Multi-modal domain translation typically refers to synthesizing a novel image that inherits certain localized attributes from a 'content' image (e.g. layout, semantics, or geometry), and inherits everything else (e.g. texture, lighting, sometimes even semantics) from a 'style' image. The dominant approach to this task is attempting to learn disentangled 'content' and 'style' representations from scratch. However, this is not only challenging, but ill-posed, as what users wish to preserve during translation varies depending on their goals. Motivated by this inherent ambiguity, we define 'content' based on conditioning information extracted by off-the-shelf pre-trained models. We then train our style extractor and image decoder with an easy to optimize set of reconstruction objectives. The wide variety of high-quality pre-trained models available and simple training procedure makes our approach straightforward to apply across numerous domains and definitions of 'content'. Additionally it offers intuitive control over which aspects of 'content' are preserved across domains. We evaluate our method on traditional, well-aligned, datasets such as CelebA-HQ, and propose two novel datasets for evaluation on more complex scenes: ClassicTV and FFHQ-Wild. Our approach, Sensorium, enables higher quality domain translation for more complex scenes.
Generating the periodic structure of stable materials is a long-standing challenge for the material design community. This task is difficult because stable materials only exist in a low-dimensional subspace of all possible periodic arrangements of atoms: 1) the coordinates must lie in the local energy minimum defined by quantum mechanics, and 2) global stability also requires the structure to follow the complex, yet specific bonding preferences between different atom types. Existing methods fail to incorporate these factors and often lack proper invariances. We propose a Crystal Diffusion Variational Autoencoder (CDVAE) that captures the physical inductive bias of material stability. By learning from the data distribution of stable materials, the decoder generates materials in a diffusion process that moves atomic coordinates towards a lower energy state and updates atom types to satisfy bonding preferences between neighbors. Our model also explicitly encodes interactions across periodic boundaries and respects permutation, translation, rotation, and periodic invariances. We significantly outperform past methods in three tasks: 1) reconstructing the input structure, 2) generating valid, diverse, and realistic materials, and 3) generating materials that optimize a specific property. We also provide several standard datasets and evaluation metrics for the broader machine learning community.
Medical image analysis continues to hold interesting challenges given the subtle characteristics of certain diseases and the significant overlap in appearance between diseases. In this work, we explore the concept of self-attention for tackling such subtleties in and between diseases. To this end, we introduce MEDUSA, a multi-scale encoder-decoder self-attention mechanism tailored for medical image analysis. While self-attention deep convolutional neural network architectures in existing literature center around the notion of multiple isolated lightweight attention mechanisms with limited individual capacities being incorporated at different points in the network architecture, MEDUSA takes a significant departure from this notion by possessing a single, unified self-attention mechanism with significantly higher capacity with multiple attention heads feeding into different scales in the network architecture. To the best of the authors' knowledge, this is the first "single body, multi-scale heads" realization of self-attention and enables explicit global context amongst selective attention at different levels of representational abstractions while still enabling differing local attention context at individual levels of abstractions. With MEDUSA, we obtain state-of-the-art performance on multiple challenging medical image analysis benchmarks including COVIDx, RSNA RICORD, and RSNA Pneumonia Challenge when compared to previous work. Our MEDUSA model is publicly available.
In contrast to Connectionist Temporal Classification (CTC) approaches, Sequence-To-Sequence (S2S) models for Handwritten Text Recognition (HTR) suffer from errors such as skipped or repeated words which often occur at the end of a sequence. In this paper, to combine the best of both approaches, we propose to use the CTC-Prefix-Score during S2S decoding. Hereby, during beam search, paths that are invalid according to the CTC confidence matrix are penalised. Our network architecture is composed of a Convolutional Neural Network (CNN) as visual backbone, bidirectional Long-Short-Term-Memory-Cells (LSTMs) as encoder, and a decoder which is a Transformer with inserted mutual attention layers. The CTC confidences are computed on the encoder while the Transformer is only used for character-wise S2S decoding. We evaluate this setup on three HTR data sets: IAM, Rimes, and StAZH. On IAM, we achieve a competitive Character Error Rate (CER) of 2.95% when pretraining our model on synthetic data and including a character-based language model for contemporary English. Compared to other state-of-the-art approaches, our model requires about 10-20 times less parameters. Access our shared implementations via this link to GitHub: https://github.com/Planet-AI-GmbH/tfaip-hybrid-ctc-s2s.
