Ada Sedova

Sanjif Shanmugavelu, Mathieu Taillefumier, Christopher Culver et al.

Run to run variability in parallel programs caused by floating-point non-associativity has been known to significantly affect reproducibility in iterative algorithms, due to accumulating errors. Non-reproducibility can critically affect the efficiency and effectiveness of correctness testing for stochastic programs. Recently, the sensitivity of deep learning training and inference pipelines to floating-point non-associativity has been found to sometimes be extreme. It can prevent certification for commercial applications, accurate assessment of robustness and sensitivity, and bug detection. New approaches in scientific computing applications have coupled deep learning models with high-performance computing, leading to an aggravation of debugging and testing challenges. Here we perform an investigation of the statistical properties of floating-point non-associativity within modern parallel programming models, and analyze performance and productivity impacts of replacing atomic operations with deterministic alternatives on GPUs. We examine the recently-added deterministic options in PyTorch within the context of GPU deployment for deep learning, uncovering and quantifying the impacts of input parameters triggering run to run variability and reporting on the reliability and completeness of the documentation. Finally, we evaluate the strategy of exploiting automatic determinism that could be provided by deterministic hardware, using the Groq accelerator for inference portions of the deep learning pipeline. We demonstrate the benefits that a hardware-based strategy can provide within reproducibility and correctness efforts.

Sanjif Shanmugavelu, Mathieu Taillefumier, Christopher Culver et al.

The ability of machine learning (ML) classification models to resist small, targeted input perturbations -- known as adversarial attacks -- is a key measure of their safety and reliability. We show that floating-point non-associativity (FPNA) coupled with asynchronous parallel programming on GPUs is sufficient to result in misclassification, without any perturbation to the input. Additionally, we show that standard adversarial robustness results may be overestimated up to 4.6 when not considering machine-level details. We develop a novel black-box attack using Bayesian optimization to discover external workloads that can change the instruction scheduling which bias the output of reductions on GPUs and reliably lead to misclassification. Motivated by these results, we present a new learnable permutation (LP) gradient-based approach to learning floating-point operation orderings that lead to misclassifications. The LP approach provides a worst-case estimate in a computationally efficient manner, avoiding the need to run identical experiments tens of thousands of times over a potentially large set of possible GPU states or architectures. Finally, using instrumentation-based testing, we investigate parallel reduction ordering across different GPU architectures under external background workloads, when utilizing multi-GPU virtualization, and when applying power capping. Our results demonstrate that parallel reduction ordering varies significantly across architectures under the first two conditions, substantially increasing the search space required to fully test the effects of this parallel scheduler-based vulnerability. These results and the methods developed here can help to include machine-level considerations into adversarial robustness assessments, which can make a difference in safety and mission critical applications.

Alex Morehead, Chen Chen, Ada Sedova et al.

How and where proteins interface with one another can ultimately impact the proteins' functions along with a range of other biological processes. As such, precise computational methods for protein interface prediction (PIP) come highly sought after as they could yield significant advances in drug discovery and design as well as protein function analysis. However, the traditional benchmark dataset for this task, Docking Benchmark 5 (DB5), contains only a modest 230 complexes for training, validating, and testing different machine learning algorithms. In this work, we expand on a dataset recently introduced for this task, the Database of Interacting Protein Structures (DIPS), to present DIPS-Plus, an enhanced, feature-rich dataset of 42,112 complexes for geometric deep learning of protein interfaces. The previous version of DIPS contains only the Cartesian coordinates and types of the atoms comprising a given protein complex, whereas DIPS-Plus now includes a plethora of new residue-level features including protrusion indices, half-sphere amino acid compositions, and new profile hidden Markov model (HMM)-based sequence features for each amino acid, giving researchers a large, well-curated feature bank for training protein interface prediction methods. We demonstrate through rigorous benchmarks that training an existing state-of-the-art (SOTA) model for PIP on DIPS-Plus yields SOTA results, surpassing the performance of all other models trained on residue-level and atom-level encodings of protein complexes to date.