Daniel Lemire's blog

, 3 min read

Science and Technology links (December 22nd, 2017)

Bitcoins are electronic records that are meant to be used as currency. They have become increasingly popular and expensive. Reportedly, the famous investment bank Goldman Sachs is planning to offer bitcoin financial services in the summer of 2018.

Doing away with television and switching to subscription services like Netflix reduces your exposure to ads. So we have Facebook and Google who earn almost all their money through ads. Netflix is effectively saving us from ads.

MagicLeap, the Florida-based augmented reality company, has announced its first product… some good looking glasses. Apparently, they are quite nice. Price remains undisclosed and you cannot buy them in store yet. It appears that will need to strap on some computer to make them work.

People are typically very interested in fairness, though it is sometimes unclear what it means in practice. What if you recruited your employees based on anonymized resumes where you can’t tell the race or the gender of the candidate? That would be “fairer” certainly, right? Maybe not:

We find that participating firms become less likely to interview and hire minority candidates when receiving anonymous résumés. We show how these unexpected results can be explained (…) by the fact that anonymization prevents the attenuation of negative signals when the candidate belongs to a minority.

It is widely believed that there is more genetic variation within human populations than between them. On this basis, it follows that the concept of “human race” is likely “socially constructed”: it has no biological basis. West Hunter has a counterpoint:

Pygmies couldn’t really be hugely shorter than most other human populations. Yet they are short (five or six standard deviations shorter)

He also points out that evolution can run its course faster than we might intuitively expect. For example, lactose tolerance is thought to have evolved in the last 10,000 years or so. Thus while it is impossible to have a rational discussion about it in today’s political climate, there are significant genetic differences between human populations, that is simply a scientific fact. It does not follow that the concept of human race is useful or relevant, but if we were a scientific civilization, we would certainly be able to talk openly about genetic differences between populations.

Awni Hannun has a post entitled Speech Recognition Is Not Solved. It is a response to same claims made this year that computers have no better speech recognition than human beings (Microsoft made such claims). He points out that performance is a major constraint when deploying new algorithms…

(Latency) Bidirectional recurrent layers are a good example of a latency killing improvement. All the recent state-of-the-art results in conversational speech use them. The problem is we can’t compute anything after the first bidirectional layer until the user is done speaking. So the latency scales with the length of the utterance.

(Compute) The amount of computational power needed to transcribe an utterance is an economic constraint. We have to consider the bang-for-buck of every accuracy improvement to a speech recognizer. If an improvement doesn’t meet an economical threshold, then it can’t be deployed.

To a lot engineers, this will sound quite familiar. Research papers often rely on trade-offs that are simply unrealistic in the real world.