[HN Gopher] ML training compute has been doubling every 6 months...
       ___________________________________________________________________
        
       ML training compute has been doubling every 6 months since 2010
        
       Author : innerfestival
       Score  : 84 points
       Date   : 2022-02-15 16:44 UTC (6 hours ago)
        
 (HTM) web link (twitter.com)
 (TXT) w3m dump (twitter.com)
        
       | egberts1 wrote:
       | 10B?
       | 
       | - Is that 2 (as in binary 10)?
       | 
       | - 108 as in a typo, or
       | 
       | - an order of magnitude of 6 thereof?
        
         | danieldk wrote:
         | 10 billion. The number of FLOPs has gone from ~1e+14 (6 layer
         | MLP MNIST) to ~1e24 (Megatron-turing NLG).
         | 
         | edit: s/parameters/FLOPs
        
           | Tamaybes wrote:
           | Not parameters, the amount of FLOPS required to train the
           | model.
        
             | danieldk wrote:
             | Whoops, thanks! That's what I get for typing too quickly
             | between household duties.
        
         | dang wrote:
         | HN's software replaces "10 billion" with "10B". Most of the
         | time it's clear.
         | 
         | However, I've replaced the submitted title ('ML training
         | compute has been growing by a factor of 10B since 2010') with
         | what the tweet says now.
        
       | gilbetron wrote:
       | Machine learning related, the transcription capability of Zoom
       | (which uses some form of machine learning) was amazing me this
       | morning. I was watching a recording of a zoom session that was a
       | mock interview to help learning a new interview question. In this
       | situation, there were two people sitting by each other in an
       | echo-y conference room, being recorded by a zoom tablet thing
       | microphone. They were talking to each other, both were french but
       | speaking english and so had moderately heavy accents, both had
       | masks on, and were talking about a coding problem. The accuracy
       | of the transcription was blowing my mind. Really the only time it
       | stumbled was for very technical terms.
        
       | Datenstrom wrote:
       | While this is true there is also a lot of interesting research
       | going on that makes model training and adaptation to new tasks
       | much more efficient. For example in meta-learning methods like
       | Model Agnostic Meta-Learning (MAML)[1] you learn a model that
       | learns to initialize weights for another model such that it
       | converges very quickly for a new task.
       | 
       | For datasets benchmarking new task adaptation there are Meta-
       | Dataset (MD)[2] for Meta-Learning, and the Visual Task Adaptation
       | Benchmark (VTAB)[3] for Representation Learning. Recently to
       | compare the two approaches VTAB+MD[4] was created. Of course,
       | there is also model quantization[5], mixed precision, sparse
       | networks[6], brain floats[7].
       | 
       | Well I've dropped a lot there, but there is a lot more outside
       | the mainstream. I've been building something on a tight budget
       | for a few years so this has been important for success there. We
       | are are very focused on being data and compute efficient.
       | 
       | [1]: https://proceedings.mlr.press/v70/finn17a/finn17a.pdf
       | 
       | [2]: https://arxiv.org/abs/1903.03096
       | 
       | [3]: https://arxiv.org/abs/1910.04867
       | 
       | [4]: https://openreview.net/pdf?id=Q0hm0_G1mpH
       | 
       | [5]: https://arxiv.org/abs/2105.08819
       | 
       | [6]: https://arxiv.org/abs/2112.13896
       | 
       | [7]: https://en.wikipedia.org/wiki/Bfloat16_floating-point_format
        
       | ellis0n wrote:
       | Growth by 1e+7 per ten years from 1e+14 2011 to 1e+21 2021 and
       | should be 1e+28 2030. Probably Doom, host OS and player bot will
       | be generated and played by AI from 1e+28 2030.
       | 
       | 2011 1e+14 = 100 000 000 000 000
       | 
       | 2021 1e+21 = 1 000 000 000 000 000 000 000 <- you are here
       | 
       | 2031 1e+28 = 10 000 000 000 000 000 000 000 000 000
        
       | joe_the_user wrote:
       | It looks like a useful article showing the trend OpenAI described
       | a few years ago has continued.
       | 
       | One thing I'd object to is coining (or maybe just using?) the
       | term "Large-Scale Era". This seems like polluting the discussion
       | with a meaningless phrase.
       | 
       | The basic message is essentially, deep learning continues to grow
       | in scale exponentially, many people consider the dividends still
       | worth the costs. No need to "sexy things up" with what no more
       | than a marketing term.
        
         | Tamaybes wrote:
         | The result about recent compute trends is different from the
         | recent trends described by OpenAI. In particular, they find a
         | 3.5-month doubling time over the Deep Learning Era, whereas the
         | paper finds a 6-month doubling time.
         | 
         | I think the Large-Scale Era does point to a new phenomenon that
         | emerged pretty discontinuously, which is that there are now
         | 'two lanes' in ML scaling. Prior to 2015, academic and industry
         | would train roughly similarly compute intensive models. Since
         | then, a small number of industry players frequently train
         | models with 10-100x more compute than what the typical
         | researcher uses.
        
           | joe_the_user wrote:
           | The thing is that the advent of deep learning was a very big
           | change in the sense you had a general purpose method appear
           | that you could use to throw computing power at many/most
           | problems (and tune a bit but still) and get results that
           | previously you couldn't get (and when did get results, you
           | required domain experts). No doubt we have changes within the
           | trajectory of this escalating brute force solutions. But
           | relative changes in this paradigm seem fundamentally
           | different than the initial advent of the paradigm.
        
