Search

Deep learning technologies, particularly deep neural networks (DNNs), have demonstrated significant success across many domains. This success has been accompanied by substantial advancements and innovations in the algorithms behind the operations required by DNNs. These enhanced algorithms hold the potential to greatly increase the performance of DNNs. However, discovering the best performing algorithm for a DNN and altering the DNN to use such algorithm is a difficult and time consuming task. To address this, we introduce an open source framework which provides easy to use fine grain algorithmic control for DNNs, enabling algorithmic exploration and selection. Along with built-in high performance implementations of common deep learning operations, the framework enables users to implement and select their own algorithms to be utilized by the DNN. The framework's built-in accelerated implementations are shown to yield outputs equivalent to and exhibit similar performance as implementations in PyTorch, a popular DNN framework. Moreover, the framework incurs no additional performance overhead, meaning that performance depends solely on the algorithms chosen by the user., Comment: IEEE HPEC 2024

Quantum circuit cutting has been proposed to help execute large quantum circuits using only small and noisy machines. Intuitively, cutting a qubit wire can be thought of as classically passing information of a quantum state along each element in a basis set. As the number of cuts increase, the number of quantum degrees of freedom needed to be passed through scales exponentially. We propose a simple reduction scheme to lower the classical and quantum resources required to perform a cut. Particularly, we recognize that for some cuts, certain basis element might pass "no information" through the qubit wire and can effectively be neglected. We empirically demonstrate our method on circuit simulators as well as IBM quantum hardware, and we observed up to 33 percent reduction in wall time without loss of accuracy., Comment: 7 pages, 5 figures, submitted to 37th IEEE International Parallel & Distributed Processing Symposium

Reinforcement Learning (RL) has achieved significant success in application domains such as robotics, games and health care. However, training RL agents is very time consuming. Current implementations exhibit poor performance due to challenges such as irregular memory accesses and thread-level synchronization overheads on CPU. In this work, we propose a framework for generating scalable reinforcement learning implementations on multi-core systems. Replay Buffer is a key component of RL algorithms which facilitates storage of samples obtained from environmental interactions and data sampling for the learning process. We define a new data structure for Prioritized Replay Buffer based on $K$-ary sum tree that supports asynchronous parallel insertions, sampling, and priority updates. To address the challenge of irregular memory accesses, we propose a novel data layout to store the nodes of the sum tree that reduces the number of cache misses. Additionally, we propose $\textit{lazy writing}$ mechanism to reduce thread-level synchronization overheads of the Replay Buffer operations. Our framework employs parallel actors to concurrently collect data via environmental interactions, and parallel learners to perform stochastic gradient descent using the collected data. Our framework supports a wide range of reinforcement learning algorithms including DQN, DDPG, etc. We demonstrate the effectiveness of our framework in accelerating RL algorithms by performing experiments on CPU + GPU platform using OpenAI benchmarks., Comment: 10 pages. HiPC21

Online interactions with the environment to collect data samples for training a Reinforcement Learning (RL) agent is not always feasible due to economic and safety concerns. The goal of Offline Reinforcement Learning is to address this problem by learning effective policies using previously collected datasets. Standard off-policy RL algorithms are prone to overestimations of the values of out-of-distribution (less explored) actions and are hence unsuitable for Offline RL. Behavior regularization, which constraints the learned policy within the support set of the dataset, has been proposed to tackle the limitations of standard off-policy algorithms. In this paper, we improve the behavior regularized offline reinforcement learning and propose BRAC+. First, we propose quantification of the out-of-distribution actions and conduct comparisons between using Kullback-Leibler divergence versus using Maximum Mean Discrepancy as the regularization protocol. We propose an analytical upper bound on the KL divergence as the behavior regularizer to reduce variance associated with sample based estimations. Second, we mathematically show that the learned Q values can diverge even using behavior regularized policy update under mild assumptions. This leads to large overestimations of the Q values and performance deterioration of the learned policy. To mitigate this issue, we add a gradient penalty term to the policy evaluation objective. By doing so, the Q values are guaranteed to converge. On challenging offline RL benchmarks, BRAC+ outperforms the baseline behavior regularized approaches by 40%~87% and the state-of-the-art approach by 6%., Comment: 16 pages. Accepted by ACML21

Hash table is a fundamental data structure for quick search and retrieval of data. It is a key component in complex graph analytics and AI/ML applications. State-of-the-art parallel hash table implementations either make some simplifying assumptions such as supporting only a subset of hash table operations or employ optimizations that lead to performance that is highly data dependent and in the worst case can be similar to a sequential implementation. In contrast, in this work we develop a dynamic hash table that supports all the hash table queries - search, insert, delete, update, while allowing us to support 'p' parallel queries (p>1) per clock cycle via p processing engines (PEs) in the worst case i.e. the performance is data agnostic. We achieve this by implementing novel XOR based multi-ported block memories on FPGAs. Additionally, we develop a technique to optimize the memory requirement of the hash table if the ratio of search to insert/update/delete queries is known beforehand. We implement our design on state-of-the-art FPGA devices. Our design is scalable to 16 PEs and supports throughput up to 5926 MOPS. It matches the throughput of the state-of-the-art hash table design - FASTHash, which only supports search and insert operations. Comparing with the best FPGA design that supports the same set of operations, our hash table achieves up to 12.3x speedup., Comment: 2020 IEEE High Performance Extreme Computing Conference (HPEC)

