A tiny one-dimensional pipeline shows the assigned filter outputs, ReLU clamp, and max-pool value without hiding any arithmetic.
Run the assigned filter
This toy one-dimensional strip is [0, 2, 5, 1]. The assigned filter is [-1, 1]. At each adjacent window, multiply and add exactly.
strip [ 0 , 2 , 5 , 1 ] , filter [ − 1 , 1 ] \text{strip }[0,2,5,1],\quad \text{filter }[-1,1] strip [ 0 , 2 , 5 , 1 ] , filter [ − 1 , 1 ] Filter ReLU Pool Keeps the Signal Visible A toy one-dimensional filter, ReLU, and max pool stay exact. visible sources source values input strip [0, 2, 5, 1] assigned filter [-1, 1] filter then ReLU then max pool step exact arithmetic filter out after ReLU / pool window 1 [0, 2] 0*-1 + 2*1 2 2 window 2 [2, 5] 2*-1 + 5*1 3 3 window 3 [5, 1] 5*-1 + 1*1 -4 0 max pool max(2, 3, 0) 3 filter outputs 2, 3, -4; ReLU outputs 2, 3, 0; pool 3 toy 1-D pipeline; assigned filter weights; one exact forward pass; head softmax is NAMED; NOT training; NOT learning; NOT image understanding; NOT generalization toy 1-D pipeline; assigned filter weights; one exact forward pass; head softmax is NAMED; NOT training; NOT learning; NOT image understanding; NOT generalization
The filter outputs are exact
The adjacent filter outputs are 2, 3, and -4.
Filter ReLU Pool Keeps the Signal Visible A toy one-dimensional filter, ReLU, and max pool stay exact. visible sources source values input strip [0, 2, 5, 1] assigned filter [-1, 1] filter then ReLU then max pool step exact arithmetic filter out after ReLU / pool window 1 [0, 2] 0*-1 + 2*1 2 2 window 2 [2, 5] 2*-1 + 5*1 3 3 window 3 [5, 1] 5*-1 + 1*1 -4 0 max pool max(2, 3, 0) 3 filter outputs 2, 3, -4; ReLU outputs 2, 3, 0; pool 3 toy 1-D pipeline; assigned filter weights; one exact forward pass; head softmax is NAMED; NOT training; NOT learning; NOT image understanding; NOT generalization toy 1-D pipeline; assigned filter weights; one exact forward pass; head softmax is NAMED; NOT training; NOT learning; NOT image understanding; NOT generalization
ReLU keeps positives and clamps the negative
ReLU keeps 2 and 3, then clamps -4 to 0. The ReLU strip is [2, 3, 0].
ReLU ( [ 2 , 3 , − 4 ] ) = [ 2 , 3 , 0 ] \operatorname{ReLU}([2,3,-4])=[2,3,0] ReLU ([ 2 , 3 , − 4 ]) = [ 2 , 3 , 0 ] Filter ReLU Pool Keeps the Signal Visible A toy one-dimensional filter, ReLU, and max pool stay exact. visible sources source values input strip [0, 2, 5, 1] assigned filter [-1, 1] filter then ReLU then max pool step exact arithmetic filter out after ReLU / pool window 1 [0, 2] 0*-1 + 2*1 2 2 window 2 [2, 5] 2*-1 + 5*1 3 3 window 3 [5, 1] 5*-1 + 1*1 -4 0 max pool max(2, 3, 0) 3 filter outputs 2, 3, -4; ReLU outputs 2, 3, 0; pool 3 toy 1-D pipeline; assigned filter weights; one exact forward pass; head softmax is NAMED; NOT training; NOT learning; NOT image understanding; NOT generalization toy 1-D pipeline; assigned filter weights; one exact forward pass; head softmax is NAMED; NOT training; NOT learning; NOT image understanding; NOT generalization
Max pool keeps the largest visible value
Max pool over [2, 3, 0] returns 3. This is a toy one-dimensional pipeline with assigned filter weights. No claim about training, image understanding, or generalization.
max ( [ 2 , 3 , 0 ] ) = 3 \max([2,3,0])=3 max ([ 2 , 3 , 0 ]) = 3 Filter ReLU Pool Keeps the Signal Visible A toy one-dimensional filter, ReLU, and max pool stay exact. visible sources source values input strip [0, 2, 5, 1] assigned filter [-1, 1] filter then ReLU then max pool step exact arithmetic filter out after ReLU / pool window 1 [0, 2] 0*-1 + 2*1 2 2 window 2 [2, 5] 2*-1 + 5*1 3 3 window 3 [5, 1] 5*-1 + 1*1 -4 0 max pool max(2, 3, 0) 3 filter outputs 2, 3, -4; ReLU outputs 2, 3, 0; pool 3 toy 1-D pipeline; assigned filter weights; one exact forward pass; head softmax is NAMED; NOT training; NOT learning; NOT image understanding; NOT generalization toy 1-D pipeline; assigned filter weights; one exact forward pass; head softmax is NAMED; NOT training; NOT learning; NOT image understanding; NOT generalization