Un-graceful degradation of ChatGPT

Connectionsim, an early wave of artificial neural network research, promised to tackle AI’s embarrassing tendency to fail catastrophically due to minor random flaws in the code or input data. Indeed, modern multi-layer neural network designs seem to exhibit “graceful degradation” in many aspects of their operation. Yet, if you push them enough you’ll get them to fail miserably, just at a higher semantic level.

Here’s an example of a brief interaction I had with ChatGPT. Notice how confident the answer sounds. Indeed, there’s depth to it. Unfortunately, the knowledge system is hapless as to the fact that it has landed in the wrong domain. It mixed up one famous name (Nikolai Bernstein) with another one (Leonard Bernstein). It also mixed up statistics with neuro/movement science; ultimately it gave a good answer to the wrong question. That’s not how humans would respond typically. If they knew enough about the topic of the question they would at least know the names and domains. In conclusion, we can conclude that ChatGPT hosts a rich body of knowledge but its retrieval of context can fail as catastrophically as we would expect from any of the earlier computer expert systems.

Q: What is the degrees of freedom problem according to Bernstein?

A: The Degrees of Freedom problem, also known as the “Bernstein’s problem,” is a statistical issue that arises when trying to estimate the parameters of a multivariate probability distribution using a limited sample size. The problem is that the number of parameters … first identified by the statistician Leonard Bernstein in the 1930s.

The energy of biological oscillations matters too, not just amplitude, phase, and frequency

On the scaling properties of oscillatory modes with balanced energy https://www.frontiersin.org/articles/10.3389/fnetp.2022.974373/full

A little theoretical side-project in complex systems. I’m trying to argue that scaling phenomena in physiology such as 1/f in the background of EEG are easier to make sense of if you think of them in terms of dimensions that have both physical and biological interpretation, such as energy. For many physical systems, to increase their frequency they either have to reduce their amplitude or start pumping in more energy. Swinging both hard and fast is not for free! Reducing amplitude to compensate for higher frequency keeps energy balanced across oscillatory action modes. So the amplitude ~ 1/frequency relation is more interesting to see as flat energy across frequencies.

Feel the beat! It resonated a lot with the media

This broke the mainstream media. Dan Cameron’s paper, Undetectable very-low frequency sound increases dancing at a live concert https://www.sciencedirect.com/science/article/abs/pii/S0960982222015354.

Daily Beast

NPR All Things Considered

Haaretz

Nature Research Highlight

Naked Scientist

Guardian

Daily Mail

Swedish Public Television

Quirks and Quarks

Australian Broadcasting Corporation

ZME Science

BBC

Vice.com

CTV

Sunday Vision