Skip to content

← Concept map

Purpose of Optimization Variants

2 explanations from your library.

  1. Gradient Descent Variants and Momentum, at 0:4910:49

    IIT Madras (NPTEL)

    Gradient Descent Variants and Momentum

    “networks converge faster basically deep learning networks converge faster to the optimal solution. Many of all of almost all of the algorithms are like you can use them like black boxes in most of the packages that I introduced last time.”

    Open full session

    Moment · 0:49

    Gradient Descent Variants and Momentum

    Full session
  2. Learning Rate Decay, at 1:5221:52

    IIT Madras (NPTEL)

    Learning Rate Decay

    “So, this is again decided by you, you have to again this is one consider this as another hyper parameter that you have to optimize, right. Another way would be to check the valid, so you have validation data and training data, so you check the performance on the validation data and whenever”

    Closest moment in this session: it teaches the idea without listing it as a key concept.

    Open full session

    Moment · 1:52

    Learning Rate Decay

    Full session
Cloudinary under the hood2 Cloudinary URLs make this page. No render servers: each one is generated on request and cached.

  • Compare reel

    2 explanations of this concept, from different sessions, spliced into one video with fl_splice. Each clip is labelled with its speaker.

    Open ↗

    …/video/upload/so_47.5,eo_62.7,w_1280,h_720,c_fill/l_video:pravaha:2b17046e-fb2d-4a2c-ad2c-e84a76f8c28a,fl_spl…/so_110.7,eo_128.6,w_1280,h_720,c_fill/fl_layer_apply/fl_layer_apply,g_north_west,x_40,y_40,so_15.2,eo_33.1/f_auto:video,q_auto/pravaha/ff2c7367-5718-4b86-9d17-04328cf12206.mp4

    so_47.5
    starts at 47.5 s · and 3 more like it
    eo_62.7
    ends at 62.7 s · and 3 more like it
    w_1280
    1280 px wide · and 1 more like it
    h_720
    720 px tall · and 1 more like it
    c_fill
    crops to fill the frame exactly · and 1 more like it
    l_video:pravaha:2b17046e-fb…
    another video (pravaha/2b17046e-fb2d-4a2c-ad2c-e84a76f8c28a) as a layer
    fl_splice
    joins the next clip onto the end of this one
    fl_layer_apply
    places the layer defined just before · and 2 more like it
    l_text:arial_34_bold:1%20%C…
    text layer “1 · IIT Madras (NPTEL)” · and 1 more like it
    co_white
    text colour white · and 1 more like it
    b_rgb:0f766ecc
    background #0f766ecc · and 1 more like it
    g_north_west
    anchored to the top-left corner · and 1 more like it
    x_40
    40 px from the side · and 1 more like it
    y_40
    40 px from the edge · and 1 more like it
    f_auto:video
    best format for each device (e.g. AV1, WebM, MP4, WebP)
    q_auto
    AI-chosen quality: smallest file that still looks right

  • Explanation thumbnail

    The frame at that moment, cropped around the speaker by AI.

    Open ↗

    …/video/upload/so_49,c_fill,ar_16:9,w_640,g_auto/f_auto,q_auto/pravaha/ff2c7367-5718-4b86-9d17-04328cf12206.jpg

    so_49
    starts at 49 s
    c_fill
    crops to fill the frame exactly
    ar_16:9
    aspect ratio 16 : 9
    w_640
    640 px wide
    g_auto
    AI picks the focus (the speaker or slide), not a blind centre crop
    f_auto
    best format for each device (e.g. AV1, WebM, MP4, WebP)
    q_auto
    AI-chosen quality: smallest file that still looks right

Want a guided order, basics first? Turn Purpose of Optimization Variants into a short course →