5 ms·
Yesss! This is what I was hoping to find last time I clicked on an HN headline "transformers from scratch", which turned out to be about machine learning: https
by myself248 5y ago
Yesss! This is what I was hoping to find last time I clicked on an HN headline "transformers from scratch", which turned out to be about machine learning: https://news.ycombinator.com/item?id=29280909 https://news.ycombinator.com/item?id=29280909
This is great. It gets into the flux density and iron loss calculations that affect the physical size of a transformer core, choosing the wire, and then exquisite detail of actually winding and finishing the thing.
I need to spend a bunch more time reading this site.
- formerly_proven 5y agoI find this trend of naming software ideas after common words really quite annoying (transformers, lenses, optics etc.)...
- bob1029 5y agoI agree, but at the same time "Sequence-to-sequence model utilizing self-attention" doesn't roll off the tongue as well and would otherwise make the thing more difficult to talk about.
- analog31 5y agoFrom what I've read, that's why physicists started giving things whimsical names like Quark, Charm, Strangeness, etc. The problem with using common words is that they always suggest an analogy, the analogy is never perfect, and often causes more confusion than understanding. (Also, Murray Gell-Mann was a whimsical character to begin with).
- amelius 5y agoOr naming companies after physical units (e.g. Tesla).
- codingdave 5y agoOr naming physical units after people (e.g. Tesla). FWIW, Tesla the company was named after Tesla the person, not the measurement.
- dylan604 5y agoYeah, the comment you replied is almost troll-like
- amelius 5y agoWhy? Many professionals use the unit Tesla on a daily basis. It's like someone claiming the word Pixel for their company or product which I'm sure already happened.
- formerly_proven 5y agohttps://store.google.com/category/phones https://store.google.com/category/phones
- myself248 5y agoOr using the word "bandwidth" to refer to "capacity" or "data throughput" rather than "the width of the band that the signal occupies". Arrrrgh. Yes the two things are the same for certain naïve modulations, but the whole point of using a better modulation is that you get more throughput for the same bandwidth. Understanding this requires first unlearning the abuse of the term, and that's hard work for a lot of engineers who really need to understand this.
- jbay808 5y agoShannon's channel theorem states that the information capacity of a channel is directly related to the bandwidth and the signal-to-noise ratio, even if you use the best possible modulation scheme: https://en.m.wikipedia.org/wiki/Shannon%E2%80%93Hartley_theorem https://en.m.wikipedia.org/wiki/Shannon%E2%80%93Hartley_theo...
- p1esk 5y agoHe’s talking about throughput, not the information capacity of the channel.
- jbay808 5y agoCan you explain the difference? They seem to me like virtually the same concept, but you seem to see an important distinction.
- p1esk 5y agoChannel capacity is an ideal theoretical concept establishing a maximum data transmission rate over a channel with strong assumptions about the noise in that channel. Throughput is what you actually get in the real world, and it depends on many things, including a modulation scheme, retransmission protocols, specific noise models, etc.
- jbay808 5y ago
- priansh 5y agoTouché, I opened this hoping to find a new method to train Transformers and got the opposite
- belter 5y agoThe Transformer Winding community knows that you won't let us down.
- scoot 5y agoA transformer and coil design guide from the same author: https://ludens.cl/Electron/Magnet.html https://ludens.cl/Electron/Magnet.html