This paper introduces a novel reference-free (RF) audio quality metric called the RF-Generative Machine Listener (RF-GML), designed to evaluate coded mono, stereo, and binaural audio at a 48 kHz sample rate. RF-GML leverages transfer learning from a state-of-the-art full-reference (FR) Generative Machine Listener (GML) with minimal architectural modifications. The term "generative" refers to the model's ability to generate an arbitrary number of simulated listening scores. Unlike existing RF models, RF-GML accurately predicts subjective quality scores across diverse content types and codecs. Extensive evaluations demonstrate its superiority in rating unencoded audio and distinguishing different levels of coding artifacts. RF-GML's performance and versatility make it a valuable tool for coded audio quality assessment and monitoring in various applications, all without the need for a reference signal.
Brain-machine interfaces (BMIs) are expanding beyond clinical settings thanks to advances in hardware and algorithms. However, they still face challenges in user-friendliness and signal variability. Classification models need periodic adaptation for real-life use, making an optimal re-training strategy essential to maximize user acceptance and maintain high performance. We propose TOR, a train-on-request workflow that enables user-specific model adaptation to novel conditions, addressing signal variability over time. Using continual learning, TOR preserves knowledge across sessions and mitigates inter-session variability. With TOR, users can refine, on demand, the model through on-device learning (ODL) to enhance accuracy adapting to changing conditions. We evaluate the proposed methodology on a motor-movement dataset recorded with a non-stigmatizing wearable BMI headband, achieving up to 92% accuracy and a re-calibration time as low as 1.6 minutes, a 46% reduction compared to a naive transfer learning workflow. We additionally demonstrate that TOR is suitable for ODL in extreme edge settings by deploying the training procedure on a RISC-V ultra-low-power SoC (GAP9), resulting in 21.6 ms of latency and 1 mJ of energy consumption per training step. To the best of our knowledge, this work is the first demonstration of an online, energy-efficient, dynamic adaptation of a BMI model to the intrinsic variability of EEG signals in real-time settings.
The training of deep and/or convolutional neural networks (DNNs/CNNs) is traditionally done on servers with powerful CPUs and GPUs. Recent efforts have emerged to localize machine learning tasks fully on the edge. This brings advantages in reduced latency and increased privacy, but necessitates working with resource-constrained devices. Approaches for inference and training in mobile and edge devices based on pruning, quantization or incremental and transfer learning require trading off accuracy. Several works have explored distributing inference operations on mobile and edge clusters instead. However, there is limited literature on distributed training on the edge. Existing approaches all require a central, potentially powerful edge or cloud server for coordination or offloading. In this paper, we describe an approach for distributed CNN training exclusively on mobile and edge devices. Our approach is beneficial for the initial CNN layers that are feature map dominated. It is based on partitioning forward inference and back-propagation operations among devices through tiling and fusing to maximize locality and expose communication and memory-aware parallelism. We also introduce the concept of layer grouping to further fine-tune performance based on computation and communication trade-off. Results show that for a cluster of 2-6 quad-core Raspberry Pi3 devices, training of an object-detection CNN provides a 2x-15x speedup with respect to a single core and up to 8x reduction in memory usage per device, all without sacrificing accuracy. Grouping offers up to 1.5x speedup depending on the reference profile and batch size.
Foundation models pre-trained using self-supervised learning have shown powerful transfer learning capabilities on various downstream tasks, including language understanding, text generation, and image recognition. The Earth observation (EO) field has produced several foundation models pre-trained directly on multispectral satellite imagery for applications like precision agriculture, wildfire and drought monitoring, and natural disaster response. However, few studies have investigated the ability of these models to generalize to new geographic locations, and potential concerns of geospatial bias -- models trained on data-rich developed nations not transferring well to data-scarce developing nations -- remain. We evaluate three popular EO foundation models, SSL4EO-S12, SatlasPretrain, and ImageNet, on five crop classification datasets across five continents. Results show that pre-trained weights designed explicitly for Sentinel-2, such as SSL4EO-S12, outperform general pre-trained weights like ImageNet. While only 100 labeled images are sufficient for achieving high overall accuracy, 900 images are required to mitigate class imbalance and improve average accuracy.
In the realm of industrial manufacturing, Artificial Intelligence (AI) is playing an increasing role, from automating existing processes to aiding in the development of new materials and techniques. However, a significant challenge arises in smaller, experimental processes characterized by limited training data availability, questioning the possibility to train AI models in such small data contexts. In this work, we explore the potential of Transfer Learning to address this challenge, specifically investigating the minimum amount of data required to develop a functional AI model. For this purpose, we consider the use case of quality control of Carbon Fiber Reinforced Polymer (CFRP) tape laying in aerospace manufacturing using optical sensors. We investigate the behavior of different open-source computer vision models with a continuous reduction of the training data. Our results show that the amount of data required to successfully train an AI model can be drastically reduced, and the use of smaller models does not necessarily lead to a loss of performance.
