A comprehensive study on deep learning bug characteristics

Abstract

Deep learning has gained substantial popularity in recent years. Developers mainly rely on libraries and tools to add deep learning capabilities to their software. What kinds of bugs are frequently found in such software? What are the root causes of such bugs? What impacts do such bugs have? Which stages of deep learning pipeline are more bug prone? Are there any antipatterns? Understanding such characteristics of bugs in deep learning software has the potential to foster the development of better deep learning platforms, debugging mechanisms, development practices, and encourage the development of analysis and verification frameworks. Therefore, we study 2716 high-quality posts from Stack Overflow and 500 bug fix commits from Github about five popular deep learning libraries Caffe, Keras, Tensorflow, Theano, and Torch to understand the types of bugs, root causes of bugs, impacts of bugs, bug-prone stage of deep learning pipeline as well as whether there are some common antipatterns found in this buggy software. The key findings of our study include: data bug and logic bug are the most severe bug types in deep learning software appearing more than 48% of the times, major root causes of these bugs are Incorrect Model Parameter (IPS) and Structural Inefficiency (SI) showing up more than 43% of the times. We have also found that the bugs in the usage of deep learning libraries have some common antipatterns that lead to a strong correlation of bug types among the libraries.

Testing machine learning based systems: a systematic mapping

“…Indeed, the very notion of a fault for an MLS is more complex than in traditional software. The code that builds the MLS may be bug-free, but it might still deviate from the expected behaviour due to faults introduced in the training phase, such as the misconfiguration of some learning parameters or the use of an unbalanced/non-representative training set (Humbatova et al 2020;Islam et al 2019).…”

“…For what concerns the actions to be taken when a fault is detected, most approaches suggest re-training and show that re-training is effective in handling the adversarial/corner-case inputs that caused the misbehaviours. However, re-training is the right corrective action only if the discovered fault is associated with the training phase, but not all ML faults are necessarily due to inadequate training, as reported in existing taxonomies of deep learning faults (Humbatova et al 2020;Islam et al 2019;Zhang et al 2018a). For instance, the model structure (not its training) may be inadequate for the task being considered.…”

“…The same issue affects existing proposals of ML mutation operators, which do not mimic real faults and introduce unrealistic issues. For instance, mutation operators that modify the weights of a neural network after it has been trained do not match any ML fault reported in the existing taxonomies (Humbatova et al 2020;Islam et al 2019;Zhang et al 2018a) and look quite unrealistic.…”

Does fixing bug increase robustness in deep learning?

“…Islam et al [12] has built the classification scheme on top of the study performed by [44] and [3] to understand the commonality and the variability of the bugs found in five deep learning libraries. The bug dataset has been utilized in their recent study [14] to understand the fixing patterns to address the bugs.…”

“…In this study, we did not execute the code; rather, we manually verified the code snippet and the discussion to understand how the changes affect the learning capability of a DNN model. Also, we referred to the root cause, bug type, and impact from the [12], which refers to the same dataset.…”