Search for: All records

Deep neural networks are increasingly required to operate across diverse hardware platforms, latency constraints, and power budgets, which motivates the need for specialized models for each scenario. However, designing and training a separate model per scenario or serving a large ensemble of models is often impractical. Weight sharing has emerged as a promising paradigm to address this challenge by training a single ''SuperNet'' that subsumes many sub-models (SubNets), and by reusing weights across those SubNets both at training and inference time. This paper provides an abridged survey of our recent advances that leverage weight sharing for efficient AI, covering both training and inference serving. In centralized once-for-all training, Delayed ε-Shrinking (DεS) improves training efficiency by strategically scheduling the introduction of smaller SubNets during training. In a federated fashion, SuperFedNas co-trains a SuperNet across distributed clients and disjoins training and searching, which enables oneshot specialization to many deployment targets at minimal cost. ∇QDARTS integrates quantization into differentiable architecture search, jointly finding neural architectures, weights, and low-precision settings to yield highly efficient models in a single search. For inference serving, SuperServe introduces a weight-shared model with dynamic SubNet routing (SubNetAct) to instantaneously switch among a spectrum of accuracy-latency operating points, coupled with a scheduler (SlackFit) for unpredictable workloads. Finally, SUSHI co-designs model, system, and accelerator to exploit weightshared SuperNets on tinyML devices, caching SubGraphs on FPGA to reduce latency and energy. Together, these works demonstrate that the weight sharing paradigm can dramatically improve the efficiency of both training and inference serving of deep models across a range of scenarios. 

The increasing deployment of ML models on the critical path of production applications requires ML inference serving systems to serve these models under unpredictable and bursty request arrival rates. Serving many models under such conditions requires a careful balance between each application's latency and accuracy requirements and the overall efficiency of utilization of scarce resources. Faced with this tension, state-of-the-art systems either choose a single model representing a static point in the latency-accuracy tradeoff space to serve all requests or incur latency target violations by loading specific models on the critical path of request serving. Our work instead resolves this tension through a resource-efficient serving of the entire range of models spanning the latency-accuracy tradeoff space. Our novel mechanism, SubNetAct, achieves this by carefully inserting specialized control-flow operators in pre-trained, weight-shared super-networks. These operators enable SubNetAct to dynamically route a request through the network to actuate a specific model that meets the request's latency and accuracy target. Thus, SubNetAct can serve a vastly higher number of models than prior systems while requiring upto 2.6\texttimes{} lower memory. More crucially, SubNetAct's near-instantaneous actuation of a wide-range of models unlocks the design space of fine-grained, reactive scheduling policies. We design one such extremely effective policy, SlackFit, and instantiate both SubNetAct and Slack-Fit in a real system, SuperServe. On real-world traces derived from a Microsoft workload, SuperServe achieves 4.67\% higher accuracy for the same latency targets and 2.85\texttimes{} higher latency target attainment for the same accuracy. 

CNNs are increasingly deployed across different hardware, dynamic environments, and low-power embedded devices. This has led to the design and training of CNN architectures with the goal of maximizing accuracy subject to such variable deployment constraints. As the number of deployment scenarios grows, there is a need to find scalable solutions to design and train specialized CNNs. Once-for-all training has emerged as a scalable approach that jointly co-trains many models (subnets) at once with a constant training cost and finds specialized CNNs later. The scalability is achieved by training the full model and simultaneously reducing it to smaller subnets that share model weights (weight-shared shrinking). However, existing once-for-all training approaches incur huge training costs reaching 1200 GPU hours. We argue this is because they either start the process of shrinking the full model too early or too late. Hence, we propose Delayed Epsilon-Shrinking (DepS) that starts the process of shrinking the full model when it is partially trained, which leads to training cost improvement and better in-place knowledge distillation to smaller models. The proposed approach also consists of novel heuristics that dynamically adjust subnet learning rates incrementally, leading to improved weight-shared knowledge distillation from larger to smaller subnets as well. As a result, DepS outperforms state-of-the-art once-for-all training techniques across different datasets including CIFAR10/100, ImageNet-100, and ImageNet-1k on accuracy and cost. It achieves higher ImageNet-1k top1 accuracy or the same accuracy with 1.3x reduction in FLOPs and 2.5x drop in training cost (GPU*hrs). 

