[HN Gopher] The Fractured Entangled Representation Hypothesis
       ___________________________________________________________________
        
       The Fractured Entangled Representation Hypothesis
        
       Author : akarshkumar0101
       Score  : 45 points
       Date   : 2025-05-20 15:56 UTC (7 hours ago)
        
 (HTM) web link (github.com)
 (TXT) w3m dump (github.com)
        
       | akarshkumar0101 wrote:
       | Tweet: https://x.com/kenneth0stanley/status/1924650124829196370
       | Arxiv: https://arxiv.org/abs/2505.11581
        
         | pvg wrote:
         | Sounds like you're one of the co-authors? Probably worth
         | mentioning if the case so people know they can discuss the work
         | with one of the work-doers.
        
           | akarshkumar0101 wrote:
           | I mentioned that in the original post, but I don't see that
           | text here anymore (thats why I added links via comment)... I
           | am new to hackernews
        
             | messe wrote:
             | I believe they just mean that you should edit the comment
             | where you added the links to mention that you are the
             | author, to add that additional context.
        
               | pvg wrote:
               | I just meant 'it's good for people to know one of the
               | authors is in the thread because it makes for more
               | interesting conversation'. Clearly did not figure out how
               | to do that without starting a bunch of meta!
        
             | macintux wrote:
             | I believe this could (or should) have been a Show HN, which
             | would have allowed you to include explanatory text. See the
             | top of this page for the rules.
             | 
             | https://news.ycombinator.com/show
             | 
             | Welcome to the site. There are a lot of features which are
             | less obvious, which you'll discover over time.
        
               | pvg wrote:
               | Reading material usually can't be a Show HN but you can
               | just post your work without that and say you're involved.
        
               | macintux wrote:
               | The repo includes runnable code.
               | 
               | > Show HN is for something you've made that other people
               | can play with... On topic: things people can run on their
               | computers or hold in their hands
        
               | pvg wrote:
               | A lot of writing includes runnable code and isn't a Show
               | HN. It's a comparatively narrow category.
        
         | ipunchghosts wrote:
         | I am interested in doing research like this. Is there any way I
         | can be a part of it or a similar group? I have been fighting
         | for funding from DoD for many years but to no avail so I
         | largely have to do this research on my own time or solve my
         | current grant's problems so that i can work on this. In my
         | mind, this kind of research is the most interesting and
         | important right now in the deep learning field. I am a hard
         | worker and a high-throughput thinking... how can i get
         | connected to otherwise with a similar mindset?
        
       | scarmig wrote:
       | Did you investigate other search processes besides SGD? I'm
       | thinking of those often termed "biologically plausible" (e.g.
       | forward-forward, FA). Are their internal representations closer
       | to the fractured or unified representations?
        
       | timewizard wrote:
       | > Much of the excitement in modern AI is driven by the
       | observation that scaling up existing systems leads to better
       | performance.
       | 
       | Scaling up almost always leads to better performance. If you're
       | only getting linear gains though then there is absolutely nothing
       | to be excited about. You are in a dead end.
        
       | goldemerald wrote:
       | This is an interesting line of research but missing a key aspect:
       | there's (almost) no references to the linear representation
       | hypothesis. Much work on neural network interpretability lately
       | has shown individual neurons are polysemantic, and therefore
       | practically useless for explainability. My hypothesis is fitting
       | linear probes (or a sparse autoencoder) would reveal linearly
       | semantic attributes.
       | 
       | It is unfortunate because they briefly mention Neel Nanda's
       | Othello experiments, but not the wide array of experiments like
       | the NeurIPS Oral "Linear Representation Hypothesis in Language
       | Models" or even golden gate Claude.
        
         | ipunchghosts wrote:
         | Is what your saying imply that there is a rotation matrix you
         | can apply to each activation output to make it less entangled?
        
           | goldemerald wrote:
           | Not quite. For an underlying semantic concept (e.g., smiling
           | face), you can go from a basis vector [0,1,0,...,0] to the
           | original latent space via a single rotation. You could then
           | induce said concept by manipulating the original latent point
           | by traversing along that linear direction.
        
             | ipunchghosts wrote:
             | I think we are saying the same thing. Please correct me
             | though where I am wrong. You could look at the maps in some
             | way but instead of the basis being one hot dimensions (the
             | standard basis), it could be rotated.
        
               | akarshkumar0101 wrote:
               | We mention this issue exactly in the fourth paragraph in
               | Section 4 and in Appendix F!
        
         | akarshkumar0101 wrote:
         | We mention this issue exactly in the fourth paragraph in
         | Section 4 and in Appendix F!
        
           | goldemerald wrote:
           | That is addressing the incomprehensibility of PCA and
           | applying a transformation to the entire latent space. I've
           | never found PCA to be meaningful for deep learning. As far as
           | I can tell, polysemous issue with neurons cannot be addressed
           | with a single linear transformation. There is no sparse
           | analysis (via linear probes or SAEs) and hence the
           | unaddressed issue.
        
       | ipunchghosts wrote:
       | I am glad they evaluated this hypothesis using weight decay which
       | is primarily thought of to induce a structured representation. My
       | first thought was that the entire paper was useless if they
       | didn't do this experiment.
       | 
       | I find it rather interesting that the structured representations
       | go from sparse to full to sparse as a function of layer depth. I
       | have noticed that applying weight decay penalty as an exponential
       | function of layer depth gives improved results over using a
       | global weight decay.
        
       | cwmoore wrote:
       | Isn't this simply mirroronic gravitation?
        
       ___________________________________________________________________
       (page generated 2025-05-20 23:01 UTC)