Labels are essential for training Deep Neural Networks (DNNs), guiding learning with fundamental ground truth. Label quality directly impacts DNN performance and generalization with accurate labels fostering robust predictions. Noisy labels introduce errors and hinder learning, affecting performance adversely. High-quality labels aid convergence, optimizing DNN training towards accurate data distribution representation. Ensuring label accuracy is vital for DNNs’ effective learning, generalization, and real-world performance. Undoubtedly, ensuring the quality of labels is not only critical but also demanding, often entailing considerable resources in terms of time and cost. As the scale of datasets grows, methods such as crowdsourcing have gained traction to expedite the labeling process. However, this approach comes with its own set of challenges, most notably the inherent susceptibility to errors and inaccuracies. For example, it was observed that the accuracy of AlexNet in classifying CIFAR-10 images plummeted from 77% to a mere 10% when labels were subjected to random flips. This stark drop in accuracy exemplifies the magnitude of influence that corrupted or erroneous labels can exert on the performance of DNNs. Such instances underscore the critical relationship between accurate labels and the efficacy of DNNs in understanding and effectively leveraging data. EnsuringDNNrobustness is vital, involving strategies like noise label identification, filtering, and integrating noise patterns into training for resilientmodels. Architectural and loss function design also combats label-related challenges, enhancing DNN adaptability across applications. This thesis investigates the pivotal role of labels in DNN training and their quality impact onmodel performance. Strategies spanning noise recovery, robust learning frameworks, andmulti-label solutions contribute toDNNresilience against noisy labels, advancing both understanding and practical applications. Chapter 1 of this thesis introduces and explains the crucial elements involved in training DNNs, which include data, DNN models, and expert participation. It highlights the complexity introduced by label noise and sets the stage for the diverse methods designed in subsequent chapters to address these aspects comprehensively. @en

Learning robust deep models against noisy labels becomes ever critical when today's data is commonly collected from open platforms and subject to adversarial corruption. The information on the label corruption process, i.e., corruption matrix, can greatly enhance the robustness of deep models but still fall behind in combating hard classes. In this paper, we propose to construct a golden symmetric loss (GSL) based on the estimated corruption matrix as to avoid overfitting to noisy labels and learn effectively from hard classes. GSL is the weighted sum of the corrected regular cross entropy and reverse cross entropy. By leveraging a small fraction of trusted clean data, we estimate the corruption matrix and use it to correct the loss as well as to determine the weights of GSL. We theoretically prove the robustness of the proposed loss function in the presence of dirty labels. We provide a heuristics to adaptively tune the loss weights of GSL according to the noise rate and diversity measured from the dataset. We evaluate our proposed golden symmetric loss on both vision and natural language deep models subject to different types of label noise patterns. Empirical results show that GSL can significantly outperform the existing robust training methods on different noise patterns, showing accuracy improvement up to 18% on CIFAR-100 and 1% on real world noisy dataset of Clothing1M.

@en

Big Data systems allow collecting massive datasets to feed the data hungry deep learning. Labelling these ever-bigger datasets is increasingly challenging and label errors affect even highly curated sets. This makes robustness to label noise a critical property for weakly-supervised classifiers. The related works on resilient deep networks tend to focus on a limited set of synthetic noise patterns, and with disparate views on their impacts, e.g., robustness against symmetric v.s. asymmetric noise patterns. In this paper, we first extend the theoretical analysis of test accuracy for any given noise patterns. Based on the insights, we design TrustNet that first learns the pattern of noise corruption, being it both symmetric or asymmetric, from a small set of trusted data. Then, TrustNet is trained via a robust loss function, which weights the given labels against the inferred labels from the learned noise pattern. The weight is adjusted based on model uncertainty across training epochs. We evaluate TrustNet on synthetic label noise for CIFAR-10, CIFAR-100 and big real-world data with label noise, i.e., Clothing1M. We compare against state-of-The-Art methods demonstrating the strong robustness of TrustNet under a diverse set of noise patterns.

@en

Noisy labeled data is more a norm than a rarity for crowd sourced contents. It is effective to distill noise and infer correct labels through aggregating results from crowd workers. To ensure the time relevance and overcome slow responses of workers, online label aggregation is increasingly requested, calling for solutions that can incrementally infer true label distribution via subsets of data items. In this paper, we propose a novel online label aggregation framework, BiLA , which employs variational Bayesian inference method and designs a novel stochastic optimization scheme for incremental training. BiLA is flexible to accommodate any generating distribution of labels by the exact computation of its posterior distribution. We also derive the convergence bound of the proposed optimizer. We compare BiLA with the state of the art based on minimax entropy, neural networks and expectation maximization algorithms, on synthetic and real-world data sets. Our evaluation results on various online scenarios show that BiLA can effectively infer the true labels, with an error rate reduction of at least 10 to 1.5 percent points for synthetic and real-world datasets, respectively.

