Lecture 7: Convolutional Networks

Lecture 7 moves from fully-connected to convolutional networks by introducing new computational primitives that respect the spatial structure of 2D image data. We discuss convolution layers, which slide a learnable filter over the input data. We discuss pooling layers, which spatially downsample their input data. We then look at normalization layers including batch, layer, and instance normalization, which normalize their input data along different axes and improve training speed.
Slides: myumi.ch/K43Zy
_________________________________________________________________________________________________
Computer Vision has become ubiquitous in our society, with applications in search, image understanding, apps, mapping, medicine, drones, and self-driving cars. Core to many of these applications are visual recognition tasks such as image classification and object detection. Recent developments in neural network approaches have greatly advanced the performance of these state-of-the-art visual recognition systems. This course is a deep dive into details of neural-network based deep learning methods for computer vision. During this course, students will learn to implement, train and debug their own neural networks and gain a detailed understanding of cutting-edge research in computer vision. We will cover learning algorithms, neural network architectures, and practical engineering tricks for training and fine-tuning networks for visual recognition tasks.
Course Website: myumi.ch/Bo9Ng
Instructor: Justin Johnson myumi.ch/QA8Pg

[@jh97jjjj]

Great lecture for free. Thank you Michigan University and professor Justin.

[@temurochilov]

Thank you I found answers to the questions that I have been looking for long time

[@faranakkarimpour3794]

Thank you for the great course.

[@jijie133]

Great.

[@tatianabellagio3107]

Amazing!
Pd: Although I am sorry for the guy with the coughing attack...........

[@kobic8]

yeah, kinda disturbed me to concentrate. 2019 it was right before covid striked the world hahah
😷

[@alokoraon1475]

I have this great package for my university course.❤

[@eurekad2070]

Thank you for exellent video! But I have a question here, at 1:05:42, after layer normalization, every sample in x has shape 1xD, while μ has shape Nx1. How do you perform the subtraction x-μ?

[@yicheng1991]

I wonder if gamma and beta with 1 x D is a typo? If it should be N x 1? If it is not a typo, doing the subtraction is just using the broadcasting mechanism like in numpy.

[@eurekad2070]

@@yicheng1991 Broadcasting mechanism makes sense. Thank you.

[@puranjitsingh1782]

Thanks for an excellent video Justin!! I had a quick question on how does the conv. filters change the 3d input into a 2d output

[@sharath_9246]

When you dot product 3d image example(3*32*32) with filter(3*5*5) gives a 2d feature map (28*28) just bcoz of the dot product operation between image and filter

[@rajivb9493]

at 35:09, the expression for output in case of stride convolution is (W - K + 2P)/S +1...for W=7, K=3, P = (K-1)/2 = 1 & S=2 we get output as (7 - 3 + 2*1)/2 + 1 = 3 +1 = 4 ...however, the slide shows the output as 3x3 instead of 4x4 at the right hand corner... is it correct..?

[@bibiworm]

I have the same question.

[@krishnatibrewal5546]

both are different situations, the calculation is done without padding whereas the formula is written considering padding

[@rajivb9493]

@@krishnatibrewal5546 ... thanks a lot, yes you're right..

[@bibiworm]

@@krishnatibrewal5546 thanks.

[@rajivb9493]

In Batch Normalization during Test time at 59:52, what are the averaging equations used to average Mean & Std deviation, sigma ..during the lecture some mention is made of exponential mean of Mean vectors & Sigma vectors...please suggest.

[@hasan0770816268]

33:10 stride
53:00 batch normalization

[@intoeleven]

why they don't use batch norm + layer norm together?

[@bibiworm]

1:01:30 what did he mean by “fusing BN with FC layer or Conv layer”?

[@krishnatibrewal5546]

You can have conv-pool-batchnorm-relu or fc- bn- relu , batch norm can be induced between any layer of the network

[@bibiworm]

@@krishnatibrewal5546 thanks a lot!

[@yahavx]

Because both are linear operators, then you can simply concat them after training (think of them as matrices A and B, in test time you multiply C=A*B and you put that instead of both)

[@ibrexg]

Well don! here is more explanation to normalization: ua-cam.com/video/sxEqtjLC0aM/v-deo.html&ab_channel=NormalizedNerd

[@magic4266]

sounds like someone was building duplo the entire lecture

[@brendawilliams8062]

Thomas the tank engine?

[@park5605]

ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem ahem .
ahem ahem.
ahe ahe he he HUUUJUMMMMMMMMMMMM