Weekly Head Voices #182: Can you predict what ...

Welcome to your home away from home folks!

This, the one hundred and eighty second edition of the WHV, looks back at the three weeks from Monday October 14 to Sunday November 3, 2019.

One of the beautiful views from our Sunday (October 20) lunch at Mont Marie in Stellenbosch.

On attempting to publish the WHV on a weekly basis.

Somewhere at the end of the first week, I sat down to write, but I soon realised that I didn't have all that much to say.

Ironically, earlier that day I had had an online discussion where I made the statement that I preferred putting out posts regularly, rather than waiting until until I had something “really good” to publish.

In the latter case, one tends to let things slip further and further, with that great internal excuse that one's ideas and writing do not yet satisfy one's otherwise unspecified but high standards.

In addition, one's performance anxiety is now fueled into a frenzy by the illusion that surely the expectations of one's audience are growing greater by the minute, and hence the next post will be measured even more strictly against the mythical yard stick of blog posts that in reality very few people, mostly friends, actually take the time to read.

So that happened.

As is the case with many other things in life that seem to follow the Sisyphean ideal, I have made peace with the fact that I will just continue working at this, until the end.

A call from the future.

I made three longer term technology-themed bets with friends and colleagues at the IEEE VisWeek conference in 2009.

The idea was that the first of the three would play out and finally be evaluated at VisWeek 2019 (10 years later), and the other two more or less 20 years later, to be decided at Visweek 2029.

Fortunately I wrote the bets up in a blog post, all the way back in 2009, titled Futuristic Betting at VisWeek 2009.

To my surprise, but far more to my delight, an overlapping subset of the involved friends and colleagues recorded an hour-long video message at VisWeek 2019 in Vancouver and sent it to me via the YouTubes.

The main message of their video was that I had lost bet #1, with which I could agree, and that they had their doubts about my chances in bets #2 and #3, with which I can not yet agree.

You can read the addendum I made to the original blog post for all of the details.

When I made those predictions, confident that we would be able to debate and laugh about them at those future conferences, I ironically had no idea that my life was going to take the turns that it did, out of academia and shortly thereafter out of Europe.

Fortunately, that futuristic betting blog post seems to have functioned as a sort of anchor in the sands of time, playing its small part in making the connection from 2009 to 2019, from Atlantic City to Vancouver and even to Somerset West.

What will happen in 2029?

Humans predicting the world in real-time.

Based on this recent EEG-based experimental study, it sounds like our brains are continuously predicting the next word that someone we're in conversation with is going to say.

Especially in noisy environments, such as IN DA CLUB, this helps us to understand what our conversation partners are saying.

What struck me when I read the linked press release, and shortly thereafter tried to assimilate the full article, quite attractively titled Neural Signal to Violations of Abstract Rules Using Speech-Like Stimuli, was that is exactly what computer-based language models do.

Language models are mathematical constructs, which mostly live in software, that attempt to predict the next word in a sentence, based on the previous words in a sequence.

Recently, deep learning approaches overtook the more classical statistical approaches for building such language models.

Given a few thousand articles in French for example, a neural network can learn how to predict the next word in any given French sentence with high accuracy.

Together with the fact that the internet is now overflowing with examples of many human languages, anyone with some spare time and GPU cycles to burn can train up a language model, or even download a pre-trained example.

Such a language model can then be fine-tuned for other language tasks such as sentiment analysis and topic classification.

I enjoyed seeing this (growing) similarity between our built-in language processing wetware and the rapidly developing world of deep learning.

Maybe you did too.

Just a little bit.

My phone camera definitely needs more of that DeepFusion to capture the amazing contrast patterns that these particular clouds were displaying at that moment. You're going to have to take my word for it.

Contents