rfmw19

This article seems to approach at a pretty low level. For another take, I recently worked on a hobby project that built a multi-label deep learning classifier using CNNs and got about 95% accuracy for the validation set. I am somewhat familiar with signal processing with my background, but really wanted to just scratch the surface with deep learning. But to be clear, it was just a fun project for me and I don't pretend to understand everything I did.

My goal was to detect instruments in complex, full-length music, as opposed to single sources with no background noise. My approach was to generate Mel spectrograms for small sections of songs and then run the deep learning classifier on these images to build a list of labels for later use, e.g. generating playlists. I more than doubled the accuracy by adding to and curating my own dataset. I used Resnet50 as the base model and found decent results even when applied to spectrograms!

I still want to do more analysis to figure out what kind of small features the model was actually picking up on. I also want to try experiments like scrambling the spectrogram time-wise and seeing if the results are still as accurate.

Great video, thanks! I have one question:

There's a part where Rand says he was asked to provide questions for the congressional Facebook hearing. He said he was shocked when his questions were not used and instead "stupid" questions were asked. I'm not sure if he meant this sarcastically or not, but isn't it clear, especially to someone like Rand being an expert in disinformation and related areas, that the questions were likely deliberately not asked? I'm not sure why he was shocked.

> The root problem is always time and money.

I agree with you. However, if you're not some high level manager who controls this, what's the next best thing you can do?

I think most problems are less technical and are more about people and processes. You can still argue you don't have enough influence there, and that's completely possible and realistic. But that should be where we direct attention.

Yes, some technical advances in better tools and languages that provide stricter proofs and so on are needed and will help. But ultimately it's still the people that need to learn, use, and enforce the processes.

As for alternatives to paper towels, here's [0] a review of different studies about drying your hands. It's not clear that alternatives to paper towel hand drying are more effective from a health perspective. But obviously cost and environment are the biggest problems.

It seems most harmful microbes are between your nails on your fingertips and paper towels might be more effective with removing microbes on your fingertips, but there is still some controversy about if this even matters overall and which is more effective in removing them from other parts of your hands. Some are also concerned about air dryers blowing up airborne microbes, and I'm personally concerned with the super loud noises (also mentioned in the research below [0])!

Finally, I only see paper towels in every medical center or hospital I've been to.

[0] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3538484/

> (3) If you've chosen not to give them your contact book, they have worsened the UX over time (for example, it only shows phone numbers now and not the display names they have set).

Adding to this, you also cannot initiate a first (new) message to anyone unless you have granted contact access. The workaround is to ask the other person to message you first...

> The web interface is completely driven by your phone, acting as a remote control and WhatsApp still doesn't have access to your conversations.

> https://signal.org uses the exact same model and it's open source, so you can review it.

To be clear, Signal's desktop interface does not work like WhatsApp's web interface: Signal is not completely driven by your phone. Signal Desktop can still send and receive messages even if the phone it is tied to is completely off.

I might have misunderstood what you're trying to say, but wanted to clarify this.

> Most companies are dictatorships.

That's an interesting conclusion. I guess one big difference is that in most places you have a choice in whether or not to work for a company, but you don't have this choice when being born in a dictatorship, unless you consider migrating elsewhere a reasonable choice. It seems there are also less good choices when considering where to work (i.e. they're all becoming dictatorships?).

It's also as bad when you consider paying or using a free service provided by a dictatorship company. Do we really have a choice anymore? You may not choose email provider X, but all your friends use X so in the end it doesn't matter, e.g. in terms of data privacy. You may not choose social network Y, but all your friends have so you're forced to participate in the dictatorship, e.g. if you want to have a normal life.

McDonalds seems to be rolling them out slowly, but are actually cutting staff while simultaneously increasing their fleet of touchscreens. I see a lot more employees alternating between doing work in the kitchen vs. taking orders at the cashier. This is even during relatively high demand times with long customer wait lines. This is just an anecdote though, the data may disagree.

I'd also be curious to see any user interface studies on these touchscreens. In my limited experiences, they are completely slow and thus an awful experience. They clearly are trying to sell more at every step instead of being useful and efficient for the end user.

My method was more specific to bills and finance documents. I used a generic photo scanner. It's not as automatic as the purpose-built document scanners that have automatic feeders and support multiple pages, but I wanted something that I could use for photography as well.

I coupled this with some very hacked together Perl scripts with Tesseract OCR[1] that fed in data to ledger-cli[2] for handling bills. I put other generic documents into folders by date.

It worked pretty well, and I was able to generate some pretty graphs from data that was fully reconciled with financial institutions like my bank, credit card, investments, etc., but still took too much time. So what do I do now? Nothing!

This was years ago. I assume there is now better support from financial institutions for extracting data and this coupled with improved OCR/machine learning might make things more robust and make it worthwhile to try again.

[1] https://en.wikipedia.org/wiki/Tesseract_(software)

[2] https://www.ledger-cli.org/