prcvd.ai is working on a software platform that will democratize speaker awareness of audience perceptions. We believe that enabling speakers to be more aware, in enough time to correct, will help smoothing the effects of bias, and that this will improve overall communication. Use the form below to signup to recieve updates as we release products and services.
Tabular datasets are an important format to consider. Though anomaly detection could be a highly interesting topic, the dataset to which this post pertains is so significantly curated that it's hard to build intuition around the anomaly detection topic area. Instead, this is a proof of concept around fastai v2. Potentially the most interesting result is that the out of the box decision tree (sklearn CART implementation) worked almost flawlessly, while the fastai `tabular_learner` struggled. This will be an area of expansion to match performance between the methods.
Accurate facial attribute estimation using digital images is important because facial attributes are among the individual features that are most commonly cited as cause for bias (and discrimination). This article describes an approach to facial attribute profiling, which is a novel application to the best knowledge of the authors. The approach uses the ubiquitous supervised learning UNet architecture first published in the Biotech community, with a ResNet34 backbone. The UNet in this exercise was trained on the Mut1ny Face dataset to produce an accurate segmentation map of the human face.