CS7643-QUIZ2 EXAM WITH ALL
NEEDED TO PASS
Convolution XFeatures X- Xcorrect Xanswers X-edges
colors
textures
motifs X(corners, Xshapes)
Receptive Xfield X- Xcorrect Xanswers X-A Xregion Xof Xan Ximage X(image Xpatch) Xfrom
Xwhich Xthe Xnode Xreceives Xinput. XUsually Xdenoted Xby Xa XK1 Xx XK2 Xmatrix.
Convolution Xvs XCross-correlation X- Xcorrect Xanswers X-Convolution: Xflip Xthe Xkernel
X(rotate X180) Xand Xtake Xthe Xdot Xproduct Xwith Ximage Xpatch
Cross-correlation: Xdo Xnot Xflip Xthe Xkernel Xto Xtake Xthe Xdot Xproduct Xwith Ximage
Xpatch
Advantage Xof Xusing Ximage Xpatch X- Xcorrect Xanswers X-1./ XReduces Xthe Xinput
Xparameters Xto X
, K1 Xx XK2 X+ X1 X(bias) X
for Xeach Xoutput Xnode. XThus, Xthe Xtotal Xnumber Xof Xinput Xparameters:
N Xx X(K1 X+ XK2 X+ X1)
2./ XExplicitly Xmaintains Xspatial Xinformation
Weight Xsharing X- Xcorrect Xanswers X-The Xweights Xwill Xrepresent Xwhat Xtypes Xof
Xfeatures Xwe Xwill Xextract. XThe Xweights X(W) Xwill Xbe Xthe Xsame Xfor Xeach Xoutput
Xnode Xwith Xrespect Xto Xa Xspecific Xkernel, Xregardless Xof Xthe Xspecific Ximage
Xpatch Xwe Xare Xlooking Xat. X
The Xtotal Xnumber Xof Xinput Xparameters: X
K1 Xx XK2 X+ X1
Input Xparameters Xwith Xmultiple Xfeature Xextractions X- Xcorrect Xanswers X-(K1 Xx XK2
X+ X1) Xx XM
where XM Xis Xthe Xnumber Xof Xfeatures
Relationship Xbetween Xconvolution Xand Xcross-correlation X- Xcorrect Xanswers X-
Duality: XIf Xcross-correlation Xis Xthe Xforward Xpass X(which Xis Xthe Xeasier
Xoperation), Xthe Xconvolution Xoperation Xis Xgoing Xto Xbe Xthe Xbackward Xpass Xto
Xcalculate Xgradients X(vice Xversa)
Valid Xconvolution X- Xcorrect Xanswers X-When Xthe Xkernel Xis Xfully Xon Xthe Ximage.
X(No Xpadding)
Output Xsize Xof Xthe Xvanilla Xconvolution, X
given XH, XW, XK1, XK2 X- Xcorrect Xanswers X-(H X- XK1 X+ X1) Xx X(W X- XK2 X+ X1)
How Xto Xadd Xpadding X- Xcorrect Xanswers X-Increases Xthe Xsize Xof Xthe Ximage Xwith
XP Xin Xboth Xdirections X(top X& Xbottom, Xleft X& Xright) X
--> X(H X+ X2P) Xx X(W X+ X2P)
Can Xbe Xfilled Xwith Xzeros Xor Xmirror Xthe Ximage
Stride Xand Xits Xconsequences X- Xcorrect Xanswers X-Number Xof Xpixels Xmoving
Xforward Xwhen Xparsing Xthe Xpatch Xthrough Ximages.
Loss Xof Xinformation
Used Xfor Xdimensionality Xreduction
Effect Xof Xchannels Xon Xoutput Xsize X- Xcorrect Xanswers X-It Xdoesn't Xhave Xeffect
Xon Xthe Xoutput Xsize: Xwe Xperform Xthe Xdot Xproduct Xfor Xeach Xchannels Xand
Xsumming Xthem Xup.
Effect Xof Xchannels Xon Xparameters X- Xcorrect Xanswers X-Each Xchannel Xmight
Xhave Xits Xown Xweights Xwith Xrespect Xto Xthe Xsame Xkernel.
