DC-CCL: Device-Cloud Collaborative Controlled Learning for Large Vision Models

Many large vision models have been deployed on the cloud for real-time services. Meanwhile, fresh samples are continuously generated on the served mobile device. How to leverage the device-side samples to improve the cloud-side large model becomes a practical requirement, but falls into the dilemma of no raw sample up-link and no large model down-link. Specifically, the user may opt out of sharing raw samples with the cloud due to the concern of privacy or communication overhead, while the size of some large vision models far exceeds the mobile device's runtime capacity. In this work, we propose a device-cloud collaborative controlled learning framework, called DC-CCL, enabling a cloud-side large vision model that cannot be directly deployed on the mobile device to still benefit from the device-side local samples. In particular, DC-CCL vertically splits the base model into two submodels, one large submodel for learning from the cloud-side samples and the other small submodel for learning from the device-side samples and performing device-cloud knowledge fusion. Nevertheless, on-device training of the small submodel requires the output of the cloud-side large submodel to compute the desired gradients. DC-CCL thus introduces a light-weight model to mimic the large cloud-side submodel with knowledge distillation, which can be offloaded to the mobile device to control its small submodel's optimization direction. Given the decoupling nature of two submodels in collaborative learning, DC-CCL also allows the cloud to take a pre-trained model and the mobile device to take another model with a different backbone architecture.

One-Time Model Adaptation to Heterogeneous Clients: An Intra-Client and Inter-Image Attention Design

The mainstream workflow of image recognition applications is first training one global model on the cloud for a wide range of classes and then serving numerous clients, each with heterogeneous images from a small subset of classes to be recognized. From the cloud-client discrepancies on the range of image classes, the recognition model is desired to have strong adaptiveness, intuitively by concentrating the focus on each individual client's local dynamic class subset, while incurring negligible overhead. In this work, we propose to plug a new intra-client and inter-image attention (ICIIA) module into existing backbone recognition models, requiring only one-time cloud-based training to be client-adaptive. In particular, given a target image from a certain client, ICIIA introduces multi-head self-attention to retrieve relevant images from the client's historical unlabeled images, thereby calibrating the focus and the recognition result. Further considering that ICIIA's overhead is dominated by linear projection, we propose partitioned linear projection with feature shuffling for replacement and allow increasing the number of partitions to dramatically improve efficiency without scarifying too much accuracy. We finally evaluate ICIIA using 3 different recognition tasks with 9 backbone models over 5 representative datasets. Extensive evaluation results demonstrate the effectiveness and efficiency of ICIIA. Specifically, for ImageNet-1K with the backbone models of MobileNetV3-L and Swin-B, ICIIA can improve the testing accuracy to 83.37% (+8.11%) and 88.86% (+5.28%), while adding only 1.62% and 0.02% of FLOPs, respectively.

Walle: An End-to-End, General-Purpose, and Large-Scale Production System for Device-Cloud Collaborative Machine Learning

To break the bottlenecks of mainstream cloud-based machine learning (ML) paradigm, we adopt device-cloud collaborative ML and build the first end-to-end and general-purpose system, called Walle, as the foundation. Walle consists of a deployment platform, distributing ML tasks to billion-scale devices in time; a data pipeline, efficiently preparing task input; and a compute container, providing a cross-platform and high-performance execution environment, while facilitating daily task iteration. Specifically, the compute container is based on Mobile Neural Network (MNN), a tensor compute engine along with the data processing and model execution libraries, which are exposed through a refined Python thread-level virtual machine (VM) to support diverse ML tasks and concurrent task execution. The core of MNN is the novel mechanisms of operator decomposition and semi-auto search, sharply reducing the workload in manually optimizing hundreds of operators for tens of hardware backends and further quickly identifying the best backend with runtime optimization for a computation graph. The data pipeline introduces an on-device stream processing framework to enable processing user behavior data at source. The deployment platform releases ML tasks with an efficient push-then-pull method and supports multi-granularity deployment policies. We evaluate Walle in practical e-commerce application scenarios to demonstrate its effectiveness, efficiency, and scalability. Extensive micro-benchmarks also highlight the superior performance of MNN and the Python thread-level VM. Walle has been in large-scale production use in Alibaba, while MNN has been open source with a broad impact in the community.