Monocular depth estimation is the base task in computer vision. It has a tremendous development in the decade with the development of deep learning. But the boundary blur of the depth map is still a serious problem. Research finds the boundary blur problem is mainly caused by two factors, first, the low-level features containing boundary and structure information may loss in deeper networks during the convolution process., second, the model ignores the errors introduced by the boundary area due to the few portions of the boundary in the whole areas during the backpropagation. In order to mitigate the boundary blur problem, we focus on the above two impact factors. Firstly, we design a scene understanding module to learn the global information with low- and high-level features, and then to transform the global information to different scales with our proposed scale transform module according to the different phases in the decoder. Secondly, we propose a boundary-aware depth loss function to pay attention to the effects of the boundary's depth value. The extensive experiments show that our method can predict the depth maps with clearer boundaries, and the performance of the depth accuracy base on NYU-depth v2 and SUN RGB-D is competitive.
Surface code error correction offers a highly promising pathway to achieve scalable fault-tolerant quantum computing. When operated as stabilizer codes, surface code computations consist of a syndrome decoding step where measured stabilizer operators are used to determine appropriate corrections for errors in physical qubits. Decoding algorithms have undergone substantial development, with recent work incorporating machine learning (ML) techniques. Despite promising initial results, the ML-based syndrome decoders are still limited to small scale demonstrations with low latency and are incapable of handling surface codes with boundary conditions and various shapes needed for lattice surgery and braiding. Here, we report the development of an artificial neural network (ANN) based scalable and fast syndrome decoder capable of decoding surface codes of arbitrary shape and size with data qubits suffering from the depolarizing error model. Based on rigorous training over 50 million random quantum error instances, our ANN decoder is shown to work with code distances exceeding 1000 (more than 4 million physical qubits), which is the largest ML-based decoder demonstration to-date. The established ANN decoder demonstrates an execution time in principle independent of code distance, implying that its implementation on dedicated hardware could potentially offer surface code decoding times of O($\mu$sec), commensurate with the experimentally realisable qubit coherence times. With the anticipated scale-up of quantum processors within the next decade, their augmentation with a fast and scalable syndrome decoder such as developed in our work is expected to play a decisive role towards experimental implementation of fault-tolerant quantum information processing.
Recently introduced implicit field representations offer an effective way of generating 3D object shapes. They leverage implicit decoder trained to take a 3D point coordinate concatenated with a shape encoding and to output a value which indicates whether the point is outside the shape or not. Although this approach enables efficient rendering of visually plausible objects, it has two significant limitations. First, it is based on a single neural network dedicated for all objects from a training set which results in a cumbersome training procedure and its application in real life. More importantly, the implicit decoder takes only points sampled within voxels (and not the entire voxels) which yields problems at the classification boundaries and results in empty spaces within the rendered mesh. To solve the above limitations, we introduce a new HyperCube architecture based on interval arithmetic network, that enables direct processing of 3D voxels, trained using a hypernetwork paradigm to enforce model convergence. Instead of processing individual 3D samples from within a voxel, our approach allows to input the entire voxel (3D cube) represented with its convex hull coordinates, while the target network constructed by a hypernet assigns it to an inside or outside category. As a result our HyperCube model outperforms the competing approaches both in terms of training and inference efficiency, as well as the final mesh quality.
Transformer-based neural models are used in many AI applications. Training these models is expensive, as it takes huge GPU resources and long duration. It is challenging because typical data like sentences have variable lengths, and Transformer's computation patterns are more complex than convolutional neural networks. Existing systems either only focus on model inference or optimization for only BERT-like encoder models. In this paper, we present LightSeq2, a system to accelerate training for a general family of Transformer models on GPUs. We propose a series of GPU optimization techniques tailored to the specific computation flow and memory access patterns of Transformer models. LightSeq2 supports many model architectures, including BERT (encoder-only), GPT (decoder-only), Transformer (encoder-decoder), and vision Transformer. Our experiments for a variety of models and benchmarks show that LightSeq2 is consistently faster (1.4-3.5x) than previous systems on different GPUs. In particular, it gains 308% training speedup compared with existing systems on a large public machine translation benchmark (WMT14 English-German).
