The discussion around creating a calendar with agenda items for each week to volunteer presentations is an interesting approach to managing projects and tasks. It suggests that the group may want to establish a system for organizing and sharing responsibilities, which can help ensure that everyone knows their role and can contribute effectively. Tweak the existing Goods idea by incorporating this calendar concept, and it could become a useful tool for planning and coordinating efforts.

The speaker notes that learning about algorithms is just one aspect of a larger ecosystem, and there are many other aspects to consider, such as deployment, testing, and multiple roles within a data science organization. This indicates that the group may want to explore different areas of expertise and find ways to collaborate effectively across these various fields.

In another conversation, a podcast about a company called Stitch Geeks is mentioned. The company has a business model where they ship boxes with clothes selected by machine learning models to customers, who can keep or return them based on their preferences. This system relies on the feedback of customers and uses it to improve its predictions. The speaker finds this concept interesting and notes that the company uses a full stack approach, with both deep-learning experts and deployment specialists working together.

The discussion then shifts to Amazon SageMaker, which is being explored as a potential tool for data science projects in medical imaging. The speaker mentions the need to copy data into a specific format for use by different models, as well as the challenge of deploying models and managing resources. However, they also note that SageMaker provides a number of features that can help with these tasks, including pre-trained models and easy deployment options.

The speaker notes that one of the benefits of using SageMaker is that it allows developers to focus on other aspects of their work, rather than having to worry about the underlying infrastructure. They also mention the ability to start a Jupyter notebook on a CPU-only instance, which can be cost-effective, as well as the option to use multiple GPUs for more complex tasks.

Finally, the speaker notes that SageMaker provides a number of other features and tools that can help with data science projects, including the ability to save and manage weights, as well as access to pre-trained models. They also mention that the system is designed to be user-friendly, making it easier for developers to get started and start working on their projects.

As the discussion comes to a close, it becomes clear that the group has a range of interests and expertise in data science, machine learning, and related areas. The idea of creating a calendar with agenda items for each week is mentioned again, this time as a way to organize and coordinate efforts across different projects and tasks. This suggests that the group may want to establish a system for planning and managing their work, which can help ensure that everyone knows their role and can contribute effectively.

The speaker notes that creating this calendar will require some initial effort, but it could become a useful tool for managing and coordinating efforts in the future. They also mention the idea of starting small and working together to get everything up and running, which suggests that the group is willing to collaborate and work together to achieve their goals.

In conclusion, the discussion around data science and machine learning covers a range of topics, from creating calendars with agenda items for each week to exploring different business models and tools like SageMaker. The group seems to be interested in finding ways to collaborate effectively across different areas of expertise and to establish systems for planning and managing their work. As they move forward, it will be important to continue this discussion and to find ways to support one another as they navigate the challenges and opportunities of working with data science and machine learning.

The speaker's experience with Amazon SageMaker is also noteworthy, as it provides a practical example of how this tool can be used in real-world applications. By exploring the different features and options available in SageMaker, developers can find ways to streamline their workflows and improve their productivity. The ability to save and manage weights, access pre-trained models, and deploy models quickly and easily are all key benefits of using SageMaker.

The podcast about Stitch Geeks provides an interesting case study of how machine learning can be used in a business context. By shipping boxes with clothes selected by machine learning models, the company is able to provide customers with a personalized shopping experience. This approach also allows the company to gather feedback and improve its predictions over time.

Overall, the discussion around data science and machine learning highlights the importance of collaboration and planning when working on complex projects. By establishing systems for organizing and coordinating efforts, developers can work together more effectively and achieve their goals. The exploration of different business models and tools like SageMaker also provides valuable insights into how these technologies can be used in real-world applications.

The speaker's comments about using a Jupyter notebook to start a project on a CPU-only instance are particularly noteworthy, as they highlight the flexibility and affordability of this approach. By starting small and working together, developers can find ways to overcome challenges and achieve their goals without breaking the bank.