We present our system (denoted as T05) for the VoiceMOS Challenge (VMC) 2024. Our system was designed for the VMC 2024 Track 1, which focused on the accurate prediction of naturalness mean opinion score (MOS) for high-quality synthetic speech. In addition to a pretrained self-supervised learning (SSL)-based speech feature extractor, our system incorporates a pretrained image feature extractor to capture the difference of synthetic speech observed in speech spectrograms. We first separately train two MOS predictors that use either of an SSL-based or spectrogram-based feature. Then, we fine-tune the two predictors for better MOS prediction using the fusion of two extracted features. In the VMC 2024 Track 1, our T05 system achieved first place in 7 out of 16 evaluation metrics and second place in the remaining 9 metrics, with a significant difference compared to those ranked third and below. We also report the results of our ablation study to investigate essential factors of our system.
Over the years, Music Information Retrieval (MIR) has proposed various models pretrained on large amounts of music data. Transfer learning showcases the proven effectiveness of pretrained backend models with a broad spectrum of downstream tasks, including auto-tagging and genre classification. However, MIR papers generally do not explore the efficiency of pretrained models for Music Recommender Systems (MRS). In addition, the Recommender Systems community tends to favour traditional end-to-end neural network learning over these models. Our research addresses this gap and evaluates the applicability of six pretrained backend models (MusicFM, Music2Vec, MERT, EncodecMAE, Jukebox, and MusiCNN) in the context of MRS. We assess their performance using three recommendation models: K-nearest neighbours (KNN), shallow neural network, and BERT4Rec. Our findings suggest that pretrained audio representations exhibit significant performance variability between traditional MIR tasks and MRS, indicating that valuable aspects of musical information captured by backend models may differ depending on the task. This study establishes a foundation for further exploration of pretrained audio representations to enhance music recommendation systems.
This study investigates the efficacy of data augmentation techniques for low-resource automatic speech recognition (ASR), focusing on two endangered Austronesian languages, Amis and Seediq. Recognizing the potential of self-supervised learning (SSL) in low-resource settings, we explore the impact of data volume on the continued pre-training of SSL models. We propose a novel data-selection scheme leveraging a multilingual corpus to augment the limited target language data. This scheme utilizes a language classifier to extract utterance embeddings and employs one-class classifiers to identify utterances phonetically and phonologically proximate to the target languages. Utterances are ranked and selected based on their decision scores, ensuring the inclusion of highly relevant data in the SSL-ASR pipeline. Our experimental results demonstrate the effectiveness of this approach, yielding substantial improvements in ASR performance for both Amis and Seediq. These findings underscore the feasibility and promise of data augmentation through cross-lingual transfer learning for low-resource language ASR.
Graph domain adaptation has recently enabled knowledge transfer across different graphs. However, without the semantic information on target graphs, the performance on target graphs is still far from satisfactory. To address the issue, we study the problem of active graph domain adaptation, which selects a small quantitative of informative nodes on the target graph for extra annotation. This problem is highly challenging due to the complicated topological relationships and the distribution discrepancy across graphs. In this paper, we propose a novel approach named Dual Consistency Delving with Topological Uncertainty (DELTA) for active graph domain adaptation. Our DELTA consists of an edge-oriented graph subnetwork and a path-oriented graph subnetwork, which can explore topological semantics from complementary perspectives. In particular, our edge-oriented graph subnetwork utilizes the message passing mechanism to learn neighborhood information, while our path-oriented graph subnetwork explores high-order relationships from sub-structures. To jointly learn from two subnetworks, we roughly select informative candidate nodes with the consideration of consistency across two subnetworks. Then, we aggregate local semantics from its K-hop subgraph based on node degrees for topological uncertainty estimation. To overcome potential distribution shifts, we compare target nodes and their corresponding source nodes for discrepancy scores as an additional component for fine selection. Extensive experiments on benchmark datasets demonstrate that DELTA outperforms various state-of-the-art approaches. The code implementation of DELTA is available at https://github.com/goose315/DELTA.
We propose a method to systematically represent both the static and the dynamic components of environments, i.e. objects and agents, as well as the changes that are happening in the environment, i.e. the actions and skills performed by agents. Our approach, the Concept Hierarchy, provides the necessary information for autonomous systems to represent environment states, perform action modeling and recognition, and plan the execution of tasks. Additionally, the hierarchical structure supports generalization and knowledge transfer to environments. We rigorously define tasks, actions, skills, and affordances that enable human-understandable action and skill recognition.