Wav2vec 2.0 is an end-to-end framework of self-supervised learning for speech representation that is successful in automatic speech recognition (ASR), but most of the work on the topic has been developed with a single language: English. Therefore, it is unclear whether the self-supervised framework is effective in recognizing other languages with different writing systems, such as Korean which uses the Hangul having a unique writing system. In this paper, we present K-Wav2Vec 2.0, which is a modified version of Wav2vec 2.0 designed for Korean automatic speech recognition by exploring and optimizing various factors of the original Wav2vec 2.0. In fine-tuning, we propose a multi-task hierarchical architecture to reflect the Korean writing structure. Moreover, a joint decoder is applied to alleviate the problem of words existing outside of the vocabulary. In pre-training, we attempted the cross-lingual transfer of the pre-trained model by further pre-training the English Wav2vec 2.0 on a Korean dataset, considering limited resources. Our experimental results demonstrate that the proposed method yields the best performance on both Korean ASR datasets: Ksponspeech (a large-scale Korean speech corpus) and Clovacall (a call-based dialog corpus). Further pre-training is also effective in language adaptation, leading to large improvements without additional data.
Aspect Sentiment Triplet Extraction (ASTE) deals with extracting opinion triplets, consisting of an opinion target or aspect, its associated sentiment, and the corresponding opinion term/span explaining the rationale behind the sentiment. Existing research efforts are majorly tagging-based. Among the methods taking a sequence tagging approach, some fail to capture the strong interdependence between the three opinion factors, whereas others fall short of identifying triplets with overlapping aspect/opinion spans. A recent grid tagging approach on the other hand fails to capture the span-level semantics while predicting the sentiment between an aspect-opinion pair. Different from these, we present a tagging-free solution for the task, while addressing the limitations of the existing works. We adapt an encoder-decoder architecture with a Pointer Network-based decoding framework that generates an entire opinion triplet at each time step thereby making our solution end-to-end. Interactions between the aspects and opinions are effectively captured by the decoder by considering their entire detected spans while predicting their connecting sentiment. Extensive experiments on several benchmark datasets establish the better efficacy of our proposed approach, especially in the recall, and in predicting multiple and aspect/opinion-overlapped triplets from the same review sentence. We report our results both with and without BERT and also demonstrate the utility of domain-specific BERT post-training for the task.
Summarization systems make numerous "decisions" about summary properties during inference, e.g. degree of copying, specificity and length of outputs, etc. However, these are implicitly encoded within model parameters and specific styles cannot be enforced. To address this, we introduce HydraSum, a new summarization architecture that extends the single decoder framework of current models to a mixture-of-experts version with multiple decoders. We show that HydraSum's multiple decoders automatically learn contrasting summary styles when trained under the standard training objective without any extra supervision. Through experiments on three summarization datasets (CNN, Newsroom and XSum), we show that HydraSum provides a simple mechanism to obtain stylistically-diverse summaries by sampling from either individual decoders or their mixtures, outperforming baseline models. Finally, we demonstrate that a small modification to the gating strategy during training can enforce an even stricter style partitioning, e.g. high- vs low-abstractiveness or high- vs low-specificity, allowing users to sample from a larger area in the generation space and vary summary styles along multiple dimensions.
There have been significant research activities in recent years to automate the design of channel encoders and decoders via deep learning. Due the dimensionality challenge in channel coding, it is prohibitively complex to design and train relatively large neural channel codes via deep learning techniques. Consequently, most of the results in the literature are limited to relatively short codes having less than 100 information bits. In this paper, we construct ProductAEs, a computationally efficient family of deep-learning driven (encoder, decoder) pairs, that aim at enabling the training of relatively large channel codes (both encoders and decoders) with a manageable training complexity. We build upon the ideas from classical product codes, and propose constructing large neural codes using smaller code components. More specifically, instead of directly training the encoder and decoder for a large neural code of dimension $k$ and blocklength $n$, we provide a framework that requires training neural encoders and decoders for the code parameters $(n_1,k_1)$ and $(n_2,k_2)$ such that $n_1 n_2=n$ and $k_1 k_2=k$. Our training results show significant gains, over all ranges of signal-to-noise ratio (SNR), for a code of parameters $(225,100)$ and a moderate-length code of parameters $(441,196)$, over polar codes under successive cancellation (SC) decoder. Moreover, our results demonstrate meaningful gains over Turbo Autoencoder (TurboAE) and state-of-the-art classical codes. This is the first work to design product autoencoders and a pioneering work on training large channel codes.