@en

Deep neural networks (DNNs) are becoming the core components of many applications running on edge devices, especially for real time image-based analysis. Increasingly, multi-faced knowledge is extracted via executing multiple DNNs inference models, e.g., identifying objects, faces, and genders from images. The response times of multi-DNN highly affect users' quality of experience and safety as well. Different DNNs exhibit diversified resource requirements and execution patterns across layers and networks, which may easily exceed the available device memory and riskily degrade the responsiveness. In this paper, we design and implement Masa, a responsive memory-aware multi-DNN execution framework, an on-device middleware featuring on modeling inter- and intra-network dependency and leveraging complimentary memory usage of each layer. Masa can consistently ensure the average response time when deterministically and stochastically executing multiple DNN-based image analyses. We extensively evaluate Masa on three configurations of Raspberry Pi and a large set of popular DNN models triggered by different generation patterns of images. Our evaluation results show that Masa can achieve lower average response times by up to 90% on devices with small memory, i.e., 512 MB to 1 GB, compared to the state of the art multi-DNN scheduling solutions.

@en

This artifact is a guideline how the Edgecaffe framework, presented in [1], can be used. Edgecaffe is an open-source Deep Neural Network framework for efficient multi-network inference on edge devices. This framework enables the layer by layer execution and fine-grained control during inference of Deep Neural Networks. Edgecaffe is created to give more fine grained-control over the execution during inference than offered by the original code of Caffe [2]. Edgecaffe made it possible for Masa to outperform Deepeye [3] and normal bulk execution. Besides the core implementation of Edgecaffe, the repository holds additional tools, Queue Runner and ModelSplitter, that make more convenient to run experiments and prepare newly trained networks

@en

Today's available datasets in the wild, e.g., from social media and open platforms, present tremendous opportunities and challenges for deep learning, as there is a significant portion of tagged images, but often with noisy, i.e. erroneous, labels. Recent studies improve the robustness of deep models against noisy labels without the knowledge of true labels. In this paper, we advocate to derive a stronger classifier which proactively makes use of the noisy labels in addition to the original images - turning noisy labels into learning features. To such an end, we propose a novel framework, LABELNET, composed of Amateur and Expert, which iteratively learn from each other. Amateur is a regular image classifier trained by the feedback of Expert, which imitates how human experts would correct the predicted labels from Amateur using the noise pattern learnt from the knowledge of both the noisy and ground truth labels. The trained Amateur and Expert proactively leverage the images and their noisy labels to infer image classes. Our empirical evaluations on noisy versions of MNIST, CIFAR-10, CIFAR-100 and real-world data of Clothing1M show that the proposed model can achieve robust classification against a wide range of noise ratios and with as little as 20-50% training data, compared to state-of-the-art deep models that solely focus on distilling the impact of noisy labels.

@en

The execution of deep neural network (DNN) inference jobs on edge devices has become increasingly popular. Multiple of such inference models can concurrently analyse the on-device data, e.g. images, to extract valuable insights. Prior art focuses on low-power accelerators, compressed neural network architectures, and specialized frameworks to reduce execution time of single inference jobs on edge devices which are resource constrained. However, it is little known how different scheduling policies can further improve the runtime performance of multi-inference jobs without additional edge resources. To enable the exploration of scheduling policies, we first develop an execution framework, EdgeCaffe, which splits the DNN inference jobs by loading and execution of each network layer. We empirically characterize the impact of loading and scheduling policies on the execution time of multi-inference jobs and point out their dependency on the available memory space. We propose a novel memory-aware scheduling policy, MemA, which opportunistically interleaves the executions of different types of DNN layers based on their estimated run-time memory demands. Our evaluation on exhaustive combinations of five networks, data inputs, and memory configurations show that MemA can alleviate the degradation of execution times of multi-inference (up to 5*) under severely constrained memory compared to standard scheduling policies without affecting accuracy.

@en

Data is generated with unprecedented speed, due to the flourishing of social media and open platforms. However, due to the lack of scrutinizing, both clean and dirty data are widely spreaded. For instance, there is a significant portion of images tagged with corrupted dirty class labels. Such dirty data sets are not only detrimental to the learning outcomes, e.g., misclassified images into the wrong classes, but also costly. It is pointed out that bad data can cost the U.S. up to a daunting 3 trillion dollars per year. In this paper, we address the following question: how prevailing (deep) machine learning models can be robustly trained given a non-negligible presence of corrupted labeled data. Dirty labels significantly increase the complexity of existing learning problems, as the ground truth of label's quality are not easily assessed. Here, we advocate to rigorously incorporate human experts into one learning framework where both artificial and human intelligence collaborate. To such an end, we combine three strategies to enhance the robustness for deep and regular machine learning algorithms, namely, (i) data filtering through additional quality model, (ii) data selection via actively learning from expert, and (iii) imitating expert's correction process. We demonstrate three strategies sequentially with examples and apply them on widely used benchmarks, such as CIFAR10 and CIFAR100. Our initial results show the effectiveness of the proposed strategies in combating dirty labels, e.g., the resulting classification can be up to 50% higher than the state-of-the-art AI-only solutions. Finally, we extend the discussion of robust learning from the trusted data to the trusted execution environment.

@en