Hacker News new | past | comments | ask | show | jobs | submit login
The Basic Ideas in Neural Networks (1994) [pdf] (stanford.edu)
181 points by sonabinu on Jan 10, 2018 | hide | past | web | favorite | 16 comments

Whoa, cool, the parent directory is publicly accessible! And it's chockablock full.

Of course it's indexed here:


With gems like: A study of rough amplitude quantization by means of Nyquist sampling theory (1956) I bet someone'll get a few days reading outta this.

I have seen somebody accessing a grading sheet from Stanford. The directory was indexed on shodan. I wonder what that was all about.

Bernard Widrow has been busy :)

I wonder why it is that articles written from way back when very often seem to have a way higher average quality than the stuff written today. I'm not sure the sole reason is that the only thing that survives is the high quality stuff, or because anyone can publish anything on the web, thus lowering the bar. It feels like even the peak of what we find today is rarely as good as the average found back then.

There is an issue of survivor bias, but people with decades of experience tend to think that quality has been going downhill. If you're at a university or institution that provides archives, try scanning journals an issue at a time.

I imagine that fields vary, but in the fields that I've worked in, the quality has gone down through the decades. Part of this is increasing pressure to publish, but another factor is the growth of fields. Growth works out very well if it is natural, with brilliant professors attracting brilliant students. But problems arise when deans and funding agencies judge professors by how many people they supervise. The pool of applicants has to be pretty deep, when professors are expected to supervise 3+ students each, and those students are expected to start doing the same after they graduate.

Many research fields come from nowhere, grow exponentially for a while, and then either decline or find a non-research way to reach a steady state (e.g. entertaining undergraduates in service classes). The early stages have high-quality work, or subsequent stages don't occur. But quality might be expected to decrease in those subsequent stages, simply because the people doing the work were not selected as much for their quality and the potential for success, but rather to keep flow going in an academic pipeline.

It also seems reasonable to put all of this in the larger context of declining levels of intellectual discourse and literacy through the years.

Nowadays, anyone who has a Medium account and took a single deep learning course can write about it and establish "authority". That's why.

Beyond survivorship bias, what was publication pressure like back then? I’ve gotten the impression that things used to be slower – tied to journal or conference schedules – and the job market wasn’t as cutthroat or focused on numbers.

I think survivorship bias, as you mention, is pretty powerful.

...maybe it's because today the higher quality ones are about stuff most of us don't understand? When a field is young you have "quality at the bottom level", but it's not really young anymore.

Rumelhart is a consistently good writer. In general though I think you're right and it's just a survival process that makes old articles seem high quality.

Rumelhart is the co-author of a multi volume work called "Parallel Distributed Processing". PDP was the old term in use when Rumelhart and McClelland repopularized artificial neural network techniques in the 80s. Interestingly enough, rumelhart and McClelland were cognitive scientists by training, not computer scientist ones.

I think the PDP books are super interesting, have you read Jeffery Elman's Rethinking Innateness kind of a successor to PDP. Elman did some of the first work on recurrent neural networks I think

Geoff Hinton, some of who's work was in PDP, was an experimental psychologist by training.

Yes, Elman also worked on recurrent neural networks applied to language/text. A lot of word2vec stuff doesn’t seem too far off when you read the articles on clustering of recurrent (hidden) layer activations.

Shockingly easy to read, explained in plain language. I didn't remember neural net articles being so accessible, especially old ones.

Hints for successful application are funny to read. Each and every word in there has been the subject of tons of papers.

* spotted Hinton in references, hehe, he was famous when neural nets were a joke

This explains the basic ideas really well without messing up with implementations and frameworks.

This is nice

Applications are open for YC Summer 2019

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact