative approach to mitigating the peak-to-average power ratio (PAPR). The proposed method uses a deep learning model called autoencoders (AEs) to simplify the process and avoid the complex calculations of traditional methods such as selective mapping (SLM). Unlike SLM, our approach does not need side information about the PAPR distribution. Through simulations of coherent optical orthogonal frequency division multiplexing (CO-OFDM) systems, the AE-based model offers substantial enhancements in both PAPR reduction and bit error rate (BER) performance when compared to conventional techniques. An error-free transmission can be acheived with a reduction in PAPR exceeding 10 dB compared to the original signal and a 1 dB advantage over SLM. In particular, the AE model achieves the best BER performance of \(2 \times 10^{-6}\) at 44 dB OSNR, surpassing traditional methods. Furthermore, the model demonstrates robustness against noise and nonlinear distortions, making it appropriate for optical channels experiencing diverse levels of impairment. This innovative technique has the potential to revolutionize next-generation optical communication systems by enabling efficient and reliable data transmission.
This thesis provides methods and analysis of models which make progress on this goal. The techniques outlined are task agnostic, and should provide benefit when used with nearly any transformer LM. We introduce two new finetuning methods which add new capabilities to the models they are used on. The first adds a recurrence mechanism, which removes the fixed-window sized constraint and improves the efficiency of a transformer decoder. The second allows masked language models (MLMs) to be used for initialization of both the encoder and decoder of a non-autoregressive sequence-to-sequence transformer, opening up generative applications of models which were previously only used for natural language understanding tasks. We also introduce two new techniques for improving the quality of predictions of any transformer decoder without additional finetuning. One, hidden state optimization, can be applied to any transformer decoder to improve the quality of predictions at inference time, especially for few-shot classification. The other, conditional beam search, allows practitioners to search for natural language generation (NLG) model outputs with high likelihood while conditioning on the event that the output is not degenerate (e.g. empty, repetitive, etc.). Finally, we provide theoretical and empirical insights on the divergence of model-likelihood and output quality which has widely been observed in prior work. These insights apply to any model which represents a distribution over text, and apply to language models which are not transformers or even autoregressive. We argue that the NLP community has, to some extent, misunderstood the implications of these findings, and encourage a point of view which has more nuance.
Identifying the physical properties of the surrounding environment is essential for robotic locomotion and navigation to deal with non-geometric hazards, such as slippery and deformable terrains. It would be of great benefit for robots to anticipate these extreme physical properties before contact; however, estimating environmental physical parameters from vision is still an open challenge. Animals can achieve this by using their prior experience and knowledge of what they have seen and how it felt. In this work, we propose a cross-modal self-supervised learning framework for vision-based environmental physical parameter estimation, which paves the way for future physical-property-aware locomotion and navigation. We bridge the gap between existing policies trained in simulation and identification of physical terrain parameters from vision. We propose to train a physical decoder in simulation to predict friction and stiffness from multi-modal input. The trained network allows the labeling of real-world images with physical parameters in a self-supervised manner to further train a visual network during deployment, which can densely predict the friction and stiffness from image data. We validate our physical decoder in simulation and the real world using a quadruped ANYmal robot, outperforming an existing baseline method. We show that our visual network can predict the physical properties in indoor and outdoor experiments while allowing fast adaptation to new environments.
Harnessing pre-trained LLMs to improve ASR systems, particularly for low-resource languages, is now an emerging area of research. Existing methods range from using LLMs for ASR error correction to tightly coupled systems that replace the ASR decoder with the LLM. These approaches either increase decoding time or require expensive training of the cross-attention layers. We propose SALSA, which couples the decoder layers of the ASR to the LLM decoder, while synchronously advancing both decoders. Such coupling is performed with a simple projection of the last decoder state, and is thus significantly more training efficient than earlier approaches. A challenge of our proposed coupling is handling the mismatch between the tokenizers of the LLM and ASR systems. We handle this mismatch using cascading tokenization with respect to the LLM and ASR vocabularies. We evaluate SALSA on 8 low-resource languages in the FLEURS benchmark, yielding substantial WER reductions of up to 38%.