Neural Architecture Search (NAS) for Federated Learning (FL) is an emerging field. It automates the design and training of Deep Neural Networks (DNNs) when data cannot be centralized due to privacy, communication costs, or regulatory restrictions. Recent federated NAS methods not only reduce manual effort but also help achieve higher accuracy than traditional FL methods like FedAvg. Despite the success, existing federated NAS methods still fall short in satisfying diverse deployment targets common in on-device inference including hardware, latency budgets, or variable battery levels. Most federated NAS methods search for only a limited range of neuro-architectural patterns, repeat them in a DNN, thereby restricting achievable performance. Moreover, these methods incur prohibitive training costs to satisfy deployment targets. They perform the training and search of DNN architectures repeatedly for each case. SuperFedNAS addresses these challenges by decoupling the training and search in federated NAS. SuperFedNAS co-trains a large number of diverse DNN architectures contained inside one supernet in the FL setting. Post-training, clients perform NAS locally to find specialized DNNs by extracting different parts of the trained supernet with no additional training. SuperFedNAS takes O(1) (instead of O(N)) cost to find specialized DNN architectures in FL for any N deployment targets. As part of SuperFedNAS, we introduce MaxNet—a novel FL training algorithm that performs multi-objective federated optimization of DNN architectures (≈5∗108) under different client data distributions. SuperFedNAS achieves upto 37.7\% higher accuracy or upto 8.13x reduction in MACs than existing federated NAS methods. 

Tiny machine learning (TinyML) applications increasingly operate in dynamically changing deployment scenarios, requiring optimization for both accuracy and latency. Existing methods mainly target a single point in the accuracy/latency tradeoff space, which is insufficient as no single static point can be optimal under variable conditions. We draw on a recently proposed weight-shared SuperNet mechanism to enable serving a stream of queries that activates different SubNets within a SuperNet. This creates an opportunity to exploit the inherent temporal locality of different queries that use the same SuperNet. We propose a hardware–software co-design called SUSHI that introduces a novel SubGraph Stationary optimization. SUSHI consists of a novel field-programmable gate array implementation and a software scheduler that controls which SubNets to serve and which SubGraph to cache in real time. SUSHI yields up to a 32% improvement in latency, 0.98% increase in served accuracy, and achieves up to 78.7% off-chip energy saved across several neural network architectures. 

There is a growing rise of applications that need to support a library of models with diverse latency-accuracy trade-offs on a Pareto frontier, especially in the health-care domain. This work presents an end-to-end system for training and serving weight-sharing models. On the training end, we leverage recent research in creating a family of models on the latency- accuracy Pareto frontier that share weights, reducing the total number of unique parameters. On the serving (inference end), we propose a novel accelerator FastSwitch that extracts weight reuse across different models, thereby providing fast real-time switching between different models. 

Machine Learning (ML) research has focused on maximizing the accuracy of predictive tasks. ML models, however, are increasingly more complex, resource intensive, and costlier to deploy in resource-constrained environments. These issues are exacerbated for prediction tasks with sequential classification on progressively transitioned stages with “happens-before” relation between them.We argue that it is possible to “unfold” a monolithic single multi-class classifier, typically trained for all stages using all data, into a series of single-stage classifiers. Each single- stage classifier can be cascaded gradually from cheaper to more expensive binary classifiers that are trained using only the necessary data modalities or features required for that stage. UnfoldML is a cost-aware and uncertainty-based dynamic 2D prediction pipeline for multi-stage classification that enables (1) navigation of the accuracy/cost tradeoff space, (2) reducing the spatio-temporal cost of inference by orders of magnitude, and (3) early prediction on proceeding stages. UnfoldML achieves orders of magnitude better cost in clinical settings, while detecting multi- stage disease development in real time. It achieves within 0.1% accuracy from the highest-performing multi-class baseline, while saving close to 20X on spatio- temporal cost of inference and earlier (3.5hrs) disease onset prediction. We also show that UnfoldML generalizes to image classification, where it can predict different level of labels (from coarse to fine) given different level of abstractions of a image, saving close to 5X cost with as little as 0.4% accuracy reduction.