keyword
deep fusion
Deep fusion is a neural network design strategy where intermediate representations or hidden features from different modalities or network components are integrated within the deeper layers of an architecture rather than at the raw input or final output stages. By combining feature representations inside the network, deep fusion enables subsequent neural layers to jointly model complex, non-linear relationships and semantic interactions across diverse information sources, such as text and imagery. This intermediate approach differs from early fusion, which combines raw input data before feature extraction, and late fusion, which only merges the final predictions of independent models. Through layered feature integration, deep fusion allows the entire model to be trained end-to-end, producing richer shared representations and improving performance on complex multimodal and cross-domain learning tasks.
2 items
DF-GAN: A Simple and Effective Baseline for Text-to-Image Synthesis
Ming Tao, Hao Tang, Fei Wu, Xiaoyuan Jing, Bing-Kun Bao, Changsheng Xu
Why you should read this
Proposes a simple one-stage text-to-image GAN architecture that directly synthesizes high-resolution images using deep feature fusion blocks and a matching-aware discriminator, outperforming complex multi-stage models without relying on auxiliary networks.
Synthesizing high-quality realistic images from text descriptions is a challenging task. Existing text-to-image Generative Adversarial Networks generally employ a stacked architecture as the backbone yet still remain three flaws. First, the stacked architecture introduces the entanglements between generators of different image scales. Second, existing studies prefer to apply and fix extra networks in adversarial learning for text-image semantic consistency, which limits the supervision capability of these networks. Third, the cross-modal attention-based text-image fusion that widely adopted by previous works is limited on several special image scales because of the computational cost. To these ends, we propose a simpler but more effective Deep Fusion Generative Adversarial Networks (DF-GAN). To be specific, we propose: (i) a novel one-stage text-to-image backbone that directly synthesizes high-resolution images without entanglements between different generators, (ii) a novel Target-Aware Discriminator composed of Matching-Aware Gradient Penalty and One-Way Output, which enhances the text-image semantic consistency without introducing extra networks, (iii) a novel deep text-image fusion block, which deepens the fusion process to make a full fusion between text and visual features. Compared with current state-of-the-art methods, our proposed DF-GAN is simpler but more efficient to synthesize realistic and text-matching images and achieves better performance on widely used datasets. Code is available at https://github.com/tobran/DF-GAN.
Added
2026-09-26
Improving Neural Machine Translation Models with Monolingual Data
Rico Sennrich, Barry Haddow, Alexandra Birch
Why you should read this
Introduces back-translation to train neural machine translation models on target-side monolingual data without architecture changes, substantially increasing translation accuracy across standard and low-resource benchmarks.
Neural Machine Translation (NMT) has obtained state-of-the art performance for several language pairs, while only using parallel data for training. Target-side monolingual data plays an important role in boosting fluency for phrase-based statistical machine translation, and we investigate the use of monolingual data for NMT. In contrast to previous work, which combines NMT models with separately trained language models, we note that encoder-decoder NMT architectures already have the capacity to learn the same information as a language model, and we explore strategies to train with monolingual data without changing the neural network architecture. By pairing monolingual training data with an automatic back-translation, we can treat it as additional parallel training data, and we obtain substantial improvements on the WMT 15 task English<->German (+2.8-3.7 BLEU), and for the low-resourced IWSLT 14 task Turkish->English (+2.1-3.4 BLEU), obtaining new state-of-the-art results. We also show that fine-tuning on in-domain monolingual and parallel data gives substantial improvements for the IWSLT 15 task English->German.
Added
2026-09-13