Martin Swany

Cheng Chu, Lei Jiang, Martin Swany et al.

We propose a circuit-level backdoor attack, \textit{QTrojan}, against Quantum Neural Networks (QNNs) in this paper. QTrojan is implemented by few quantum gates inserted into the variational quantum circuit of the victim QNN. QTrojan is much stealthier than a prior Data-Poisoning-based Backdoor Attack (DPBA), since it does not embed any trigger in the inputs of the victim QNN or require the access to original training datasets. Compared to a DPBA, QTrojan improves the clean data accuracy by 21\% and the attack success rate by 19.9\%.

Malintha Fernando, Ransalu Senanayake, Heeyoul Choi et al.

Autonomous mobility is emerging as a new disruptive mode of urban transportation for moving cargo and passengers. However, designing scalable autonomous fleet coordination schemes to accommodate fast-growing mobility systems is challenging primarily due to the increasing heterogeneity of the fleets, time-varying demand patterns, service area expansions, and communication limitations. We introduce the concept of partially observable advanced air mobility games to coordinate a fleet of aerial vehicles by accounting for the heterogeneity of the interacting agents and the self-interested nature inherent to commercial mobility fleets. To model the complex interactions among the agents and the observation uncertainty in the mobility networks, we propose a novel heterogeneous graph attention encoder-decoder (HetGAT Enc-Dec) neural network-based stochastic policy. We train the policy by leveraging deep multi-agent reinforcement learning, allowing decentralized decision-making for the agents using their local observations. Through extensive experimentation, we show that the learned policy generalizes to various fleet compositions, demand patterns, and observation topologies. Further, fleets operating under the HetGAT Enc-Dec policy outperform other state-of-the-art graph neural network policies by achieving the highest fleet reward and fulfillment ratios in on-demand mobility networks.

Sahil Tyagi, Martin Swany

Distributed deep learning (DDL) training systems are designed for cloud and data-center environments that assumes homogeneous compute resources, high network bandwidth, sufficient memory and storage, as well as independent and identically distributed (IID) data across all nodes. However, these assumptions don't necessarily apply on the edge, especially when training neural networks on streaming data in an online manner. Computing on the edge suffers from both systems and statistical heterogeneity. Systems heterogeneity is attributed to differences in compute resources and bandwidth specific to each device, while statistical heterogeneity comes from unbalanced and skewed data on the edge. Different streaming-rates among devices can be another source of heterogeneity when dealing with streaming data. If the streaming rate is lower than training batch-size, device needs to wait until enough samples have streamed in before performing a single iteration of stochastic gradient descent (SGD). Thus, low-volume streams act like stragglers slowing down devices with high-volume streams in synchronous training. On the other hand, data can accumulate quickly in the buffer if the streaming rate is too high and the devices can't train at line-rate. In this paper, we introduce ScaDLES to efficiently train on streaming data at the edge in an online fashion, while also addressing the challenges of limited bandwidth and training with non-IID data. We empirically show that ScaDLES converges up to 3.29 times faster compared to conventional distributed SGD.

Sahil Tyagi, Martin Swany

In distributed training, deep neural networks (DNNs) are launched over multiple workers concurrently and aggregate their local updates on each step in bulk-synchronous parallel (BSP) training. However, BSP does not linearly scale-out due to high communication cost of aggregation. To mitigate this overhead, alternatives like Federated Averaging (FedAvg) and Stale-Synchronous Parallel (SSP) either reduce synchronization frequency or eliminate it altogether, usually at the cost of lower final accuracy. In this paper, we present \texttt{SelSync}, a practical, low-overhead method for DNN training that dynamically chooses to incur or avoid communication at each step either by calling the aggregation op or applying local updates based on their significance. We propose various optimizations as part of \texttt{SelSync} to improve convergence in the context of \textit{semi-synchronous} training. Our system converges to the same or better accuracy than BSP while reducing training time by up to 14$\times$.