Current Open-Domain Question Answering (ODQA) model paradigm often contains a retrieving module and a reading module. Given an input question, the reading module predicts the answer from the relevant passages which are retrieved by the retriever. The recent proposed Fusion-in-Decoder (FiD), which is built on top of the pretrained generative model T5, achieves the state-of-the-art performance in the reading module. Although being effective, it remains constrained by inefficient attention on all retrieved passages which contain a lot of noise. In this work, we propose a novel method KG-FiD, which filters noisy passages by leveraging the structural relationship among the retrieved passages with a knowledge graph. We initiate the passage node embedding from the FiD encoder and then use graph neural network (GNN) to update the representation for reranking. To improve the efficiency, we build the GNN on top of the intermediate layer output of the FiD encoder and only pass a few top reranked passages into the higher layers of encoder and decoder for answer generation. We also apply the proposed GNN based reranking method to enhance the passage retrieval results in the retrieving module. Extensive experiments on common ODQA benchmark datasets (Natural Question and TriviaQA) demonstrate that KG-FiD can improve vanilla FiD by up to 1.5% on answer exact match score and achieve comparable performance with FiD with only 40% of computation cost.
Current state-of-the-art approaches for image captioning typically adopt an autoregressive manner, i.e., generating descriptions word by word, which suffers from slow decoding issue and becomes a bottleneck in real-time applications. Non-autoregressive image captioning with continuous iterative refinement, which eliminates the sequential dependence in a sentence generation, can achieve comparable performance to the autoregressive counterparts with a considerable acceleration. Nevertheless, based on a well-designed experiment, we empirically proved that iteration times can be effectively reduced when providing sufficient prior knowledge for the language decoder. Towards that end, we propose a novel two-stage framework, referred to as Semi-Autoregressive Image Captioning (SAIC), to make a better trade-off between performance and speed. The proposed SAIC model maintains autoregressive property in global but relieves it in local. Specifically, SAIC model first jumpily generates an intermittent sequence in an autoregressive manner, that is, it predicts the first word in every word group in order. Then, with the help of the partially deterministic prior information and image features, SAIC model non-autoregressively fills all the skipped words with one iteration. Experimental results on the MS COCO benchmark demonstrate that our SAIC model outperforms the preceding non-autoregressive image captioning models while obtaining a competitive inference speedup. Code is available at https://github.com/feizc/SAIC.
This report describes our solution to the VALUE Challenge 2021 in the captioning task. Our solution, named CLIP4Caption++, is built on X-Linear/X-Transformer, which is an advanced model with encoder-decoder architecture. We make the following improvements on the proposed CLIP4Caption++: We employ an advanced encoder-decoder model architecture X-Transformer as our main framework and make the following improvements: 1) we utilize three strong pre-trained CLIP models to extract the text-related appearance visual features. 2) we adopt the TSN sampling strategy for data enhancement. 3) we involve the video subtitle information to provide richer semantic information. 3) we introduce the subtitle information, which fuses with the visual features as guidance. 4) we design word-level and sentence-level ensemble strategies. Our proposed method achieves 86.5, 148.4, 64.5 CIDEr scores on VATEX, YC2C, and TVC datasets, respectively, which shows the superior performance of our proposed CLIP4Caption++ on all three datasets.
Image-to-image (I2I) translation is usually carried out among discrete domains. However, image domains, often corresponding to a physical value, are usually continuous. In other words, images gradually change with the value, and there exists no obvious gap between different domains. This paper intends to build the model for I2I translation among continuous varying domains. We first divide the whole domain coverage into discrete intervals, and explicitly model the latent style code for the center of each interval. To deal with continuous translation, we design the editing modules, changing the latent style code along two directions. These editing modules help to constrain the codes for domain centers during training, so that the model can better understand the relation among them. To have diverse results, the latent style code is further diversified with either the random noise or features from the reference image, giving the individual style code to the decoder for label-based or reference-based synthesis. Extensive experiments on age and viewing angle translation show that the proposed method can achieve high-quality results, and it is also flexible for users.
