How Deep Neural Networks Work - Full Course for Beginners

I just woke up. I am very confused. Why am i here-

dunno why but this video was playing when i woke up in the middle of the night

Just woke up, don’t know where I am or how I ended up here

So, I just woke up to this video on my phone but the ironic part is I just learned about this yesterday.

Watching this on my way to sleep for all the people who are waking up to this, it might break the cycle. 🙏💪

Let me guess: you just woke up and this video was playing

i slept watching a different completely unrelated video and woke up on this what just happened

Somehow this autoplayed on my phone while I was sleeping.

I came with an interest in neutral networks. I left feeling well rested.

I fell asleep watching a very simple maths video and woke up to this after dreaming that me and my friends were studying its contents. I’ve never done anything to do with this before but I understood it when I was dreaming about it so will probably give it another listen. It reminds me of being in College/University when SWIM was doing a bunch of drugs and accidentally designed a computer brain. Score for drugs 1,264,273,995,267,177, score for sobriety: still zero LOL

From sleeping on a Geopolitics video to landing here, I am stunned😅

I woke up and this was playing on the background

Just woke to this playing. It was the catalyst to the craziest most vivid dream since childhood...im in my 30's.

So to be the first, I’d just like to say my journey consisted of falling asleep to a video about why a magnet on the front of a car wouldn’t work, then it went to cursed units of measurement, then it went to professor Dave explains and then I ended up here, all in all I’ve been asleep for about 3 hours and I need more sleep… Anyone else wanna share the journey?

what is this, i just woke up..

Assuming that everyone has had or currently has a learning capacity you realize that environment plays a huge part.

Very surprised to see everyone woke up to this video as well. The algorithm strikes again!

Apparently everyone waking up to this including myself 😂

Youtube is a good detector of sleep

If you're an audio guy, Squash functions are just compression by factor of Ratio (r). Threshold is the pickup weight input and knee is smoothing of weights between input and output over a certain range. And there you go. Compression in a nutshell. However the dB peak scale is non-linear. The dB scale is power of 2x10dB. That's what makes it the most confusing. So a ratio of 10 to keep it simpler is double the volume at the threshold gradually weighting less until the set peak where compression is zero. The knee rolls off that effect by a dB factor at a specified loudness and breadth of its impact. Seems gaussian to me. I don't know how the math works at the knee but it gives a smoother transition from boosted to left alone. So in a typical simple compression threshold at -24dB with 10 ratio would result in threshold at -12dB tapering to -10dB, -8dB, -6 and so on until you hit zero assuming your highest peaks are 0dB which is bad. Then you adjust the output to -8 or -14 depending on the sound and that scales the whole curve downward unaltered relatively by whatever output dB you set. If your threshold was boosted by compression to -12dB and you scale it down in output by -8dB then your threshold after processing will be -20dB tapering off up to -8dB in the same curve it had before the output was scaled down. That's why you have to adjust input vs threshold vs ratio vs knee vs output to get the best out of simple compression. Multiband compression is the same thing just much more complicated as it accounts for frequency where you can specify within a certain frequency range how much compression you'd like. Overlap them and yeah that gets quite complicated but it's super useful to getting the right sound especially in dialogue to grab and manipulate the loudness of tonality and sibilance while rejecting the background noise or any echo or unwanted reverb. The same principles apply in NNs in more of a deterministic and mathematical way. It entirely depends on the architecture and what it is used for as you are taking a larger dynamic range of inputs and compressing them to a smaller range of outputs. That's why CDs in the 90's Redbook audio was 16 bits wide. 2^16 made for 65536 levels of volume for any given sample. That was enough because it was replacing cassette tape which had horrible dynamic range. Now it's standard to have 24 bit audio which has a vastly higher dynamic range of 16,777,216 levels of volume at any given sample. For production and processing it's common to have 96 bit audio which has 7.92281625 x 10^28 levels of loudness. That's technically not better than analog but no human would ever be able to tell the difference. It helps computers and audio processing make very very accurate changes.