Sahil Tyagi, Martin Swany

Gradient compression alleviates expensive communication in distributed deep learning by sending fewer values and its corresponding indices, typically via Allgather (AG). Training with high compression ratio (CR) achieves high accuracy like DenseSGD, but has lower parallel scaling due to high communication cost (i.e., parallel efficiency). Using lower CRs improves parallel efficiency by lowering synchronization cost, but degrades model accuracy as well (statistical efficiency). Further, speedup attained with different models and CRs also varies with network latency, effective bandwidth and collective op used for aggregation. In many cases, collectives like Allreduce (AR) have lower cost than AG to exchange the same amount of data. In this paper, we propose an AR-compatible Topk compressor that is bandwidth-optimal and thus performs better than AG in certain network configurations. We develop a flexible communication strategy that switches between AG and AR based on which collective is optimal in the current settings, and model the pareto-relationship between parallel and statistical efficiency as a multi-objective optimization (MOO) problem to dynamically adjust CR and accelerate training while still converging to high accuracy.

Bibrak Qamar Chandio, Prateek Srivastava, Maciej Brodowicz et al.

The paper provides a unified co-design of 1) a programming and execution model that allows spawning tasks from within the vertex data at runtime, 2) language constructs for \textit{actions} that send work to where the data resides, combining parallel expressiveness of local control objects (LCOs) to implement asynchronous graph processing primitives, 3) and an innovative vertex-centric data-structure, using the concept of Rhizomes, that parallelizes both the out and in-degree load of vertex objects across many cores and yet provides a single programming abstraction to the vertex objects. The data structure hierarchically parallelizes the out-degree load of vertices and the in-degree load laterally. The rhizomes internally communicate and remain consistent, using event-driven synchronization mechanisms, to provide a unified and correct view of the vertex. Simulated experimental results show performance gains for BFS, SSSP, and Page Rank on large chip sizes for the tested input graph datasets containing highly skewed degree distribution. The improvements come from the ability to express and create fine-grain dynamic computing task in the form of \textit{actions}, language constructs that aid the compiler to generate code that the runtime system uses to optimally schedule tasks, and the data structure that shares both in and out-degree compute workload among memory-processing elements.

Sahil Tyagi, Martin Swany

Distributed data-parallel (DDP) training improves overall application throughput as multiple devices train on a subset of data and aggregate updates to produce a globally shared model. The periodic synchronization at each iteration incurs considerable overhead, exacerbated by the increasing size and complexity of state-of-the-art neural networks. Although many gradient compression techniques propose to reduce communication cost, the ideal compression factor that leads to maximum speedup or minimum data exchange remains an open-ended problem since it varies with the quality of compression, model size and structure, hardware, network topology and bandwidth. We propose GraVAC, a framework to dynamically adjust compression factor throughout training by evaluating model progress and assessing gradient information loss associated with compression. GraVAC works in an online, black-box manner without any prior assumptions about a model or its hyperparameters, while achieving the same or better accuracy than dense SGD (i.e., no compression) in the same number of iterations/epochs. As opposed to using a static compression factor, GraVAC reduces end-to-end training time for ResNet101, VGG16 and LSTM by 4.32x, 1.95x and 6.67x respectively. Compared to other adaptive schemes, our framework provides 1.94x to 5.63x overall speedup.

Malintha Fernando, Ransalu Senanayake, Martin Swany

We propose a novel framework for real-time communication-aware coverage control in networked robot swarms. Our framework unifies the robot dynamics with network-level message-routing to reach consensus on swarm formations in the presence of communication uncertainties by leveraging local information. Specifically, we formulate the communication-aware coverage as a cooperative graphical game, and use variational inference to reach mixed strategy Nash equilibria of the stage games. We experimentally validate the proposed approach in a mobile ad-hoc wireless network scenario using teams of aerial vehicles and terrestrial user equipment (UE) operating over a large geographic region of interest. We show that our approach can provide wireless coverage to stationary and mobile UEs under realistic network conditions.