Data-Efficient Image Transformers

Martins Bruveris

This is the next post in the series on the ImageNet leaderboard and it takes us to place #71 – Training data-efficient image transformers & distillation through attention. The visual transformers paper showed that it is possible for transformers to surpass CNNs on visual tasks, but doing so takes hundreds of millions of images and hundreds if not thousands of compute-days on TPUs. This paper shows how to train transformers using only ImageNet data on GPUs in a few days.

Continue reading “Data-Efficient Image Transformers” →

A Journey Along the ImageNet Leaderboard – Transformers

Martins Bruveris

This post is the first of a (planned) series that looks at the papers from the ImageNet leaderboard. As the top places are occupied by Transformers and EfficientNets, we will start our exploration with the 17th place – An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale – the first paper on visual transformers. Continue reading “A Journey Along the ImageNet Leaderboard – Transformers” →

TensorFlow Model Surgery

Martins Bruveris

There are things that are very easy to achieve in TensorFlow 2 and there are things that are annoyingly hard and often the dividing line is surprisingly narrow. For example, it is easy to combine two models into a bigger model, but splitting a model into two parts is difficult. Here we explore some ways of achieving the latter. Continue reading “TensorFlow Model Surgery” →

Exploring Bias using NIST’s FRVT

Martins Bruveris

“All unbiased algorithms are alike; each biased algorithm is biased in its own way.” In this post we will see how valid Tolstoy’s quote is when applied to face recognition. There is no doubt that face recognition algorithms are biased1: their accuracy is different for different population groups. In fact, it is hard to check the first part of Tolstoy’s quote, because we don’t know any unbiased algorithms. But we can investigate the second part. How does bias differ across algorithms? Continue reading “Exploring Bias using NIST’s FRVT” →

Have you heard of NCRFRILS?

Martins Bruveris

Neither had I. And the police in Washington, D.C. would have preferred to keep it that way. But since NCRFRILS is an automatic face recognition system used by the police to identify protestors accused of committing acts of violence, maybe some discussion is warranted. We are seeing more and more reports of police using facial recognition technology to identify people at protests.

What is worrying is the lack of public scrutiny of this use of facial recognition and the arguments advanced by the police to avoid such scrutiny. What are these arguments? Continue reading “Have you heard of NCRFRILS?” →

The Growing Threat of Smart Surveillance

Martins Bruveris
Face Recognition

In the UK we have become used to surveillance cameras. It is estimated that there are 1.85 million CCTV cameras in the UK or one camera for every 36 people. As we go about our lives we are, on average, seen by 70 cameras every day. There are 309 cameras at the Oxford Circus tube station alone. And while the cost-to-benefit ratio can be debated, CCTV cameras have been credited with a small but statistically significant reduction in crime and they helped police identify the Novichok killers in Salisbury. Continue reading “The Growing Threat of Smart Surveillance” →