The visual system processes a scene using a sequence of selective glimpses, each driven by spatial and object-based attention. These glimpses reflect what is relevant to the ongoing task and are selected through recurrent processing and recognition of the objects in the scene. In contrast, most models treat attention selection and recognition as separate stages in a feedforward process. Here we show that using capsule networks to create an object-centric hidden representation in an encoder-decoder model with iterative glimpse attention yields effective integration of attention and recognition. We evaluate our model on three multi-object recognition tasks; highly overlapping digits, digits among distracting clutter and house numbers, and show that it learns to effectively move its glimpse window, recognize and reconstruct the objects, all with only the classification as supervision. Our work takes a step toward a general architecture for how to integrate recurrent object-centric representation into the planning of attentional glimpses.
This paper presents 6D-ViT, a transformer-based instance representation learning network, which is suitable for highly accurate category-level object pose estimation on RGB-D images. Specifically, a novel two-stream encoder-decoder framework is dedicated to exploring complex and powerful instance representations from RGB images, point clouds and categorical shape priors. For this purpose, the whole framework consists of two main branches, named Pixelformer and Pointformer. The Pixelformer contains a pyramid transformer encoder with an all-MLP decoder to extract pixelwise appearance representations from RGB images, while the Pointformer relies on a cascaded transformer encoder and an all-MLP decoder to acquire the pointwise geometric characteristics from point clouds. Then, dense instance representations (i.e., correspondence matrix, deformation field) are obtained from a multi-source aggregation network with shape priors, appearance and geometric information as input. Finally, the instance 6D pose is computed by leveraging the correspondence among dense representations, shape priors, and the instance point clouds. Extensive experiments on both synthetic and real-world datasets demonstrate that the proposed 3D instance representation learning framework achieves state-of-the-art performance on both datasets, and significantly outperforms all existing methods.
Sequence expansion between encoder and decoder is a critical challenge in sequence-to-sequence TTS. Attention-based methods achieve great naturalness but suffer from unstable issues like missing and repeating phonemes, not to mention accurate duration control. Duration-informed methods, on the contrary, seem to easily adjust phoneme duration but show obvious degradation in speech naturalness. This paper proposes PAMA-TTS to address the problem. It takes the advantage of both flexible attention and explicit duration models. Based on the monotonic attention mechanism, PAMA-TTS also leverages token duration and relative position of a frame, especially countdown information, i.e. in how many future frames the present phoneme will end. They help the attention to move forward along the token sequence in a soft but reliable control. Experimental results prove that PAMA-TTS achieves the highest naturalness, while has on-par or even better duration controllability than the duration-informed model.
Suffering from limited singing voice corpus, existing singing voice synthesis (SVS) methods that build encoder-decoder neural networks to directly generate spectrogram could lead to out-of-tune issues during the inference phase. To attenuate these issues, this paper presents a novel acoustic model with independent pitch encoder and phoneme encoder, which disentangles the phoneme and pitch information from music score to fully utilize the corpus. Specifically, according to equal temperament theory, the pitch encoder is constrained by a pitch metric loss that maps distances between adjacent input pitches into corresponding frequency multiples between the encoder outputs. For the phoneme encoder, based on the analysis that same phonemes corresponding to varying pitches can produce similar pronunciations, this encoder is followed by an adversarially trained pitch classifier to enforce the identical phonemes with different pitches mapping into the same phoneme feature space. By these means, the sparse phonemes and pitches in original input spaces can be transformed into more compact feature spaces respectively, where the same elements cluster closely and cooperate mutually to enhance synthesis quality. Then, the outputs of the two encoders are summed together to pass through the following decoder in the acoustic model. Experimental results indicate that the proposed approaches can characterize intrinsic structure between pitch inputs to obtain better pitch synthesis accuracy and achieve superior singing synthesis performance against the advanced baseline system.