As the discussion comes to a close, it becomes clear that the group has a deep understanding of data science and machine learning concepts. The exploration of different business models and tools like SageMaker provides valuable insights into how these technologies can be used in real-world applications. By establishing systems for organizing and coordinating efforts, developers can work together more effectively and achieve their goals.

The calendar concept mentioned earlier takes on new significance as the group moves forward. By creating a system for planning and managing their work, developers can ensure that everyone knows their role and can contribute effectively. This will be crucial in achieving the group's goals and overcoming challenges along the way.

In conclusion, the discussion around data science and machine learning highlights the importance of collaboration, planning, and flexibility when working on complex projects. By exploring different business models and tools like SageMaker, developers can find ways to streamline their workflows and improve their productivity. The exploration of different concepts and ideas also provides valuable insights into how these technologies can be used in real-world applications.

Overall, the discussion highlights the importance of finding ways to collaborate effectively across different areas of expertise. By establishing systems for organizing and coordinating efforts, developers can work together more effectively and achieve their goals. The exploration of different business models and tools like SageMaker provides valuable insights into how these technologies can be used in real-world applications.

The speaker's experience with Amazon SageMaker is also noteworthy, as it provides a practical example of how this tool can be used in real-world applications. By exploring the different features and options available in SageMaker, developers can find ways to streamline their workflows and improve their productivity.

As the discussion comes to a close, it becomes clear that the group has a deep understanding of data science and machine learning concepts. The exploration of different business models and tools like SageMaker provides valuable insights into how these technologies can be used in real-world applications. By establishing systems for organizing and coordinating efforts, developers can work together more effectively and achieve their goals.

Overall, the discussion highlights the importance of collaboration, planning, and flexibility when working on complex projects. By exploring different business models and tools like SageMaker, developers can find ways to streamline their workflows and improve their productivity.

