Fairness without demographics in repeated loss minimization
Fairness without demographics in repeated loss minimization Hashimoto et al., ICML’18
When we train machine learning models and optimise for average loss it is possible to obtain systems with very high overall accuracy, but which perform poorly on under-represented subsets of the input space. For example, a speech recognition system that performs poorly with minority accents.
We refer to this phenomenon of high overall accuracy but low minority accuracy as a representation disparity… This representation disparity forms our definition of unfairness, and has been observed in face recognition, language identification, dependency parsing, part-of-speech tagging, academic recommender systems, and automatic video captioning.
For systems that are continually trained and evolved based on data collected from their users, the poor performance for a minority group can set in place a vicious cycle in which members of such a group use the system less (because it doesn’t work as well for them), causing them to provide less data and hence to be further under-represented in the training set…
… this problem of disparity amplification is a possibility in any machine learning system that is retrained on user data.
An interesting twist in the problem is that the authors assume neither the Continue reading
Some view the platform as the death knell for serverless platforms not based on Kubernetes, while others tout patience.
Leading cloud access security broker (CASB) vendors McAfee and Bitglass talk security and cloud-native attacks at Black Hat.
At the Alibaba Cloud Summit this week, the cloud provider launched a wide breadth of products that span serverless, data, search and analytics, IoT, and machine learning.
Nokia’s cloud-native core technology will help the Indian operator evolve from LTE to a 5G core.
