Deep learning (DL) training algorithms utilize nondeterminism to improve models’ accuracy and training efficiency. Hence, multiple identical training runs (e.g., identical training data, algorithm, and network) produce different models with different accuracies and training times. In addition to these algorithmic factors, DL libraries (e.g., TensorFlow and cuDNN) introduce additional variance (referred to as implementation-level variance) due to parallelism, optimization, and floating-point computation.
This work is the first to study the variance of DL systems and the awareness of this variance among researchers and practitioners. Our experiments on three datasets with six popular networks show large overall accuracy differences among identical training runs. Even after excluding weak models, the accuracy difference is 10.8%. In addition, implementation-level factors alone cause the accuracy difference across identical training runs to be up to 2.9%, the per-class accuracy difference to be up to 52.4%, and the training time difference to be up to 145.3%. All core libraries (TensorFlow, CNTK, and Theano) and low-level libraries (e.g., cuDNN) exhibit implementation-level variance across all evaluated versions.
Our researcher and practitioner survey shows that 83.8% of the 901 participants are unaware of or unsure about any implementation-level variance. In addition, our literature survey shows that only 19.5±3% of papers in recent top software engineering (SE), artificial intelligence (AI), and systems conferences use multiple identical training runs to quantify the variance of their DL approaches. This paper raises awareness of DL variance and directs SE researchers to challenging tasks such as creating deterministic DL implementations to facilitate debugging and improving the reproducibility of DL software and results.