Unnaturally Long Attention Span

When it’s Appropriate to Use Machine Superintelligence

I refer to Superintelligent systems as those that are considered in popular parlance to be “AI”, but go beyond the computational capabilities of the human brain.  Such as one Deep Blue chess playing system, or a web search engine, or even a pocket calculator.  These systems perform intelligent tasks, but in a very different way than a human does.  These systems can achieve high performance by taking advantage of computational or storage facilities that humans do not have in their biological arsenal.

The question that inevitably arises for AI practitioners in the process of building intelligent systems is: when should Superintelligent techniques be used?

For machine learning engineers working at a company where percentage points of classification accuracy directly translate to revenues, the right answer may be “whenever possible”.  Yet for an AI purist trying to build human-like agents the answer is “almost never”.  Russell and Norvig identified this fundamental dichotomy in their treatise on AI as the diametric goals of building systems that think like humans versus building systems that think rationally.

I’m currently working on a problem in which the best approaches that are inspired by human intutions, and hence generalize to a wide range of situations, have led performance to a plateau.  A typical machine learning expert is trained in this regard to examine their dataset and misclassification matrix, make diagnostic measurements, such as variance, bias, learning rate, and make the appropriate adjustments to either their choice of algorithm or feature set.

However, before one proceeds with these diagnostics I’d advocate another way of thinking about the problem for the practitioner concerned with both performance and generalization. One that has to do with considering information representation.

We often expect our intelligent systems to behave in response to us in a human way, but too easily forget that the representation of the world that the machine recieves is very different than the one we as humans receive.  In a sense, the machine lives in a different Universe and all of his connections to reality are gated by these artificial inputs.  The reason Stanford’s autonomous SUV Stanley was able to navigate through the Mojave desert and win the $1 million DARPA grand challenge was not due to a breakthrough in better, more human-like decision making, but a barrage of on-board sensors including 5 roof-mounted laser range finders, dual 24GHz RADARs, GPS, IMUs, all fed to a supercomputer in the trunk. It would have been a different matter entirely if Stanley was able to achieve this feat with just two cameras and two audio sensors.

A single event in objective reality is very different when projected into the human or machine experience.  A human rating movies on Netflix gets a very different psychological and phenomenological experience than he gets from classifying rows of numbers by staring at a spreadsheet of vectors, even though both acts produce the same functional result.

In cases where the representation of the Universe that a machine gets is an undersampled version of the minimal feature set that a human needs to perform the same task, the machine needs to use Superintelligence in order to achieve the same functional result. Many different information-theoretic measures can be used to test whether the input signals correspond well when it’s not obviously apparent. Besides, achieving human-like response from a machine is quite trivial with the right inputs; it’s with the wrong inputs that we have to use more sophisticated techniques.

Vocals Remover for Windows Media Player

Last week, I was looking for some software to remove the voice tracks from mp3s. There are plenty of independent audio editing programs that can filter out vocals from sound files, but I didn’t want to modify my mp3s or keep a whole ‘nother voice-stripped version of my music library. What I really wanted was just a minimal plugin for my Windows Media Player that I could enable to turn off vocals. There didn’t seem to be anything out there freely available.

So, I started hacking my own plugin, and I’m releasing it here for free download in case it could be useful for others. (Hi, Google!)

Vocals Remover is an Audio DSP plugin for Windows Media Player that cancels out the voice track in real-time so that you can sing along. It supports all major file formats (including video) and lets you adjust the amount of voice removal and gain compensation. While the plugin works pretty reliably, it can get confused on songs where the foreground singer and instrumentals are hard to tell apart.  So, don’t expect it to work well on your experimental house mix or just about anything by T-Pain.

Download
Vocals Remover for  Windows Media Player.msi (44 KB)
Requirements: WMP 11 or greater

To install, simply run the downloaded package and uninstall works in the usual way via Windows control panel.

To enable or disable, just see whether the plugin is selected in WMP.  In WMP12, this is found by right-click -> Tools -> Plugins.  There is also a properties panel in Tools > Options >Plug-ins > Audio DSP -> Properties button, where you can customize the amount of the effect to apply.  A 0.0 means the effect is effectively off and a 1.0 means the maximum amount of voice cancellation is applied.

A short demo:

The above video also features the highly recommended lyrics plugin, which in combination with the Vocals Remover, turns your PC into a hacktastic DIY karaoke box!