On-Device Learning with Cloud-Coordinated Data Augmentation for Extreme Model Personalization in Recommender Systems

Data heterogeneity is an intrinsic property of recommender systems, making models trained over the global data on the cloud, which is the mainstream in industry, non-optimal to each individual user's local data distribution. To deal with data heterogeneity, model personalization with on-device learning is a potential solution. However, on-device training using a user's small size of local samples will incur severe overfitting and undermine the model's generalization ability. In this work, we propose a new device-cloud collaborative learning framework, called CoDA, to break the dilemmas of purely cloud-based learning and on-device learning. The key principle of CoDA is to retrieve similar samples from the cloud's global pool to augment each user's local dataset to train the recommendation model. Specifically, after a coarse-grained sample matching on the cloud, a personalized sample classifier is further trained on each device for a fine-grained sample filtering, which can learn the boundary between the local data distribution and the outside data distribution. We also build an end-to-end pipeline to support the flows of data, model, computation, and control between the cloud and each device. We have deployed CoDA in a recommendation scenario of Mobile Taobao. Online A/B testing results show the remarkable performance improvement of CoDA over both cloud-based learning without model personalization and on-device training without data augmentation. Overhead testing on a real device demonstrates the computation, storage, and communication efficiency of the on-device tasks in CoDA.

Federated Submodel Averaging

We study practical data characteristics underlying federated learning, where non-i.i.d. data from clients have sparse features, and a certain client's local data normally involves only a small part of the full model, called a submodel. Due to data sparsity, the classical federated averaging (FedAvg) algorithm or its variants will be severely slowed down, because when updating the global model, each client's zero update of the full model excluding its submodel is inaccurately aggregated. Therefore, we propose federated submodel averaging (FedSubAvg), ensuring that the expectation of the global update of each model parameter is equal to the average of the local updates of the clients who involve it. We theoretically proved the convergence rate of FedSubAvg by deriving an upper bound under a new metric called the element-wise gradient norm. In particular, this new metric can characterize the convergence of federated optimization over sparse data, while the conventional metric of squared gradient norm used in FedAvg and its variants cannot. We extensively evaluated FedSubAvg over both public and industrial datasets. The evaluation results demonstrate that FedSubAvg significantly outperforms FedAvg and its variants.

Toward Understanding the Influence of Individual Clients in Federated Learning

Federated learning allows mobile clients to jointly train a global model without sending their private data to a central server. Despite that extensive works have studied the performance guarantee of the global model, it is still unclear how each individual client influences the collaborative training process. In this work, we defined a novel notion, called {\em Fed-Influence}, to quantify this influence in terms of model parameter, and proposed an effective and efficient estimation algorithm. In particular, our design satisfies several desirable properties: (1) it requires neither retraining nor retracing, adding only linear computational overhead to clients and the server; (2) it strictly maintains the tenet of federated learning, without revealing any client's local data; and (3) it works well on both convex and non-convex loss functions and does not require the final model to be optimal. Empirical results on a synthetic dataset and the FEMNIST dataset show that our estimation method can approximate Fed-Influence with small bias. Further, we demonstrated an application of client-level model debugging.

