Valentin Radu

Hongrui Shi, Valentin Radu, Po Yang

With the rapid expansion of edge devices, such as IoT devices, where crucial data needed for machine learning applications is generated, it becomes essential to promote their participation in privacy-preserving Federated Learning (FL) systems. The best way to achieve this desiderate is by reducing their training workload to match their constrained computational resources. While prior FL research has address the workload constrains by introducing lightweight models on the edge, limited attention has been given to optimizing on-device training efficiency through reducing the amount of data need during training. In this work, we propose FedFT-EDS, a novel approach that combines Fine-Tuning of partial client models with Entropy-based Data Selection to reduce training workloads on edge devices. By actively selecting the most informative local instances for learning, FedFT-EDS reduces training data significantly in FL and demonstrates that not all user data is equally beneficial for FL on all rounds. Our experiments on CIFAR-10 and CIFAR-100 show that FedFT-EDS uses only 50% user data while improving the global model performance compared to baseline methods, FedAvg and FedProx. Importantly, FedFT-EDS improves client learning efficiency by up to 3 times, using one third of training time on clients to achieve an equivalent performance to the baselines. This work highlights the importance of data selection in FL and presents a promising pathway to scalable and efficient Federate Learning.

Hongrui Shi, Valentin Radu

The Federated Learning (FL) workflow of training a centralized model with distributed data is growing in popularity. However, until recently, this was the realm of contributing clients with similar computing capability. The fast expanding IoT space and data being generated and processed at the edge are encouraging more effort into expanding federated learning to include heterogeneous systems. Previous approaches distribute light-weight models to clients are rely on knowledge transfer to distil the characteristic of local data in partitioned updates. However, their additional knowledge exchange transmitted through the network degrades the communication efficiency of FL. We propose to reduce the size of knowledge exchanged in these FL setups by clustering and selecting only the most representative bits of information from the clients. The partitioned global update adopted in our work splits the global deep neural network into a lower part for generic feature extraction and an upper part that is more sensitive to this selected client knowledge. Our experiments show that only 1.6% of the initially exchanged data can effectively transfer the characteristic of the client data to the global model in our FL approach, using split networks. These preliminary results evolve our understanding of federated learning by demonstrating efficient training using strategically selected training samples.

Rik Mulder, Valentin Radu, Christophe Dubach

The choice of convolutional routines (primitives) to implement neural networks has a tremendous impact on their inference performance (execution speed) on a given hardware platform. To optimise a neural network by primitive selection, the optimal primitive is identified for each layer of the network. This process requires a lengthy profiling stage, iterating over all the available primitives for each layer configuration, to measure their execution time on the target platform. Because each primitive exploits the hardware in different ways, new profiling is needed to obtain the best performance when moving to another platform. In this work, we propose to replace this prohibitively expensive profiling stage with a machine learning based approach of performance modeling. Our approach speeds up the optimisation time drastically. After training, our performance model can estimate the performance of convolutional primitives in any layer configuration. The time to optimise the execution of large neural networks via primitive selection is reduced from hours to just seconds. Our performance model is easily transferable to other target platforms. We demonstrate this by training a performance model on an Intel platform and performing transfer learning to AMD and ARM processor devices with minimal profiled samples.

Yuan Wen, Andrew Anderson, Valentin Radu et al.

Convolutional neural networks (CNNs) are used in many embedded applications, from industrial robotics and automation systems to biometric identification on mobile devices. State-of-the-art classification is typically achieved by large networks, which are prohibitively expensive to run on mobile and embedded devices with tightly constrained memory and energy budgets. We propose an approach for ahead-of-time domain specific optimization of CNN models, based on an integer linear programming (ILP) for selecting primitive operations to implement convolutional layers. We optimize the trade-off between execution time and memory consumption by: 1) attempting to minimize execution time across the whole network by selecting data layouts and primitive operations to implement each layer; and 2) allocating an appropriate workspace that reflects the upper bound of memory footprint per layer. These two optimization strategies can be used to run any CNN on any platform with a C compiler. Our evaluation with a range of popular ImageNet neural architectures (GoogleNet, AlexNet, VGG, ResNet and SqueezeNet) on the ARM Cortex-A15 yields speedups of 8x compared to a greedy algorithm based primitive selection, reduces memory requirement by 2.2x while sacrificing only 15% of inference time compared to a solver that considers inference time only. In addition, our optimization approach exposes a range of optimal points for different configurations across the Pareto frontier of memory and latency trade-off, which can be used under arbitrary system constraints.

Valentin Radu, Kuba Kaszyk, Yuan Wen et al.

Convolutional Neural Networks (CNN) are becoming a common presence in many applications and services, due to their superior recognition accuracy. They are increasingly being used on mobile devices, many times just by porting large models designed for server space, although several model compression techniques have been considered. One model compression technique intended to reduce computations is channel pruning. Mobile and embedded systems now have GPUs which are ideal for the parallel computations of neural networks and for their lower energy cost per operation. Specialized libraries perform these neural network computations through highly optimized routines. As we find in our experiments, these libraries are optimized for the most common network shapes, making uninstructed channel pruning inefficient. We evaluate higher level libraries, which analyze the input characteristics of a convolutional layer, based on which they produce optimized OpenCL (Arm Compute Library and TVM) and CUDA (cuDNN) code. However, in reality, these characteristics and subsequent choices intended for optimization can have the opposite effect. We show that a reduction in the number of convolutional channels, pruning 12% of the initial size, is in some cases detrimental to performance, leading to 2x slowdown. On the other hand, we also find examples where performance-aware pruning achieves the intended results, with performance speedups of 3x with cuDNN and above 10x with Arm Compute Library and TVM. Our findings expose the need for hardware-instructed neural network pruning.

Jack Turner, José Cano, Valentin Radu et al.

Convolutional Neural Networks (CNNs) are extremely computationally demanding, presenting a large barrier to their deployment on resource-constrained devices. Since such systems are where some of their most useful applications lie (e.g. obstacle detection for mobile robots, vision-based medical assistive technology), significant bodies of work from both machine learning and systems communities have attempted to provide optimisations that will make CNNs available to edge devices. In this paper we unify the two viewpoints in a Deep Learning Inference Stack and take an across-stack approach by implementing and evaluating the most common neural network compression techniques (weight pruning, channel pruning, and quantisation) and optimising their parallel execution with a range of programming approaches (OpenMP, OpenCL) and hardware architectures (CPU, GPU). We provide comprehensive Pareto curves to instruct trade-offs under constraints of accuracy, execution time, and memory space.