The essence of machine learning is learning. Unfortunately, machines don’t know how to evaluate what they are learning, and if the training is biased, the machine’s output will reflect that bias.
Chiral Software’s work is primarily in security images, and bias in public safety is unacceptable. Therefore, Chiral Software has implemented several steps to avoid creating biased systems:
- Training data must be diverse. Chiral Software purposefully selects a highly diverse set of training data. For example, if all the training images of weapons feature men, a system runs a risk of associating males with weapons. We make sure this does not happen by creating balanced imagery.
- Where possible, self-training can remove bias by removing humans from the training loop. Chiral’s patented anomaly detection system learns from real behavior with no biases. The only biases it learns are what it sees.
- At an organizational level, every member of the Chiral Software team understands how training data impact results and we are always considering, is this something we want our software to learn?
We are happy to provide an analysis of data sets and evaluate them for bias, and review organizational practices to ensure machine learning benefits all of us equally.