Automating child speech analysis is crucial for applications such as neurocognitive assessments. Speaker diarization, which identifies ``who spoke when'', is an essential component of the automated analysis. However, publicly available child-adult speaker diarization solutions are scarce due to privacy concerns and a lack of annotated datasets, while manually annotating data for each scenario is both time-consuming and costly. To overcome these challenges, we propose a data-efficient solution by creating simulated child-adult conversations using AudioSet. We then train a Whisper Encoder-based model, achieving strong zero-shot performance on child-adult speaker diarization using real datasets. The model performance improves substantially when fine-tuned with only 30 minutes of real train data, with LoRA further improving the transfer learning performance. The source code and the child-adult speaker diarization model trained on simulated conversations are publicly available.
Human activity recognition (HAR) holds immense potential for transforming health and fitness monitoring, yet challenges persist in achieving personalized outcomes and sustainability for on-device continuous inferences. This work introduces a wrist-worn smart band designed to address these challenges through a novel combination of on-device TinyML-driven computing and cloud-enabled auto-deployment. Leveraging inertial measurement unit (IMU) sensors and a customized 1D Convolutional Neural Network (CNN) for personalized HAR, users can tailor activity classes to their unique movement styles with minimal calibration. By utilising TinyML for local computations, the smart band reduces the necessity for constant data transmission and radio communication, which in turn lowers power consumption and reduces carbon footprint. This method also enhances the privacy and security of user data by limiting its transmission. Through transfer learning and fine-tuning on user-specific data, the system achieves a 37\% increase in accuracy over generalized models in personalized settings. Evaluation using three benchmark datasets, WISDM, PAMAP2, and the BandX demonstrates its effectiveness across various activity domains. Additionally, this work presents a cloud-supported framework for the automatic deployment of TinyML models to remote wearables, enabling seamless customization and on-device inference, even with limited target data. By combining personalized HAR with sustainable strategies for on-device continuous inferences, this system represents a promising step towards fostering healthier and more sustainable societies worldwide.
In the realm of digital music, using tags to efficiently organize and retrieve music from extensive databases is crucial for music catalog owners. Human tagging by experts is labor-intensive but mostly accurate, whereas automatic tagging through supervised learning has approached satisfying accuracy but is restricted to a predefined set of training tags. Few-shot learning offers a viable solution to expand beyond this small set of predefined tags by enabling models to learn from only a few human-provided examples to understand tag meanings and subsequently apply these tags autonomously. We propose to integrate few-shot learning methodology into multi-label music auto-tagging by using features from pre-trained models as inputs to a lightweight linear classifier, also known as a linear probe. We investigate different popular pre-trained features, as well as different few-shot parametrizations with varying numbers of classes and samples per class. Our experiments demonstrate that a simple model with pre-trained features can achieve performance close to state-of-the-art models while using significantly less training data, such as 20 samples per tag. Additionally, our linear probe performs competitively with leading models when trained on the entire training dataset. The results show that this transfer learning-based few-shot approach could effectively address the issue of automatically assigning long-tail tags with only limited labeled data.
Objective: Head impact information including impact directions, speeds and force are important to study traumatic brain injury, design and evaluate protective gears. This study presents a deep learning model developed to accurately predict head impact information, including location, speed, orientation, and force, based on head kinematics during helmeted impacts. Methods: Leveraging a dataset of 16,000 simulated helmeted head impacts using the Riddell helmet finite element model, we implemented a Long Short-Term Memory (LSTM) network to process the head kinematics: tri-axial linear accelerations and angular velocities. Results: The models accurately predict the impact parameters describing impact location, direction, speed, and the impact force profile with R2 exceeding 70% for all tasks. Further validation was conducted using an on-field dataset recorded by instrumented mouthguards and videos, consisting of 79 head impacts in which the impact location can be clearly identified. The deep learning model significantly outperformed existing methods, achieving a 79.7% accuracy in identifying impact locations, compared to lower accuracies with traditional methods (the highest accuracy of existing methods is 49.4%). Conclusion: The precision underscores the model's potential in enhancing helmet design and safety in sports by providing more accurate impact data. Future studies should test the models across various helmets and sports on large in vivo datasets to validate the accuracy of the models, employing techniques like transfer learning to broaden its effectiveness.
Foundation models like ChatGPT and Sora that are trained on a huge scale of data have made a revolutionary social impact. However, it is extremely challenging for sensors in many different fields to collect similar scales of natural images to train strong foundation models. To this end, this work presents a simple and effective framework SimMAT to study an open problem: the transferability from vision foundation models trained on natural RGB images to other image modalities of different physical properties (e.g., polarization). SimMAT consists of a modality-agnostic transfer layer (MAT) and a pretrained foundation model. We apply SimMAT to a representative vision foundation model Segment Anything Model (SAM) to support any evaluated new image modality. Given the absence of relevant benchmarks, we construct a new benchmark to evaluate the transfer learning performance. Our experiments confirm the intriguing potential of transferring vision foundation models in enhancing other sensors' performance. Specifically, SimMAT can improve the segmentation performance (mIoU) from 22.15% to 53.88% on average for evaluated modalities and consistently outperforms other baselines. We hope that SimMAT can raise awareness of cross-modal transfer learning and benefit various fields for better results with vision foundation models.