"WEBVTTKind: captionsLanguage: enwhen scream and today I will talk about this paper called a imagenet train CNN's are biased towards texture and increasing the shape bias improves accuracy and robustness so this was presented at at a major conference it's called I CLR I don't remember the full form of I Sierra and this is just presented last week this is one of my favorite papers from this conference and if you have any questions to stop me at any time more interactive it is the more fun the discussion will be so just my acknowledgments I basically present used his talk and the first author is Robert Garros and I use screenshots from his talk and also from his blog and and from these two courses if you are interested in cnn's at the Stanford 231 n is the best convolutional neural network courses I think now we have intuition CNN's they have different versions on this course and all the material is online the videos are on YouTube and we have it for 2016 17 18 and this is just an older more traditional course where it's doing texture and handles okay so the title of this talk is you know we know CNN's are doing really well so for example in object recognition we have in the image net competition their rate is below 5% so one of the things this paper is trying to understand is what is the CNN learning why is it doing so well is it looking at the shape of objects or is it learning focusing on the texture so just to define image texture here you see three examples so you have some so image texture gives us information of pixels or intensity when an image and you can have say artificial texture like say this image of bricks or you have you found find this in the nature to like say the skin the four of a cat or the floor of a dog or the skin of any animal or basically that's that's the general idea about image texture and if you just do a google search on so before the deep learning error there were lot of image processing methods to find image texture features so there's both they're looking at how frequently say image intensity is up here how pure periodic things are so you know this is this is a very periodic pattern or it could be just some random texture so there's a whole bunch of literature on image temperature features okay so so far you know like if you if you ask the question like how do cnn's recognize objects a the standard in explanation we have got from various sources it's also layer one you have edges clear to you build more complex shapes so in this example author this is from the author stock so he says layer three you have like a door and as you build up there you get you get more complex objects so in the air Phi has the house and we have something similar in in different questions too so this is showing the features in different layers of vgg from cs2 31 and of course that the ottoman has one more example where maybe it's more easier for me to describe this ideas that the earlier layers have some basic shapes and then cure layer see is getting to a wheel and then layers I might to the complex object and this is from a paper in 2015 there's also some so people are trying to think that or is CNN's recognizing objects as how humans recognize objects but nobody knows that for sure so if so you see these three images right do you do you see any any object in these three images you can type or see it in the chat yeah so this is just a texturized version of okay okay great great guesses so this looks like you know like just something but I'll show you the original image from which these are created so these are created from the dog image and when I saw this example I was quite surprised like if I see these three images I can never say there is a dog but the CNN is able to say that oh there is a dog in all these three images which was quite surprising okay so this is so this is the main question authors try to answer okay you have an image of a cat you have the texture which is the local features and you have the shape so what is the CNN trying how did it recognize the object and there are some papers that said that all CNN is focusing on on the texture and so there was like constituting papers so one one set of people like know CNN focusing on shape and the other was like CNN is focusing on texture so this whole object of objective of this authors paper is can we figure out which is it is a shape or texture okay so in any question so far okay so so what the author did is this this is also nice way like I like the scientific contribution to he's like okay I want to answer this question what is the deep learning experimental setup so he took images from basically imagenet so he took the cat shape and the elephant texture and then just combined it into into this new image basically using standard style transfer methods okay it's a method from 2016 so here's the first result from his work is basically when they they take the original image the texture image the cat image or the new modified image you see the CNN is completely confusion it thinks it's an elephant so they did this on thousands of images pretty much like every image in image and now they also do like what it's called as they also did human experiments so when we created these thousands of images they showed it to a standard CNN and they also had like a bunch of students a bunch of people in the lab look at the images and try and say whether it's a cat or an elephant so what dish means the red is the results for the humans so no matter what you did the human observers had a strong shape biased so they can look at it and say oh this this is a cat and on the right hand side you see what happens to like last night 51 of the most popular CNN architectures is that it has a very strong textured bias so it's it's getting it's getting misleaded by by texture and so the question might be okay this is only a test at 50 how about like google net or alex net or all the other architectures so they all one of the things one of the findings reported is that all of them have have this big texture bias so they are more focused on the local texture rather than learning the shape and the authors say maybe maybe it is because it's easier to learn the local texture than to learn the shape which might be more complex okay so that was their first conclusion was that they mostly recognized cnn's mostly recognized textures and broad shapes so now the second the second question is can we can we make cnn's or have a bigger gift shape a bigger importance and how would you design that experiment what would your training set be so the idea is pretty I think it's a clever idea is you develop something called there's a new data set big made it's called stylize imagenet so this is the original image on the left and what they did is they they replace the texture with random stuff right so you are basically getting rid of the texture but in all these images you