Exploring System Performance of Continual Learning for Mobile and Embedded Sensing Applications

Continual learning approaches help deep neural network models adapt and learn incrementally by trying to solve catastrophic forgetting. However, whether these existing approaches, applied traditionally to image-based tasks, work with the same efficacy to the sequential time series data generated by mobile or embedded sensing systems remains an unanswered question. To address this void, we conduct the first comprehensive empirical study that quantifies the performance of three predominant continual learning schemes (i.e., regularization, replay, and replay with examples) on six datasets from three mobile and embedded sensing applications in a range of scenarios having different learning complexities. More specifically, we implement an end-to-end continual learning framework on edge devices. Then we investigate the generalizability, trade-offs between performance, storage, computational costs, and memory footprint of different continual learning methods. Our findings suggest that replay with exemplars-based schemes such as iCaRL has the best performance trade-offs, even in complex scenarios, at the expense of some storage space (few MBs) for training examples (1% to 5%). We also demonstrate for the first time that it is feasible and practical to run continual learning on-device with a limited memory budget. In particular, the latency on two types of mobile and embedded devices suggests that both incremental learning time (few seconds - 4 minutes) and training time (1 - 75 minutes) across datasets are acceptable, as training could happen on the device when the embedded device is charging thereby ensuring complete data privacy. Finally, we present some guidelines for practitioners who want to apply a continual learning paradigm for mobile sensing tasks.

Implicit Rate-Constrained Optimization of Non-decomposable Objectives

We consider a popular family of constrained optimization problems arising in machine learning that involve optimizing a non-decomposable evaluation metric with a certain thresholded form, while constraining another metric of interest. Examples of such problems include optimizing the false negative rate at a fixed false positive rate, optimizing precision at a fixed recall, optimizing the area under the precision-recall or ROC curves, etc. Our key idea is to formulate a rate-constrained optimization that expresses the threshold parameter as a function of the model parameters via the Implicit Function theorem. We show how the resulting optimization problem can be solved using standard gradient based methods. Experiments on benchmark datasets demonstrate the effectiveness of our proposed method over existing state-of-the art approaches for these problems. The code for the proposed method is available at https://github.com/google-research/google-research/tree/master/implicit_constrained_optimization .

Wavelet Design in a Learning Framework

Wavelets have proven to be highly successful in several signal and image processing applications. Wavelet design has been an active field of research for over two decades, with the problem often being approached from an analytical perspective. In this paper, we introduce a learning based approach to wavelet design. We draw a parallel between convolutional autoencoders and wavelet multiresolution approximation, and show how the learning angle provides a coherent computational framework for addressing the design problem. We aim at designing data-independent wavelets by training filterbank autoencoders, which precludes the need for customized datasets. In fact, we use high-dimensional Gaussian vectors for training filterbank autoencoders, and show that a near-zero training loss implies that the learnt filters satisfy the perfect reconstruction property with very high probability. Properties of a wavelet such as orthogonality, compact support, smoothness, symmetry, and vanishing moments can be incorporated by designing the autoencoder architecture appropriately and with a suitable regularization term added to the mean-squared error cost used in the learning process. Our approach not only recovers the well known Daubechies family of orthogonal wavelets and the Cohen-Daubechies-Feauveau family of symmetric biorthogonal wavelets, but also learns wavelets outside these families.

Retrieve in Style: Unsupervised Facial Feature Transfer and Retrieval

We present Retrieve in Style (RIS), an unsupervised framework for fine-grained facial feature transfer and retrieval on real images. Recent work shows that it is possible to learn a catalog that allows local semantic transfers of facial features on generated images by capitalizing on the disentanglement property of the StyleGAN latent space. RIS improves existing art on: 1) feature disentanglement and allows for challenging transfers (i.e., hair and pose) that were not shown possible in SoTA methods. 2) eliminating the need for per-image hyperparameter tuning, and for computing a catalog over a large batch of images. 3) enabling face retrieval using the proposed facial features (e.g., eyes), and to our best knowledge, is the first work to retrieve face images at the fine-grained level. 4) robustness and natural application to real images. Our qualitative and quantitative analyses show RIS achieves both high-fidelity feature transfers and accurate fine-grained retrievals on real images. We discuss the responsible application of RIS.