This paper presents Dolphin, a novel decoder-decoder architecture for energy-efficient processing of long contexts in language models. Our approach addresses the significant energy consumption and latency challenges inherent in on-device models. Dolphin employs a compact 0.5B parameter decoder to distill extensive contextual information into a memory embedding, substantially reducing the input length for the primary 7B parameter decoder model. Inspired by vision-language models, we repurpose the image embedding projector to encode long textual contexts, effectively treating extended context as a distinct modality. This innovative method enables processing of substantially longer contexts without the typical computational overhead associated with extended input sequences. Empirical evaluations demonstrate a 10-fold improvement in energy efficiency and a 5-fold reduction in latency compared to conventional full-length context processing methods without losing quality of the response. Our work contributes to the development of more sustainable and scalable language models for on-device applications, addressing the critical need for energy-efficient and responsive AI technologies in resource-constrained environments while maintaining the accuracy to understand long contexts. This research has implications for the broader field of natural language processing, particularly in the domain of efficient model design for resource-limited settings. By enabling more sophisticated AI capabilities on edge devices, Dolphin paves the way for advanced language processing in a wide range of applications where computational resources are at a premium. The Dolphin model is publicly available at https://huggingface.co/NexaAIDev/Dolphin.
A text-to-speech (TTS) model trained to reconstruct speech given text tends towards predictions that are close to the average characteristics of a dataset, failing to model the variations that make human speech sound natural. This problem is magnified for zero-shot voice cloning, a task that requires training data with high variance in speaking styles. We build off of recent works which have used Generative Advsarial Networks (GAN) by proposing a Transformer encoder-decoder architecture to conditionally discriminates between real and generated speech features. The discriminator is used in a training pipeline that improves both the acoustic and prosodic features of a TTS model. We introduce our novel adversarial training technique by applying it to a FastSpeech2 acoustic model and training on Libriheavy, a large multi-speaker dataset, for the task of zero-shot voice cloning. Our model achieves improvements over the baseline in terms of speech quality and speaker similarity. Audio examples from our system are available online.
LiDAR panoptic segmentation, which jointly performs instance and semantic segmentation for things and stuff classes, plays a fundamental role in LiDAR perception tasks. While most existing methods explicitly separate these two segmentation tasks and utilize different branches (i.e., semantic and instance branches), some recent methods have embraced the query-based paradigm to unify LiDAR panoptic segmentation. However, the distinct spatial distribution and inherent characteristics of objects(things) and their surroundings(stuff) in 3D scenes lead to challenges, including the mutual competition of things/stuff and the ambiguity of classification/segmentation. In this paper, we propose decoupling things/stuff queries according to their intrinsic properties for individual decoding and disentangling classification/segmentation to mitigate ambiguity. To this end, we propose a novel framework dubbed DQFormer to implement semantic and instance segmentation in a unified workflow. Specifically, we design a decoupled query generator to propose informative queries with semantics by localizing things/stuff positions and fusing multi-level BEV embeddings. Moreover, a query-oriented mask decoder is introduced to decode corresponding segmentation masks by performing masked cross-attention between queries and mask embeddings. Finally, the decoded masks are combined with the semantics of the queries to produce panoptic results. Extensive experiments on nuScenes and SemanticKITTI datasets demonstrate the superiority of our DQFormer framework.
Referring image segmentation (RIS) requires dense vision-language interactions between visual pixels and textual words to segment objects based on a given description. However, commonly adapted dual-encoders in RIS, e.g., Swin transformer and BERT (uni-modal encoders) or CLIP (a multi-modal dual-encoder), lack dense multi-modal interactions during pre-training, leading to a gap with a pixel-level RIS task. To bridge this gap, existing RIS methods often rely on multi-modal fusion modules that interact two encoders, but this approach leads to high computational costs. In this paper, we present a novel RIS method with a single-encoder, i.e., BEiT-3, maximizing the potential of shared self-attention across all framework components. This enables seamless interactions of two modalities from input to final prediction, producing granularly aligned multi-modal features. Furthermore, we propose lightweight yet effective decoder modules, a Shared FPN and a Shared Mask Decoder, which contribute to the high efficiency of our model. Our simple baseline with a single encoder achieves outstanding performances on the RIS benchmark datasets while maintaining computational efficiency, compared to the most recent SoTA methods based on dual-encoders.