UMAP Uniform Manifold Approximation and Projection for Dimension Reduction | SciPy 2018 |

Поділитися
Вставка
  • Опубліковано 15 тра 2024
  • This talk will present a new approach to dimension reduction called UMAP. UMAP is grounded in manifold learning and topology, making an effort to preserve the topological structure of the data. The resulting algorithm can provide both 2D visualisations of data of comparable quality to t-SNE, and general purpose dimension reduction. UMAP has been implemented as a (scikit-learn compatible) python library that can perform efficient dimension reduction, scaling out to much larger datasets than t-SNE or other comparable algorithms (see github.com/lmcinnes/umap).
    See the full SciPy 2018 playlist here: • SciPy 2018: Scientific...
  • Наука та технологія

КОМЕНТАРІ • 56

  • @MegaStatis
    @MegaStatis 5 років тому +13

    Thanks for the nice and clean explanation for UMAP!

  • @kentvandervelden
    @kentvandervelden 5 років тому +11

    Wow, fantastic talk!! Trying out several ideas from the talk immediately.

  •  5 років тому +14

    Greatest ML innovation of the last decade.

  • @dwinsemius
    @dwinsemius 4 роки тому +4

    Nice explanation of connection between topological notions and algorithm implementation.

  • @FougaFrancois
    @FougaFrancois 5 років тому +1

    I will be trying that shortly! Fantastic ideas !

  • @jukebox54
    @jukebox54 5 років тому +22

    Thanks for the Python implementation!

  • @shakedg2956
    @shakedg2956 3 роки тому

    Loved the talk! Thank you!

  • @soothingTrench
    @soothingTrench 5 років тому +4

    Amazing talk

  • @kvafsu225
    @kvafsu225 Рік тому

    Very illustrative presentation. Thanks

  • @mishalthapa4643
    @mishalthapa4643 Рік тому

    wonderful and informative presentation!

  • @inflivia
    @inflivia 2 роки тому +1

    Cool talk, well delivered.

  • @9164312
    @9164312 5 років тому +6

    Amazing,Genius!!

  • @matveyshishov
    @matveyshishov Рік тому

    Loved it so much, thanks!

  • @avacdoos3101
    @avacdoos3101 11 місяців тому

    Amazingly Explained

  • @unoqualsiasi7341
    @unoqualsiasi7341 5 років тому +1

    Amazing, thanks!

  • @NoNTr1v1aL
    @NoNTr1v1aL 2 роки тому +1

    Amazing video!

  • @maziyarbaranpouyan3543
    @maziyarbaranpouyan3543 5 років тому +3

    Perfect Talk!!!

  • @almeidaalin3
    @almeidaalin3 4 роки тому

    very clear, thank you!

  • @phpngpl
    @phpngpl 5 років тому +2

    You are a great golfer.

  • @itsme.junaway
    @itsme.junaway 5 років тому +10

    This is too good to be true!

  • @asishswain1259
    @asishswain1259 2 роки тому

    incredible !

  • @user-sv5vb1mj1q
    @user-sv5vb1mj1q 3 роки тому

    great success, my congrats

  • @ValeryKhamenya
    @ValeryKhamenya 3 роки тому

    Brilliant!!!

  • @kochasaito4628
    @kochasaito4628 3 роки тому

    The Singular simplices also permitted, but the homotopy quotient is required there.

  • @AnIceCrasher
    @AnIceCrasher 4 роки тому

    Supercool, thank you :D

  • @ProfessionalTycoons
    @ProfessionalTycoons 5 років тому +1

    amazing

  • @ryanbae9563
    @ryanbae9563 5 років тому

    Good talk.

  • @BadGatewayMusic
    @BadGatewayMusic 5 років тому

    fantastic

  • @MaskedPixel
    @MaskedPixel 3 місяці тому

    Yup. Using this one tomorrow.

  • @kelixoderamirez
    @kelixoderamirez Рік тому

    permission to learn sir. thank you

  • @vedantsrivastava5754
    @vedantsrivastava5754 5 років тому +8

    I am using the UMAP for Dimensionality Reduction. It's implementation really bods well with my data. I have a question regarding the feature importance. Suppose, after the dimensionality reduction, the UMAP separates out the labels clearly and creates the 3 subgroups containing all labels. Can I somehow get the list of features corresponding to each group? Which features were more relevant to
    which group?

    • @lelandmcinnes9501
      @lelandmcinnes9501 5 років тому +7

      One option that I have seen done is to build a classifier based on the clusters and then use SHAP or LIME to get interpretations of what separates the clusters.

    • @vedantsrivastava5754
      @vedantsrivastava5754 4 роки тому +1

      Thanks @@lelandmcinnes9501 , it is really helpful. Now I am looking for features contribution to each classes ? Does the earlier question of mine make any sense? With tuning the different hyperparamter , I will get different groups...Does these 2 dimensional cluster have any meaning of original features?

  • @endingalaporte
    @endingalaporte 3 роки тому

    very very very nice thank you so much

  • @aug_st
    @aug_st 3 роки тому

    Brilliant

  • @weiguo9183
    @weiguo9183 7 місяців тому

    Does anyone know which kind of normalized distance is used in 11:29?

  • @jaritos675
    @jaritos675 5 років тому +1

    Does anyone have a reference for the triplet networks he talks about?

