[HN Gopher] Deep learning model compression methods
       ___________________________________________________________________
        
       Deep learning model compression methods
        
       Author : singhrac
       Score  : 75 points
       Date   : 2021-04-05 00:57 UTC (22 hours ago)
        
 (HTM) web link (rachitsingh.com)
 (TXT) w3m dump (rachitsingh.com)
        
       | bitforger wrote:
       | I wrote a blog post a while back about whether methods like these
       | will be necessary in the future:
       | 
       | http://mitchgordon.me/machine/learning/2020/01/13/do-we-real...
        
       | mkaic wrote:
       | Excellent write-up, I learned some new things today!
        
       | mkaic wrote:
       | It seems absolutely wild to me that there are cases where you can
       | prune 90% of neurons and still get a functional network out the
       | other side. I'm definitely interested in experimenting with some
       | of these techniques myself soon!
        
         | sillysaurusx wrote:
         | It's not so strange if you visualize the actual weights of a
         | model. Notice that most of the visualizations on
         | https://battle.shawwn.com/sdb/visualizations/2020-09-16-117m...
         | are solid colors.
        
           | yorwba wrote:
           | If you zoom in until you can distinguish individual pixels,
           | it's clear that they aren't actually solid colors. It's just
           | that all the variability is at the pixel level and there are
           | no large-scale structures. (There couldn't possibly be,
           | because the ordering of rows and columns is arbitrary: you
           | could shuffle them around and get an equivalent model. The
           | most you could expect is some banding where values in the
           | same row or column are related.)
        
         | mjburgess wrote:
         | Measure the position of a pen relative to 1000 other objects in
         | your room.
         | 
         | Move the pen.
         | 
         | How many of those measurements co-vary? Nearly all of them.
         | 
         | Can you reduce that down to a subset of measurements that vary
         | always independently?
         | 
         | With care, you only need 3.
        
           | robertacion wrote:
           | This is very interesting, can you recommend any link where I
           | can learn more about this concept?
        
             | mjburgess wrote:
             | ggl factor analysis
             | 
             | also: degree of freedom, dimension
        
         | elteto wrote:
         | Please go easy on yourself...
        
           | perl4ever wrote:
           | "A man with an unusually tiny brain manages to live an
           | entirely normal life despite his condition, which was caused
           | by a fluid build-up in his skull.
           | 
           | Scans of the 44-year-old man's brain showed that a huge
           | fluid-filled chamber called a ventricle took up most of the
           | room in his skull, leaving little more than a thin sheet of
           | actual brain tissue"
           | 
           | An article goes on to say that he was tested at an IQ of 75,
           | below normal but not disabled; he was employed as a civil
           | servant, and married with children.
           | 
           | https://web.archive.org/web/20150712092909if_/http://www.new.
           | ..
        
             | mjburgess wrote:
             | biology has no relevance here
             | 
             | why the body can be regulated with a subset of brain
             | function has nothing to do with why parameters covary in
             | large statistical models
        
       | posharma wrote:
       | I'm surprised there's no mention of sparse formats to store the
       | weights of the model.
        
         | antixk wrote:
         | Becuase unless you have a certain (high) level of sparsity,
         | sparse formats are infact ineffective in storing. There are
         | cases where sparse formats take more memory than storing dense
         | tensors.
        
           | posharma wrote:
           | Sure. It's still one of the methods to compress models. You
           | might be surprised how often the weights are really quite
           | sparse.
        
       | woliveirajr wrote:
       | Always thought of the opposite: DL model as a compression method.
       | 
       | Like taking a initial frame of a movie and training a DL to
       | reproduce the following frames. And then transmitting the frame +
       | parameters to reproduce the movie.
        
         | unemphysbro wrote:
         | Yeah, I've thought something along those lines, maybe applied
         | to a layer(s)/architectural element-wise basis. I'm sure
         | someone's done it because it; just seems like another
         | optimization problem.
        
         | abyesilyurt wrote:
         | It is not video, but I have worked on image compression using
         | neural networks on my master thesis.
         | 
         | It was an improvement over an existing method, so there is
         | quite a bit of research going on in this area.
         | 
         | Briefly, the idea was to use an auto-encoder to transform the
         | image, then quantize and encode the transform coefficients. So,
         | you can actually "teach" the network to be resilient against
         | quantization operation. Very similar to what the author
         | describes.
        
         | astrange wrote:
         | Deep learning for video compression seems like it'll have a
         | natural advantage just because video codecs have never been
         | particularly large (maybe 100 KBs?) but everyone expects an ML
         | model to be enormous. That's a lot of space to store common
         | data in.
        
         | wenc wrote:
         | That's what NVIDIA Maxine does.
         | 
         | https://www.dpreview.com/news/5756257699/nvidia-research-dev...
        
         | chestervonwinch wrote:
         | Any dimensionality reduction method (DL-based or otherwise) can
         | be viewed as a form of compression. But, the end goal in data
         | dimensionality reduction is usually quite different than with
         | data compression.
         | 
         | Either way, this is much different than model
         | compression/distillation, which is compression of the
         | parametrized functional mapping itself. As a silly example,
         | imagine you fit a 100-degree polynomial to noisy linear data
         | using some proper regularization. You would find that can
         | distill/compress your 100-degree polynomial model into a
         | 1-degree model with comparable accuracy.
        
         | fantod wrote:
         | Maybe this could be the beginning of a plausible explanation
         | for why knowledge distillation works.
        
       ___________________________________________________________________
       (page generated 2021-04-05 23:03 UTC)