Deep neural networks with short residual connections have demonstrated remarkable success across domains, but increasing depth often introduces computational redundancy without corresponding improvements in representation quality. We introduce Auto-Compressing Networks (ACNs), an architectural variant where additive long feedforward connections from each layer to the output replace traditional short residual connections. By analyzing the distinct dynamics induced by this modification, we reveal a unique property we coin as *auto-compression*—the ability of a network to organically compress information during training with gradient descent, through architectural design alone. Through auto-compression, information is dynamically "pushed" into early layers during training, enhancing their representational quality and revealing potential redundancy in deeper ones. We theoretically show that this property emerges from layer-wise training patterns found only in ACNs, where layers are dynamically utilized during training based on task requirements. We also find that ACNs exhibit enhanced noise robustness compared to residual networks, superior performance in low-data settings, improved transfer learning capabilities, and mitigate catastrophic forgetting suggesting that they learn representations that generalize better despite using fewer parameters. Our results demonstrate up to 18\% reduction in catastrophic forgetting and 30-80\% architectural compression while maintaining accuracy across vision transformers, MLP-mixers, and BERT architectures. These findings establish ACNs as a practical approach to developing efficient neural architectures that automatically adapt their computational footprint to task complexity, while learning robust representations suitable for noisy real-world tasks and continual learning scenarios.

In this work, we investigate the personalization of text-to-music diffusion models in a few-shot setting. Motivated by recent advances in the computer vision domain, we are the first to explore the combination of pre-trained text-to-audio diffusers with two established personalization methods. We experiment with the effect of audio-specific data augmentation on the overall system performance and assess different training strategies. For evaluation, we construct a novel dataset with prompts and music clips. We consider both embedding-based and music-specific metrics for quantitative evaluation, as well as a user study for qualitative evaluation. Our analysis shows that similarity metrics are in accordance with user preferences and that current personalization approaches tend to learn rhythmic music concepts more easily than melody. The code, dataset, and example material of this study are open to the research community1.

Modern speech recognition systems exhibit rapid performance degradation under domain shift. This issue is especially prevalent in data-scarce settings, such as low-resource languages, where the diversity of training data is limited. In this work, we propose M2DS2, a simple and sample-efficient fine-tuning strategy for large pre-trained speech models, based on mixed source and target domain self-supervision. We find that including source domain self-supervision stabilizes training and avoids mode collapse of the latent representations. For evaluation, we collect HParl, a 120-hour speech corpus for Greek, consisting of plenary sessions in the Greek Parliament. We merge HParl with two popular Greek corpora to create GREC-MD, a test-bed for multi-domain evaluation of Greek ASR systems. In our experiments, we find that, while other Unsupervised Domain Adaptation baselines fail in this resource-constrained environment, M2DS2 yields significant improvements for cross-domain adaptation, even when only a few hours of in-domain audio are available. When we relax the problem in a weakly supervised setting, we find that independent adaptation for audio using M2DS2 and language using simple LM augmentation techniques is particularly effective, yielding word error rates comparable to the fully supervised baselines.

Multimodal learning pipelines have benefited from the success of pretrained language models. However, this comes at the cost of increased model parameters. In this work, we propose Adapted Multimodal BERT (AMB), a BERT-based architecture for multimodal tasks that uses a combination of adapter modules and intermediate fusion layers. The adapter adjusts the pretrained language model for the task at hand, while the fusion layers perform task-specific, layer-wise fusion of audio-visual information with textual BERT representations. During the adaptation process the pre-trained language model parameters remain frozen, allowing for fast, parameter-efficient training. In our ablations we see that this approach leads to efficient models, that can outperform their fine-tuned counterparts and are robust to input noise. Our experiments on sentiment analysis with CMU-MOSEI show that AMB outperforms the current state-of-the-art across metrics, with 3.4% relative reduction in the resulting error and 2.1% relative improvement in 7−class classification accuracy.

Combining simple concepts to form structured thoughts and decomposing complex concepts into their constituents is one key characteristic of human cognition. In this work we extract video representations by combining multi-scale processing with compositional constraints, i.e., we constrain the latent space created by the network so that coarse grained video features are composed from a set of fine-grained video features using simple functions. We integrate the proposed constraints in a state-of-the-art contrastive learning frame-work. In our ablations, we evaluate different formulations of the compositional constraints and composition functions. We evaluate the proposed approach for the downstream tasks of action detection in UCF-101, and video summarization in the SumMe dataset. We achieve significant improvements over the baseline, i.e., 3.9% and 6.3% relative improvements for UCF-101 and SumMe respectively, showcasing the importance of compositional video representations.

In this paper, we propose a new deep unfolding neural network based on the ADMM algorithm for analysis Compressed Sensing. The proposed network jointly learns a redundant analysis operator for sparsification and reconstructs the signal of interest. We compare our proposed network with a state-of-the-art unfolded ISTA decoder, that also learns an orthogonal sparsifier. Moreover, we consider not only image, but also speech datasets as test examples. Computational experiments demonstrate that our proposed network outperforms the state-of-the-art deep unfolding network, consistently for both real-world image and speech datasets.