       | axg11 wrote:
       | For those in the field, the progress of model performance feels
       | just as rapid. Language models today are incredibly powerful
       | compared to language models in the early 2010s.
       | 
       | I think the advent of retrieval models (retrieval transformers)
       | will continue this compute trend in a more efficient manner. They
       | allow focusing of the compute onto indexing of the knowledge.
        
       | uses wrote:
       | What's a good next step after learning "classical" neural nets,
       | i.e. backpropagation ANNs?
       | 
       | I've been working with MNIST with C and CUDA with dynamic
       | parallelism for a couple months and it's been extremely
       | enlightening but I'm kind of ready to move on.
       | 
       | CNNs maybe?
        
         | nefitty wrote:
         | What would you recommend for a complete beginner? I found these
         | lectures that have helped me start sketching out the field in
         | my mind, but it's still difficult:
         | https://www.davidsilver.uk/teaching/
        
       | chickenpotpie wrote:
       | https://en.m.wikipedia.org/wiki/Wirth%27s_law
        
         | wirthjason wrote:
         | This is one of my favorite laws!
        
       | Victerius wrote:
       | Why spend billions of dollars on armies of PhDs when you can
       | create a better organic intelligence in 9 months at a fraction of
       | the cost?
       | 
       | Why try to beat nature?
       | 
       | I'm being only semi sarcastic here.
        
         | LesZedCB wrote:
         | because the kind of work we want the machines to do is
         | degrading and monotonous
         | 
         | oscar wilde's _the soul of man under socialism_ probably better
         | describes the idea.
        
           | AnEro wrote:
           | I wish that's why we all are doing this
        
         | spyder wrote:
         | Because it's software and once you created it you can duplicate
         | it, customize it with the fraction of the cost and time.
         | 
         |  _" Why try to beat nature?"_
         | 
         | If we wouldn't try "beat nature" we would be still in caves or
         | would be extinct. Nature is "beating" itself with evolution and
         | we are part of that.
        
         | cossatot wrote:
         | Plus 20 years of intensive training after that 9 months...
         | 
         | Or maybe your newborn is more capable than mine!
        
         | AnEro wrote:
         | Nature's $7 an hour general compute nodes are still too
         | expensive to not burn money to try to make an equivalent at the
         | cost of electricity and initial purchase.
        
         | throwawaynay wrote:
         | AI is already more (cost)efficient at some tasks than humans
        
         | nootropicat wrote:
         | 1. Because technology to clone an existing human (not just the
         | body) doesn't exist, especially into a more durable substrate
         | 
         | 2. Because ai doesn't have biological scaling limits. There's
         | nothing particularly special about human level of intelligence.
        
           | rowanG077 wrote:
           | > There's nothing particularly special about human level of
           | intelligence.
           | 
           | Big claim since nothing in the known universe has human level
           | intelligence besides humans. It's like holding the original
           | declaration of independence and saying that there is nothing
           | particularly special about this piece of paper.
        
             | nefitty wrote:
             | This analogy breaks down easily.
             | 
             | What do you propose is special about human-level
             | intelligence? If anything, it is in constant struggle with
             | biological and emotional needs. Yes, of course, human
             | brains are amazing, the declaration of independence is
             | great, but the entire human project is the push for
             | progress.
        
               | rowanG077 wrote:
               | I can't say what is special about human-level
               | intelligence. Most importantly because the scientific
               | community(and by extension me) can't define what human-
               | level intelligence means. It's just too illusive at this
               | point. That something is undefined by humans doesn't mean
               | you can simply say it's not special.
        
         | mistrial9 wrote:
         | deep, real-time scans of digital media streams, subscription
         | behavior, market responses.. things like that, where the inputs
         | are digital already
        
         | anonymouse008 wrote:
         | I honestly thought you meant waiting for the next iteration
         | doubling of ML model performance/training...
         | 
         | which I thought was brilliant, ha!
        
         | ska wrote:
         | > Why spend billions of dollars on armies ...
         | 
         | The vast majority of AI research (over all time) has not been
         | seriously intended or expected to arrive at a general
         | intelligence.
         | 
         | Sure there are some true believers there in some sort of
         | handwavy construction argument ("and then a miracle occurs")
         | and there have been serious research attempts to understand
         | intelligence, of limited scope an resources. But to be clear,
         | whether this wave or previous ones, the money washes in and the
         | army of PhD's happens when there is a sniff of practical,
         | usable algorithmic results from machine learning. It's not the
         | same thing.
         | 
         | Parallel application of even relatively stupid algorithms
         | targeted at the right tasks will rapidly outproduce any
         | reasonable number of organic intelligences of the type you
         | describe :)
        
         | jacquesm wrote:
         | Because a newborn needs another 18 years or so of nurturing and
         | education to become useful, this costs so much that quite a few
         | countries see no qualms in saddling these new humans with the
         | debt of their education.
        
       | mistrial9 wrote:
       | hehe - DeepLearning that is, my supervised models work fine
       | without a compute infrastructure the size of a major naval vessel
       | thanks!
        
       | itronitron wrote:
       | At what point should this start to have a noticeable result
       | outside of the profession?
        
       ___________________________________________________________________
       (page generated 2022-02-15 23:01 UTC)