M Xx X(Ch Xx XK1 Xx XK2 X+ X1)
NEEDED TO PASS
Convolution XFeatures X- Xcorrect Xanswers X-edges
colors
textures
motifs X(corners, Xshapes)
Receptive Xfield X- Xcorrect Xanswers X-A Xregion Xof Xan Ximage X(image Xpatch) Xfrom
Xwhich Xthe Xnode Xreceives Xinput. XUsually Xdenoted Xby Xa XK1 Xx XK2 Xmatrix.
Convolution Xvs XCross-correlation X- Xcorrect Xanswers X-Convolution: Xflip Xthe Xkernel
X(rotate X180) Xand Xtake Xthe Xdot Xproduct Xwith Ximage Xpatch
Cross-correlation: Xdo Xnot Xflip Xthe Xkernel Xto Xtake Xthe Xdot Xproduct Xwith Ximage
Xpatch
Advantage Xof Xusing Ximage Xpatch X- Xcorrect Xanswers X-1./ XReduces Xthe Xinput
Xparameters Xto X
, K1 Xx XK2 X+ X1 X(bias) X
for Xeach Xoutput Xnode. XThus, Xthe Xtotal Xnumber Xof Xinput Xparameters:
N Xx X(K1 X+ XK2 X+ X1)
2./ XExplicitly Xmaintains Xspatial Xinformation
Weight Xsharing X- Xcorrect Xanswers X-The Xweights Xwill Xrepresent Xwhat Xtypes Xof
Xfeatures Xwe Xwill Xextract. XThe Xweights X(W) Xwill Xbe Xthe Xsame Xfor Xeach Xoutput
Xnode Xwith Xrespect Xto Xa Xspecific Xkernel, Xregardless Xof Xthe Xspecific Ximage
Xpatch Xwe Xare Xlooking Xat. X
The Xtotal Xnumber Xof Xinput Xparameters: X
K1 Xx XK2 X+ X1
Input Xparameters Xwith Xmultiple Xfeature Xextractions X- Xcorrect Xanswers X-(K1 Xx XK2
X+ X1) Xx XM
where XM Xis Xthe Xnumber Xof Xfeatures
Relationship Xbetween Xconvolution Xand Xcross-correlation X- Xcorrect Xanswers X-
Duality: XIf Xcross-correlation Xis Xthe Xforward Xpass X(which Xis Xthe Xeasier
Xoperation), Xthe Xconvolution Xoperation Xis Xgoing Xto Xbe Xthe Xbackward Xpass Xto
Xcalculate Xgradients X(vice Xversa)
Valid Xconvolution X- Xcorrect Xanswers X-When Xthe Xkernel Xis Xfully Xon Xthe Ximage.
X(No Xpadding)
Output Xsize Xof Xthe Xvanilla Xconvolution, X
given XH, XW, XK1, XK2 X- Xcorrect Xanswers X-(H X- XK1 X+ X1) Xx X(W X- XK2 X+ X1)
How Xto Xadd Xpadding X- Xcorrect Xanswers X-Increases Xthe Xsize Xof Xthe Ximage Xwith
XP Xin Xboth Xdirections X(top X& Xbottom, Xleft X& Xright) X
--> X(H X+ X2P) Xx X(W X+ X2P)
Can Xbe Xfilled Xwith Xzeros Xor Xmirror Xthe Ximage
Stride Xand Xits Xconsequences X- Xcorrect Xanswers X-Number Xof Xpixels Xmoving
Xforward Xwhen Xparsing Xthe Xpatch Xthrough Ximages.
Loss Xof Xinformation
Used Xfor Xdimensionality Xreduction
Effect Xof Xchannels Xon Xoutput Xsize X- Xcorrect Xanswers X-It Xdoesn't Xhave Xeffect
Xon Xthe Xoutput Xsize: Xwe Xperform Xthe Xdot Xproduct Xfor Xeach Xchannels Xand
Xsumming Xthem Xup.
Effect Xof Xchannels Xon Xparameters X- Xcorrect Xanswers X-Each Xchannel Xmight
Xhave Xits Xown Xweights Xwith Xrespect Xto Xthe Xsame Xkernel.
M Xx X(Ch Xx XK1 Xx XK2 X+ X1)