  • @argolo9512
    @argolo9512 4 роки тому +1

    Whats new about UMAP as compared to other TDA methods? The approach is just like Gunnar Carlson's mapper

  • @kochasaito4628
    @kochasaito4628 3 роки тому

    How deal with the obstruction?

  • @shubhamtalks9718
    @shubhamtalks9718 2 роки тому

    Wow!

  • @Level6
    @Level6 3 роки тому +1

    19:10 t-sne vs UMAP

  • @a.opterix3576
    @a.opterix3576 2 роки тому +1

    Do we know why it is the case, that the local connectivity assumption solves the curse of dimensionality?

    • @marcomoreno6748
      @marcomoreno6748 3 місяці тому

      Does anyone have something they'd like to share on this question?

  • @chitralalawat8106
    @chitralalawat8106 4 роки тому

    The sinosuidal type wave only comes if you've 2D data.. & I think when you started, you said that it's a high dimensional data?

  • @just4onecomment
    @just4onecomment 3 роки тому

    Why does he say the "distributions look nice" at 12:10 ? They look pretty close/overlapping and indistinguishable to me

    • @xy9439
      @xy9439 2 роки тому

      That's the point

  • @user-mv3im2fi4f
    @user-mv3im2fi4f 19 днів тому

    Comecei sem entender nada, terminei sem entender nada e mais um pouco

  • @tinakeshav1576
    @tinakeshav1576 3 роки тому

    Why are we taking 20 nearest neighbours? ~ 12:00

  • @aug_st
    @aug_st 3 роки тому

    Never using tSNE again 😎

  • @frankdimeglio8216
    @frankdimeglio8216 3 роки тому

    THE ULTIMATE, THE CLEAR, AND THE TRUE MATHEMATICAL UNIFICATION REGARDING PHYSICS/PHYSICAL EXPERIENCE IS PROVEN:
    Time DILATION proves that electromagnetism/ENERGY IS GRAVITY, AS E=mc2 is DIRECTLY and fundamentally derived from F=ma. Indeed, gravity/acceleration involves BALANCED inertia/INERTIAL RESISTANCE; AS ELECTROMAGNETISM/energy is gravity. Therefore, the rotation of WHAT IS THE MOON matches it's revolution. Moreover, objects fall at the SAME RATE (neglecting air resistance, of course); AS ELECTROMAGNETISM/ENERGY IS GRAVITY. (Gravity IS ELECTROMAGNETISM/energy. "Mass"/ENERGY IS GRAVITY. ELECTROMAGNETISM/energy is gravity.) E=mc2 IS F=ma. The stars AND PLANETS are POINTS in the night sky. (Very importantly, outer "space" involves full inertia; AND it is fully invisible AND black.) INSTANTANEITY is thus FUNDAMENTAL to what is the FULL and proper understanding of physics/physical experience. This NECESSARILY represents, INVOLVES, AND DESCRIBES what is possible/potential AND actual IN BALANCE, AS ELECTROMAGNETISM/ENERGY IS GRAVITY. SO, time is NECESSARILY possible/potential AND actual IN BALANCE; AS ELECTROMAGNETISM/ENERGY IS GRAVITY. BALANCE and completeness go hand in hand. Thoughts are invisible. THINK. NOW, the INTEGRATED EXTENSIVENESS of THOUGHT (AND description) is improved in the truly superior mind. INDEED, the ability of THOUGHT to describe OR reconfigure sensory experience is ULTIMATELY dependent upon the extent to which THOUGHT IS SIMILAR TO sensory experience. I have explained the fourth spatial dimension, as E=mc2 IS F=ma; as this then also explains the significance of c4. Gravity IS ELECTROMAGNETISM/ENERGY. It is proven. Again, BALANCE and completeness go hand in hand. It ALL CLEARLY makes perfect sense. Great !!!
    E=mc2 IS F=ma. Very carefully consider what is the Moon. I have explained perpetual motion. The planets are not falling in what is "curved" "SPACE" in relation to what is the Sun. SO, a given PLANET sweeps out equal areas in equal times; AS E=mc2 IS F=ma. Consider the man who is standing on what is the Earth/ground. Touch AND feeling BLEND, AS ELECTROMAGNETISM/ENERGY IS GRAVITY. OVERLAY WHAT IS THE EYE in BALANCED RELATION to/with what is the Earth. E=mc2 IS F=ma. LOOK at what is the translucent, semi-spherical, AND BLUE sky. The Earth is ALSO blue ! Great. It is ALL CLEARLY proven. ELECTROMAGNETISM/energy is gravity. Gravity IS ELECTROMAGNETISM/ENERGY. Gravity/acceleration involves BALANCED inertia/INERTIAL RESISTANCE, AS E=mc2 IS F=ma; AS ELECTROMAGNETISM/ENERGY IS GRAVITY. Again, the stars AND PLANETS are POINTS in the night sky. Great !!! BALANCE and completeness go hand in hand. THINK !!! "Mass"/energy is gravity. ELECTROMAGNETISM/energy is gravity. The tides NECESSARILY constitute a BALANCED electromagnetic/gravitational manifestation, form, or effect. In fact, stellar clustering proves that electromagnetism/ENERGY IS GRAVITY.
    By Frank DiMeglio

  • @chilliking3424
    @chilliking3424 3 роки тому

    fantastic

  • @uzulim9234
    @uzulim9234 3 роки тому

    amazing