Depth estimation on embedded computers for robot swarms in forest

Robot swarms to date are not prepared for autonomous navigation such as path planning and obstacle detection in forest floor, unable to achieve low-cost. The development of depth sensing and embedded computing hardware paves the way for swarm of terrestrial robots. The goal of this research is to improve this situation by developing low cost vision system for small ground robots to rapidly perceive terrain. We develop two depth estimation models and evaluate their performance on Raspberry Pi 4 and Jetson Nano in terms of accuracy, runtime and model size of depth estimation models, as well as memory consumption, power draw, temperature, and cost of above two embedded on-board computers. Our research demonstrated that auto-encoder network deployed on Raspberry Pi 4 runs at a power consumption of 3.4 W, memory consumption of about 200 MB, and mean runtime of 13 ms. This can be to meet our requirement for low-cost swarm of robots. Moreover, our analysis also indicated multi-scale deep network performs better for predicting depth map from blurred RGB images caused by camera motion. This paper mainly describes depth estimation models trained on our own dataset recorded in forest, and their performance on embedded on-board computers.

Low-viewpoint forest depth dataset for sparse rover swarms

Rapid progress in embedded computing hardware increasingly enables on-board image processing on small robots. This development opens the path to replacing costly sensors with sophisticated computer vision techniques. A case in point is the prediction of scene depth information from a monocular camera for autonomous navigation. Motivated by the aim to develop a robot swarm suitable for sensing, monitoring, and search applications in forests, we have collected a set of RGB images and corresponding depth maps. Over 100k images were recorded with a custom rig from the perspective of a small ground rover moving through a forest. Taken under different weather and lighting conditions, the images include scenes with grass, bushes, standing and fallen trees, tree branches, leafs, and dirt. In addition GPS, IMU, and wheel encoder data was recorded. From the calibrated, synchronized, aligned and timestamped frames about 9700 image-depth map pairs were selected for sharpness and variety. We provide this dataset to the community to fill a need identified in our own research and hope it will accelerate progress in robots navigating the challenging forest environment. This paper describes our custom hardware and methodology to collect the data, subsequent processing and quality of the data, and how to access it.

Distributed Optimization over Block-Cyclic Data

We consider practical data characteristics underlying federated learning, where unbalanced and non-i.i.d. data from clients have a block-cyclic structure: each cycle contains several blocks, and each client's training data follow block-specific and non-i.i.d. distributions. Such a data structure would introduce client and block biases during the collaborative training: the single global model would be biased towards the client or block specific data. To overcome the biases, we propose two new distributed optimization algorithms called multi-model parallel SGD (MM-PSGD) and multi-chain parallel SGD (MC-PSGD) with a convergence rate of $O(1/\sqrt{NT})$, achieving a linear speedup with respect to the total number of clients. In particular, MM-PSGD adopts the block-mixed training strategy, while MC-PSGD further adds the block-separate training strategy. Both algorithms create a specific predictor for each block by averaging and comparing the historical global models generated in this block from different cycles. We extensively evaluate our algorithms over the CIFAR-10 dataset. Evaluation results demonstrate that our algorithms significantly outperform the conventional federated averaging algorithm in terms of test accuracy, and also preserve robustness for the variance of critical parameters.

Distributed Non-Convex Optimization with Sublinear Speedup under Intermittent Client Availability

Federated learning is a new distributed machine learning framework, where a bunch of heterogeneous clients collaboratively train a model without sharing training data. In this work, we consider a practical and ubiquitous issue in federated learning: intermittent client availability, where the set of eligible clients may change during the training process. Such an intermittent client availability model would significantly deteriorate the performance of the classical Federated Averaging algorithm (FedAvg for short). We propose a simple distributed non-convex optimization algorithm, called Federated Latest Averaging (FedLaAvg for short), which leverages the latest gradients of all clients, even when the clients are not available, to jointly update the global model in each iteration. Our theoretical analysis shows that FedLaAvg attains the convergence rate of $O(1/(N^{1/4} T^{1/2}))$, achieving a sublinear speedup with respect to the total number of clients. We implement and evaluate FedLaAvg with the CIFAR-10 dataset. The evaluation results demonstrate that FedLaAvg indeed reaches a sublinear speedup and achieves 4.23% higher test accuracy than FedAvg.