Feedforward monocular face capture methods seek to reconstruct posed faces from a single image of a person. Current state of the art approaches have the ability to regress parametric 3D face models in real-time across a wide range of identities, lighting conditions and poses by leveraging large image datasets of human faces. These methods however suffer from clear limitations in that the underlying parametric face model only provides a coarse estimation of the face shape, thereby limiting their practical applicability in tasks that require precise 3D reconstruction (aging, face swapping, digital make-up, ...). In this paper, we propose a method for high-precision 3D face capture taking advantage of a collection of unconstrained videos of a subject as prior information. Our proposal builds on a two stage approach. We start with the reconstruction of a detailed 3D face avatar of the person, capturing both precise geometry and appearance from a collection of videos. We then use the encoder from a pre-trained monocular face reconstruction method, substituting its decoder with our personalized model, and proceed with transfer learning on the video collection. Using our pre-estimated image formation model, we obtain a more precise self-supervision objective, enabling improved expression and pose alignment. This results in a trained encoder capable of efficiently regressing pose and expression parameters in real-time from previously unseen images, which combined with our personalized geometry model yields more accurate and high fidelity mesh inference. Through extensive qualitative and quantitative evaluation, we showcase the superiority of our final model as compared to state-of-the-art baselines, and demonstrate its generalization ability to unseen pose, expression and lighting.
Knowledge Distillation (KD) transfers knowledge from a large pre-trained teacher network to a compact and efficient student network, making it suitable for deployment on resource-limited media terminals. However, traditional KD methods require balanced data to ensure robust training, which is often unavailable in practical applications. In such scenarios, a few head categories occupy a substantial proportion of examples. This imbalance biases the trained teacher network towards the head categories, resulting in severe performance degradation on the less represented tail categories for both the teacher and student networks. In this paper, we propose a novel framework called Knowledge Rectification Distillation (KRDistill) to address the imbalanced knowledge inherited in the teacher network through the incorporation of the balanced category priors. Furthermore, we rectify the biased predictions produced by the teacher network, particularly focusing on the tail categories. Consequently, the teacher network can provide balanced and accurate knowledge to train a reliable student network. Intensive experiments conducted on various long-tailed datasets demonstrate that our KRDistill can effectively train reliable student networks in realistic scenarios of data imbalance.
We present DreamBeast, a novel method based on score distillation sampling (SDS) for generating fantastical 3D animal assets composed of distinct parts. Existing SDS methods often struggle with this generation task due to a limited understanding of part-level semantics in text-to-image diffusion models. While recent diffusion models, such as Stable Diffusion 3, demonstrate a better part-level understanding, they are prohibitively slow and exhibit other common problems associated with single-view diffusion models. DreamBeast overcomes this limitation through a novel part-aware knowledge transfer mechanism. For each generated asset, we efficiently extract part-level knowledge from the Stable Diffusion 3 model into a 3D Part-Affinity implicit representation. This enables us to instantly generate Part-Affinity maps from arbitrary camera views, which we then use to modulate the guidance of a multi-view diffusion model during SDS to create 3D assets of fantastical animals. DreamBeast significantly enhances the quality of generated 3D creatures with user-specified part compositions while reducing computational overhead, as demonstrated by extensive quantitative and qualitative evaluations.
This paper introduces Kolmogorov-Arnold Networks (KAN) as an enhancement to the traditional linear probing method in transfer learning. Linear probing, often applied to the final layer of pre-trained models, is limited by its inability to model complex relationships in data. To address this, we propose substituting the linear probing layer with KAN, which leverages spline-based representations to approximate intricate functions. In this study, we integrate KAN with a ResNet-50 model pre-trained on ImageNet and evaluate its performance on the CIFAR-10 dataset. We perform a systematic hyperparameter search, focusing on grid size and spline degree (k), to optimize KAN's flexibility and accuracy. Our results demonstrate that KAN consistently outperforms traditional linear probing, achieving significant improvements in accuracy and generalization across a range of configurations. These findings indicate that KAN offers a more powerful and adaptable alternative to conventional linear probing techniques in transfer learning.
The field of Computer Vision (CV) has faced challenges. Initially, it relied on handcrafted features and rule-based algorithms, resulting in limited accuracy. The introduction of machine learning (ML) has brought progress, particularly Transfer Learning (TL), which addresses various CV problems by reusing pre-trained models. TL requires less data and computing while delivering nearly equal accuracy, making it a prominent technique in the CV landscape. Our research focuses on TL development and how CV applications use it to solve real-world problems. We discuss recent developments, limitations, and opportunities.