can see most of them at least I can say that there is a dog present so they've created this stylized image net with random textures and this data set is available online any any question so far yeah by shape by shape we mean the shape of the object yeah so then it's an elephant or car or let's say all cats have some specific shapes yeah kind of a Leslie yeah yeah it's like you're hallucinating so then it's if you retrain less than 15 with these with this new data set you see your shape bias you start getting a little bit closer to you get performance closer to humans but it is also interesting like for some shapes it's it's still naughty before like like an aeroplane it's still not as good as the red was the human absorb performance and orange is when you train it on the stylized image net and the black is original this lot so one of the questions of the chat is exactly I think yeah that that's exactly one way to think about it this is not data augmentation we don't want the model to learn texture we wanted to learn the shape and I think in the talk the author says like yeah maybe it's easier for the model to learn the texture so let's remove the texture right so we don't want the model to focus on it's exactly what you said okay so then so the two things is okay we think this is kind of a clever idea right okay I want this it says we see a way of how to make the CNN learn something you are interested in so maybe here its shape but maybe you have some other thing in your data set that you are or in your problem you're interested so far the cnn's mostly recognized textures not shapes and then if you drain on a suitable data set you can make them develop better have a shape bias instead so we are able to do that but then the question is is there any benefit of this is this really this is useful in any Cell so what they did is they the first go is the traditional CNN and you have the image net top one performance so if you just took the first output from image net what was your accuracy and if you took image net top Phi is basically just I should took the top five what is the chance and say your input image was a dog and what is the chance you you are correct if you get a dog in the top fire disposal so it's pretty hard to it improve image net so they still get like a point six improvement for the top one so this is for object recognition and there is another task called object detection where you are not trying to say it's when you're trying to put you're trying to find where the object is in the image so you're trying to put a box around around the object right so you use some methods called phone star or CNN or faster CNN or retina net those are some of the methods for object detection so you're trying to say oh is this is there a person in the image and you try to put a box around it and you see a significant improvement for object detection from 70 and for object detection the metric is different it's map ma P so they show there is since there was an improvement in both recognition and object detection so one of if you have any questions you can just speak up right I don't see anything new in the chat so now we have the question why do we there's another interesting question why do we have a shape bias so here you know so you start on the left hand side you see this image of a person on a bicycle and as you move through you're just adding more noise and as you so first four images I can still see the bicycle maybe by this this image there's just all noise the last two images so since they're trying to see okay why do we have a shape why do humans have a shape bias and they think they don't have any evidence but they say you know distortion will destroy texture but even when you add noise you might still be able to see the shape so if you were trying to develop a more robust system maybe it's better to have give shape more importance so again they do like another experiment with lesson 850 and so on the y-axis is the classification accuracy and the x-axis is the amount of noise that's added so initially Nimet and 0.9 knowing noise strength is I think it's from zero to one point nine my strength is very noisy and you see the rest net fall starts falling really quickly right so by 0.2 noise you are below 20 percent classification like this so let's say you collected images right you have your training data set that you have connected and I don't know you are doing driving right we're trying to do automated driving and you collect real images from from a live stream or a video and they might not be as clean or as you know as we well created as your training data set so you might start seeing a big drop in your desk perform so if you if we just focus on texture when they try the same things with like a bunch of people and they see that we have pretty good so even with 0.2 percent noise humans much better than the rest at 50 the question is can we can we make the rest let go more towards this red curve so this is the the rest net the yellow is the rest net trained on their stylized image net they are moving towards the human but it's still a little far away there's still a gap so one of so the link I shed it has it has the talk and it has its presentation and it also has like question or an answer session yeah that's so one question of the chat is resonate 50 texture biases yeah is it 50 that's right it's just at fifty trained on imagenet so if you add a little bit of noise or you know your test images change change a little bit if you can see one of the question somebody asked him was like we still have this gap right so his answer was right now we only changed the training set if we only made this stylized image net but we haven't if you maybe we could do some things with the architecture maybe we need a new architecture for improving shape bias and or maybe there are some other knobs like that you can trained because this is the first time you're seeing like a paper like this where you know you're comparing doing all these experiments in the lab and showing the same images with a large set of images to humans and the CNN yeah so you know that's a good question what type of changes to the CNN eye yeah maybe different corners maybe different number of layers I'm not sure we can look at the different architectures like if you see there's one architecture called inception and there are a lot of things in there that I would have never thought about I think that that's still an open topic like what type of architecture would would work here just that'sthat's interesting too like he was they think like it's it's learning like local texture like you know it's learning let me see yeah they think it's it's somehow focusing on in a menace in the CNN saying can't it's focusing on like say the local texture on the fur and yeah maybe if you show them enough cats you can learn that like okay cats have this