Skip to content

Plenary Session Details

Plenary Session 1

Details

Discussion Notes


Plenary Session 2

Details

Discussion Notes


Plenary Session 3

Details

  • Topic: Cross Cutting Topics in Deep Learning and Software Engineering
  • Time: 4:30pm - 6:00pm
  • Room: Cortez 3
  • Session Lead: Denys Poshyvanyk
  • Session Scribes: Xiangyu Zhang, Kevin Moran
  • All Participants

Discussion Points

In DL-based systems it is very difficult to control where and how information is stored (knowledge is stored as weights of the network in all the layers with some specialization in specific regions of NN). Data safety regulations (e.g., European General Data Protection Regulation) may require anonymized/aggregated statistics potentially impacting the performance of the system, trouble-shooting and data exploration.

  1. “Explainability of DL models”

    • Deep learning models raise ethical questions when applied to specific tasks, if we are unable to explain why a specific classification occurred then we are powerless to control the moral and ethical dilemmas that DL introduces.
    • To progress the effectiveness of DL models we need to understand the inter-workings of the DL components (neurons, backprop, gradient descent) but more importantly, how they impact and influence one another leading to an input resulting in an output.
    • Reduction of complexity is an important concept in SE. Since we are unable to determine exactly what and why certain features are valued to create a relationship between the input and output, we are also unable to reduce the complexity of the model while maintaining its effectiveness.
  2. “Security implications”

    • Deep learning poses a grave security threat using adversarial examples, which are small perturbations to an input which cause a very unexpected output.
    • Security has two research directions: 1. the process of creating new, adversarial data examples to confuse and break DL models and 2. the process of defending against these adversarial examples using unique training techniques.
    • Security is also directly tied to the variety of data given for training, therefore, unexpected or unique instances (even when not intentionally generated) can be problematic. When datapoints fall out of their expected distribution into outlier territory, there is no guarantee that the model will react as we expect which can have security risks for whatever the task is that is model dependent.
  3. “Testing for Bias in training data”

    • DL models are data dependent, but humans don't always collect data with equality and a removal of bias in mind, therefore, we can expect DL models to react in similar ways when trained on this type of data
    • Since we are unable to control the features the model extracts, or understand why they extract them, models have free reign to consider discriminatory variables in order to make classifications.
    • Testing for bias data is complex because it is dependent on human judgement. Ideally, we could train a DL model to tell us when data is bias, but that would require having a dataset of what is considered diverse and without bias. We desperately need an objective way to determine when data is unbias and more importantly understand how bias data may affect classification of DL models.
  4. Reviewing expertise for papers that sit at the intersection of AI/DL & SE

    • As AI/DL related papers in SE become a more popular topic, we need to address issues related to reviewing expertise.

Discussion Notes