FAIR’s RegNet outperforms SOTA EfficientNet and runs 5x faster on GPUs

A team of researchers from FAIR (Facebook AI Research) has recently developed RegNet, a low-dimensional design space. The new design space outperforms various existing traditional models and functions 5x faster on the Graphics Processing Units.

FAIR, which was developed by the social networking giant in 2014, has been openly collaborating with the community to advance artificial intelligence through applied and fundamental research. It has become an international research organization since its inception, with laboratories in London, Pittsburgh, Seattle, Tel Aviv, Montreal, Paris, New York, and Menlo Park.

Researchers of FAIR have reportedly stated that RegNet produces fast, simple, and versatile networks with general design principles, and outperforms SOTA EfficientNet models from Google. EfficientNet, introduced by Google in 2019, has adopted a combination of model scaling rules and network attached storage (NAS). It also represents the current Summits on the Air (SOTA).

A team of researchers in the Facebook AI team has conducted controlled comparisons of RegNet with EfficientNet following the same training setup without any training-time enhancement. The team focused on delivering actual network design spaces that are comprised of huge & potentially infinite model architecture populations. The quality of this design space is analyzed by EDF (empirical distribution function). Researchers also have reportedly discovered unexpected insights into network design while analyzing RegNet. For instance, they identified that the depth of the models is stable across various compute regimes, with 60 layers or 20 blocks as an optimal depth.

Researchers further noticed that the usage of inverted bottlenecks can degrade the performance of mobile networks. Neither an inverted bottleneck nor a bottleneck is being used in the best models. They also recently built an AI tool that tricks the facial recognition system known as the de-identification system to wrongly identify a subject in a video. The system also changes the key facial features of the person in the video by using machine learning technology.

Source credit: