An NN is like a stack of pancakes, with computation flowing up when we make predictions.
How does it all work?
THREAD: How is it possible to train a well-performing, advanced Computer Vision model πΌπ» ππ΅π² ππ£π¨? π€
At the heart of this lies the most important technique in modern deep learning - transfer learning.
Let's analyze how it
THREAD: Can you start learning cutting-edge deep learning without specialized hardware? \U0001f916
— Radek Osmulski (@radekosmulski) February 11, 2021
In this thread, we will train an advanced Computer Vision model on a challenging dataset. \U0001f415\U0001f408 Training completes in 25 minutes on my 3yrs old Ryzen 5 CPU.
Let me show you how...
An NN is like a stack of pancakes, with computation flowing up when we make predictions.
How does it all work?
An image is a collection of pixels. Each pixel is just a bunch of numbers describing its color.
Here is what it might look like for a black and white image
Each layer performs computation on the image, transforming it and passing it upwards.
The outputs of a layer are called activations, and the outputs of the last layer have a special meaning... they are the predictions!
We would like for it to have a high value, and for other activations in the last layer to be small...
So far so good! But what about transfer learning?
We know that these layers evolve during training to become feature detectors.
What do we mean by that?
A layer above might have windows that construct shapes from these lines.
We might have a window light up when it sees a square, another when it sees a colorful blob.
Maybe one sliding window will combine lines and detect text... maybe another one will learn to detect faces.
Does all of this sound like a hard task?
But, presumably, once we detect all these lower-level features, we can combine them in a plethora of interesting ways? π€
Or trains, planes, and ships. Or blood cell boundaries. Or aneurysms in x-rays. The possibilities are endless!
We let researchers, large corporations, spend millions of dollars to train very complex models.
And then we get to build on top of their work! π
But so much for the theory. How does it all work in practice?
The @fastdotai framework downloaded the model for us and removed the top of it (the part responsible for predicting 1 of 1000 classes).
During training, we kept nearly the entire model frozen, and only trained the uppermost part, making use of all the lower level features that were being detected.
Ingenius! π
We will hopefully get a chance to explore all of them π
If you enjoyed this thread, let me know please and help me reach others who might also be interested ππ
And the visualizations of what the layers can detect?
Next stop - deciphering how it all works in code and finding ways to further improve our model!
Stay tuned for more π
More from Tech
The entire discussion around Facebookβs disclosures of what happened in 2016 is very frustrating. No exec stopped any investigations, but there were a lot of heated discussions about what to publish and when.
In the spring and summer of 2016, as reported by the Times, activity we traced to GRU was reported to the FBI. This was the standard model of interaction companies used for nation-state attacks against likely US targeted.
In the Spring of 2017, after a deep dive into the Fake News phenomena, the security team wanted to publish an update that covered what we had learned. At this point, we didnβt have any advertising content or the big IRA cluster, but we did know about the GRU model.
This report when through dozens of edits as different equities were represented. I did not have any meetings with Sheryl on the paper, but I canβt speak to whether she was in the loop with my higher-ups.
In the end, the difficult question of attribution was settled by us pointing to the DNI report instead of saying Russia or GRU directly. In my pre-briefs with members of Congress, I made it clear that we believed this action was GRU.
The story doesn\u2019t say you were told not to... it says you did so without approval and they tried to obfuscate what you found. Is that true?
— Sarah Frier (@sarahfrier) November 15, 2018
In the spring and summer of 2016, as reported by the Times, activity we traced to GRU was reported to the FBI. This was the standard model of interaction companies used for nation-state attacks against likely US targeted.
In the Spring of 2017, after a deep dive into the Fake News phenomena, the security team wanted to publish an update that covered what we had learned. At this point, we didnβt have any advertising content or the big IRA cluster, but we did know about the GRU model.
This report when through dozens of edits as different equities were represented. I did not have any meetings with Sheryl on the paper, but I canβt speak to whether she was in the loop with my higher-ups.
In the end, the difficult question of attribution was settled by us pointing to the DNI report instead of saying Russia or GRU directly. In my pre-briefs with members of Congress, I made it clear that we believed this action was GRU.
You May Also Like
BREAKING: @CommonsCMS @DamianCollins just released previously sealed #Six4Three @Facebook documents:
Some random interesting tidbits:
1) Zuck approves shutting down platform API access for Twitter's when Vine is released #competition
2) Facebook engineered ways to access user's call history w/o alerting users:
Team considered access to call history considered 'high PR risk' but 'growth team will charge ahead'. @Facebook created upgrade path to access data w/o subjecting users to Android permissions dialogue.
3) The above also confirms @kashhill and other's suspicion that call history was used to improve PYMK (People You May Know) suggestions and newsfeed rankings.
4) Docs also shed more light into @dseetharaman's story on @Facebook monitoring users' @Onavo VPN activity to determine what competitors to mimic or acquire in 2013.
https://t.co/PwiRIL3v9x
Some random interesting tidbits:
1) Zuck approves shutting down platform API access for Twitter's when Vine is released #competition
2) Facebook engineered ways to access user's call history w/o alerting users:
Team considered access to call history considered 'high PR risk' but 'growth team will charge ahead'. @Facebook created upgrade path to access data w/o subjecting users to Android permissions dialogue.
3) The above also confirms @kashhill and other's suspicion that call history was used to improve PYMK (People You May Know) suggestions and newsfeed rankings.
4) Docs also shed more light into @dseetharaman's story on @Facebook monitoring users' @Onavo VPN activity to determine what competitors to mimic or acquire in 2013.
https://t.co/PwiRIL3v9x