Most of the existing work on FPGA acceleration of Convolutional Neural Network (CNN) focus on employing a single strategy (algorithm, dataflow, etc.) across all the layers. Such an approach does not achieve optimal latency on complex and deep CNNs. Emerging CNNs have diverse per-layer computation characteristics including parallelism, arithmetic intensity, locality, and memory footprint. Per-layer strategy selection and fine-grained tuning are required to achieve low end-to-end latency. However, specialized hardware modules dedicated to each layer limit the per-layer utilization and adversely affect end-to-end latency. In this paper, we address these problems by an algorithm-architecture co-optimization framework, DYNAMAP, consisting of (1) a unified hardware overlay that can be reused across layers, supporting dynamic mapping of all three families of popular convolution algorithms, and further allowing flexible dataflow switching to maximize hardware utilization for each layer; (2) a novel software Design Space Exploration (DSE) flow that customizes the hardware overlay and chooses optimal strategy mapping. We show that the algorithm mapping space increases exponentially with network depth, and while the optimal algorithm selection problem is NP-hard in general, by exploiting the series-parallel structure of CNN models, we demonstrate a polynomial-time solution for optimal algorithm mapping. DYNAMAP is optimized for any CNN, including those having diverse computation and memory requirements across the layers. We demonstrate DYNAMAP using two state-of-the-art CNNs - GoogleNet and Inception-V4. The generated accelerators achieve up to $2.8\times$ and $1.4\times$ speedups, respectively, wrt inference latency compared with the state-of-the-art FPGA implementations., Comment: Published in ACM/SIGDA FPGA '21

Model usage is the central challenge of model-based reinforcement learning. Although dynamics model based on deep neural networks provide good generalization for single step prediction, such ability is over exploited when it is used to predict long horizon trajectories due to compounding errors. In this work, we propose a Dyna-style model-based reinforcement learning algorithm, which we called Maximum Entropy Model Rollouts (MEMR). To eliminate the compounding errors, we only use our model to generate single-step rollouts. Furthermore, we propose to generate \emph{diverse} model rollouts by non-uniform sampling of the environment states such that the entropy of the model rollouts is maximized. We mathematically derived the maximum entropy sampling criteria for one data case under Gaussian prior. To accomplish this criteria, we propose to utilize a prioritized experience replay. Our preliminary experiments in challenging locomotion benchmarks show that our approach achieves the same sample efficiency of the best model-based algorithms, matches the asymptotic performance of the best model-free algorithms, and significantly reduces the computation requirements of other model-based methods., Comment: ICML BIG Workshop 2020, camera ready version

Buildings sector is one of the major consumers of energy in the United States. The buildings HVAC (Heating, Ventilation, and Air Conditioning) systems, whose functionality is to maintain thermal comfort and indoor air quality (IAQ), account for almost half of the energy consumed by the buildings. Thus, intelligent scheduling of the building HVAC system has the potential for tremendous energy and cost savings while ensuring that the control objectives (thermal comfort, air quality) are satisfied. Recently, several works have focused on model-free deep reinforcement learning based techniques such as Deep Q-Network (DQN). Such methods require extensive interactions with the environment. Thus, they are impractical to implement in real systems due to low sample efficiency. Safety-aware exploration is another challenge in real systems since certain actions at particular states may result in catastrophic outcomes. To address these issues and challenges, we propose a model-based reinforcement learning approach that learns the system dynamics using a neural network. Then, we adopt Model Predictive Control (MPC) using the learned system dynamics to perform control with random-sampling shooting method. To ensure safe exploration, we limit the actions within safe range and the maximum absolute change of actions according to prior knowledge. We evaluate our ideas through simulation using widely adopted EnergyPlus tool on a case study consisting of a two zone data-center. Experiments show that the average deviation of the trajectories sampled from the learned dynamics and the ground truth is below $20\%$. Compared with baseline approaches, we reduce the total energy consumption by $17.1\% \sim 21.8\%$. Compared with model-free reinforcement learning approach, we reduce the required number of training steps to converge by 10x., Comment: 10 pages, 13 figures, to be appear in ACM BuildSys '19, November 13-14, 2019, New York, NY, USA

Mitigating Supply-Demand mismatch is critical for smooth power grid operation. Traditionally, load curtailment techniques such as Demand Response (DR) have been used for this purpose. However, these cannot be the only component of a net-load balancing framework for Smart Grids with high PV penetration. These grids can sometimes exhibit supply surplus causing over-voltages. Supply curtailment techniques such as Volt-Var Optimizations are complex and computationally expensive. This increases the complexity of net-load balancing systems used by the grid operator and limits their scalability. Recently new technologies have been developed that enable the rapid and selective connection of PV modules of an installation to the grid. Taking advantage of these advancements, we develop a unified optimal net-load balancing framework which performs both load and solar curtailment. We show that when the available curtailment values are discrete, this problem is NP-hard and develop bounded approximation algorithms for minimizing the curtailment cost. Our algorithms produce fast solutions, given the tight timing constraints required for grid operation. We also incorporate the notion of fairness to ensure that curtailment is evenly distributed among all the nodes. Finally, we develop an online algorithm which performs net-load balancing using only data available for the current interval. Using both theoretical analysis and practical evaluations, we show that our net-load balancing algorithms provide solutions which are close to optimal in a small amount of time., Comment: 11 pages. To be published in the 4th ACM International Conference on Systems for Energy-Efficient Built Environments (BuildSys 17) Changes from previous version: Fixed a bug in Algorithm 1 which was causing some min cost solutions to be missed