Current deep learning approaches for multimodal fusion rely on bottom-up fusion of high and mid-level latent modality representations (late/mid fusion) or low level sensory inputs (early fusion). Models of human perception highlight the importance of top-down fusion, where high-level representations affect the way sensory inputs are perceived, i.e. cognition affects perception. These top-down interactions are not captured in current deep learning models. In this work we propose a neural architecture that captures top-down cross-modal interactions, using a feedback mechanism in the forward pass during network training. The proposed mechanism extracts high-level representations for each modality and uses these representations to mask the sensory inputs, allowing the model to perform top-down feature masking. We apply the proposed model for multimodal sentiment recognition on CMU-MOSEI. Our method shows consistent improvements over the well established MulT and over our strong late fusion baseline, achieving state-of-the-art results.

We propose a novel deep architecture for the task of reasoning about social interactions in videos. We leverage the multi-step reasoning capabilities of Compositional Attention Networks (MAC), and propose a multimodal extension (MAC-X). MAC-X is based on a recurrent cell that performs iterative mid-level fusion of input modalities (visual, auditory, text) over multiple reasoning steps, by use of a temporal attention mechanism. We then combine MAC-X with LSTMs for temporal input processing in an end-to-end architecture. Our ablation studies show that the proposed MAC-X architecture can effectively leverage multimodal input cues using mid-level fusion mechanisms. We apply MAC-X to the task of Social Video Question Answering in the Social IQ dataset and obtain a 2.5% absolute improvement in terms of binary accuracy over the current state-of-the-art.

Deep learning Text-to-Speech (TTS) systems have achieved impressive generated speech quality, close to human parity. However, they suffer from training stability issues and in-correct alignment between the intermediate acoustic representation and the text input. In this work, we propose Regotron, a regularized Tacotron2 version which alleviates the training issues by augmenting the objective function with an additional term, which penalizes non-monotonic alignments in the location-sensitive attention mechanism. By introducing this regularization term we demonstrate its effectiveness to stabilize the training process, produce a monotonic attention quicker (13% of the total number of epochs compared to Tacotron2) and reduce the alignment errors during inference. Moreover, Regotron has minimal additional computational overhead, reduces common TTS mistakes and at the same time achieves improved speech naturalness according to subjective mean opinion scores (MOS) collected from 50 evaluators.

Constantinos Karouzos, Georgios Paraskevopoulos, Alexandros Potamianos. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

This paper presents an audio visual automatic speech recognition (AV-ASR) system using a Transformer-based architecture.We particularly focus on the scene context provided by the visual information, to ground the ASR.We extract representations for audio features in the encoder layers of the transformer and fuse video features using an additional crossmodal multihead attention layer.Additionally, we incorporate a multitask training criterion for multiresolution ASR, where we train the model to generate both character and subword level transcriptions.Experimental results on the How2 dataset, indicate that multiresolution training can speed up convergence by around 50% and relatively improves word error rate (WER) performance by upto 18% over subword prediction models.Further, incorporating visual information improves performance with relative gains upto 3.76% over audio only models.Our results are comparable to state-of-the-art Listen, Attend and Spell-based architectures.

In this work we propose a machine learning model for depression detection from transcribed clinical interviews. Depression is a mental disorder that impacts not only the subject's mood but also the use of language. To this end we use a Hierarchical Attention Network to classify interviews of depressed subjects. We augment the attention layer of our model with a conditioning mechanism on linguistic features, extracted from affective lexica. Our analysis shows that individuals diagnosed with depression use affective language to a greater extent than not-depressed. Our experiments show that external affective information improves the performance of the proposed architecture in the General Psychotherapy Corpus and the DAIC-WoZ 2017 depression datasets, achieving state-of-the-art 71.6 and 68.6 F1 scores respectively.

We examine the use of linear and non-linear dimensionality reduction algorithms for extracting low-rank feature representations for speech emotion recognition. Two feature sets are used, one based on low-level descriptors and their aggregations (IS10) and one modeling recurrence dynamics of speech (RQA), as well as their fusion. We report speech emotion recognition (SER) results for learned representations on two databases using different classification methods. Classification with low-dimensional representations yields performance improvement in a variety of settings. This indicates that dimensionality reduction is an effective way to combat the curse of dimensionality for SER. Visualization of features in two dimensions provides insight into discriminatory abilities of reduced feature sets.

We investigate the performance of features that can capture nonlinear recurrence dynamics embedded in the speech signal for the task of Speech Emotion Recognition (SER). Reconstruction of the phase space of each speech frame and the computation of its respective Recurrence Plot (RP) reveals complex structures which can be measured by performing Recurrence Quantification Analysis (RQA). These measures are aggregated by using statistical functionals over segment and utterance periods. We report SER results for the proposed feature set on three databases using different classification methods. When fusing the proposed features with traditional feature sets, we show an improvement in unweighted accuracy of up to 5.7% and 10.7% on Speaker-Dependent (SD) and Speaker-Independent (SI) SER tasks, respectively, over the baseline. Following a segment-based approach we demonstrate state-of-the-art performance on IEMOCAP using a Bidirectional Recurrent Neural Network.