Risk score learning for COVID-19 contact tracing apps

Digital contact tracing apps for COVID-19, such as the one developed by Google and Apple, need to estimate the risk that a user was infected during a particular exposure, in order to decide whether to notify the user to take precautions, such as entering into quarantine, or requesting a test. Such risk score models contain numerous parameters that must be set by the public health authority. Although expert guidance for how to set these parameters has been provided (e.g. https://github.com/lfph/gaen-risk-scoring/blob/main/risk-scoring.md), it is natural to ask if we could do better using a data-driven approach. This can be particularly useful when the risk factors of the disease change, e.g., due to the evolution of new variants, or the adoption of vaccines. In this paper, we show that machine learning methods can be used to automatically optimize the parameters of the risk score model, provided we have access to exposure and outcome data. Although this data is already being collected in an aggregated, privacy-preserving way by several health authorities, in this paper we limit ourselves to simulated data, so that we can systematically study the different factors that affect the feasibility of the approach. In particular, we show that the parameters become harder to estimate when there is more missing data (e.g., due to infections which were not recorded by the app). Nevertheless, the learning approach outperforms a strong manually designed baseline.

Score-Based Generative Modeling through Stochastic Differential Equations

Creating noise from data is easy; creating data from noise is generative modeling. We present a stochastic differential equation (SDE) that smoothly transforms a complex data distribution to a known prior distribution by slowly injecting noise, and a corresponding reverse-time SDE that transforms the prior distribution back into the data distribution by slowly removing the noise. Crucially, the reverse-time SDE depends only on the time-dependent gradient field (a.k.a., score) of the perturbed data distribution. By leveraging advances in score-based generative modeling, we can accurately estimate these scores with neural networks, and use numerical SDE solvers to generate samples. We show that this framework encapsulates previous approaches in diffusion probabilistic modeling and score-based generative modeling, and allows for new sampling procedures. In particular, we introduce a predictor-corrector framework to correct errors in the evolution of the discretized reverse-time SDE. We also derive an equivalent neural ODE that samples from the same distribution as the SDE, which enables exact likelihood computation, and improved sampling efficiency. In addition, our framework enables conditional generation with an unconditional model, as we demonstrate with experiments on class-conditional generation, image inpainting, and colorization. Combined with multiple architectural improvements, we achieve record-breaking performance for unconditional image generation on CIFAR-10 with an Inception score of 9.89 and FID of 2.20, a competitive likelihood of 3.10 bits/dim, and demonstrate high fidelity generation of $1024 \times 1024$ images for the first time from a score-based generative model.

Adaptive Contention Window Design using Deep Q-learning

We study the problem of adaptive contention window (CW) design for random-access wireless networks. More precisely, our goal is to design an intelligent node that can dynamically adapt its minimum CW (MCW) parameter to maximize a network-level utility knowing neither the MCWs of other nodes nor how these change over time. To achieve this goal, we adopt a reinforcement learning (RL) framework where we circumvent the lack of system knowledge with local channel observations and we reward actions that lead to high utilities. To efficiently learn these preferred actions, we follow a deep Q-learning approach, where the Q-value function is parametrized using a multi-layer perception. In particular, we implement a rainbow agent, which incorporates several empirical improvements over the basic deep Q-network. Numerical experiments based on the NS3 simulator reveal that the proposed RL agent performs close to optimal and markedly improves upon existing learning and non-learning based alternatives.

Few-Shot Adaptation of Generative Adversarial Networks

Generative Adversarial Networks (GANs) have shown remarkable performance in image synthesis tasks, but typically require a large number of training samples to achieve high-quality synthesis. This paper proposes a simple and effective method, Few-Shot GAN (FSGAN), for adapting GANs in few-shot settings (less than 100 images). FSGAN repurposes component analysis techniques and learns to adapt the singular values of the pre-trained weights while freezing the corresponding singular vectors. This provides a highly expressive parameter space for adaptation while constraining changes to the pretrained weights. We validate our method in a challenging few-shot setting of 5-100 images in the target domain. We show that our method has significant visual quality gains compared with existing GAN adaptation methods. We report qualitative and quantitative results showing the effectiveness of our method. We additionally highlight a problem for few-shot synthesis in the standard quantitative metric used by data-efficient image synthesis works. Code and additional results are available at http://e-271.github.io/few-shot-gan.