kind of texture and then I think there's actually like maybe after the talk I'll show you the paper and then as like at least ten different papers trying to figure this out like is it shape or texture the CNN God okay that's benefits of shape bias that was yeah if Distortion yeah that was the last thing we discusses if you add noise humans are still pretty good but we just met shifty trained on the new data set it gets a little better this is one part still I didn't fully understand the paper he says like they didn't give it like images with lot of noise but it was still able to be more robustness be more robust to different types of noise and fresh image here was wasn't too helpful to me so this is something I'll try to figure out right exactly that there might be more information on shape okay so they were that was the three main conclusion from the paper that because before this paper we did not know so we know like CNN's I'm working really well but for object recognition but we didn't have like the clear evidence like how how is it learning you just we were just putting the images in and we get the output so this is the first time we see like result like this that it's mostly textures or not shape and then you can make you can make it learn shapes if you have a data set they have something about noise robustness right and yeah that's a good way to think about is just like you can think like maybe shape is a better discriminator between objects for for this task the people there was some interesting go back yes I'm just trying to share the people to show you some things about so do you see a PDF right so yeah that that's the people it's actually surprising it's like 22 pages for a conference paper which is just pretty nice so this is some more motivation you know like you even though we had the first successful CNN in 2012 a lot of people trying to figure out what what is doing so this guy right a little he thought like that CNN's have shaped like bias like children but in this paper it shows it's not true and these are this is basically the background like why the author is very interested in this topic and so there's one group that's what it was shaped before and this other group here which thinks it's was a texture but they kind of answer this question this was another nice thing I liked in this paper if you see okay so you see you see finger to write the show to make sure so the column each column is like how the CNN does and the top row is the performance of different networks and bottom row is the type of input so even though yes so from original to risk in CNN's are pretty good especially rest at 50 so the one way to read it this is 2012 2013 2014 and less than 50 is 2015 or 16 but if you start giving edges or thing else you start seeing drops in the performance see what else and if good is all available online for this yeah and this was this isn't this graphs a very this figure is very interesting too so like we're seeing it like in in production you might not get good quality data right maybe maybe something is you have your camera has some noise and then your test images have noise or maybe you know the contrast changes right of your image is very possible or there is some other other type of noise so each ABCD is a different type of noise and the red is basically the humans how the humans do with the different type of noise and again the yellow is with this shape eyes right and the black is the original resonate 50 so what you see is like the original ResNet 15 very very susceptible to different types of noise and they're trying to show how can we be more robust and so I just showed you the noise in that slide I just showed no plot but for uniform noise but then this is for contrast and I think it's it's very hard to get like oral presentation at this conference so you can see how much work they did to get get a oral talk they were huge pattex number of trials see if there's anything interesting this is the same plot I was showing in the presentation but now I try to I was trying to see if I can use this idea for my for a classification task but I have not been successful and I can I can show you what I did so far it's actually a kaggle competition so in this competition the goal was to Claire classify images from like from a museum let me show you some images and then basically what I'm doing is I I didn't write the kernels from scratch it's if you have done cattle before you will know there may be you know whenever there's a competition people will share it share some kernels for say like exploratory data analysis and in this case this is a Museum of Art in New York and they have a lot of objects that have been categorized by a subject matter expert and there are many different mabel's it's a very unbalanced data set so the reason I pick this to try it is there are a lot of kernels so kernels are somebody is written like starter kernel like either with frosty eye or just exploratory data analysis cotton so this let me show you one of them I try and show you what I did and I haven't been able to implement the idea but I can show you what I'm trying to do so there are different images of artifacts right so there are paintings jars objects some some objects are like this they have different shapes they're weird shapes so some objects have like weird height here with and so here in in this one they are trying to so we are trying to tag apply tags to these images and then this exploratory data analysis nice I also use it to like learn more Python stuff because I'm more of a MATLAB MATLAB person so I'm trying to learn how to do these good plots with with Python or so they're using Seaborn so you can quickly look at the code and see how to make these plots okay so this is mostly for the data statistics so they are trying to look at what is the mean value in each Channel and so on so for example you see there is some object with some writing and the image is mostly green this weird and this data set I think the training data said almost it has more than hundred thousand images let me show you the classification code it's basically using fastly ice or something we have discussed it in the previous session different bad size the size of the images are 250 I'm mostly running on on the Kaggle curtains the reason I'm doing Kaggle is because they they give me they give you a free GPU so for example I have run about six or seven experiments and I can run up to let me see what was the see there so 32,000 seconds so that is how many hours it's almost 10 hours so 1 hour is 3600 so 10 hours is 3600 so I have you know this it's only one GPU but I just let it run overnight so these are some of the different experiments I did with different faster parameters okay so what I'm trying to do is show you I'm just using resonate 15 similar to what Jeremy said and using transformations for data augmentation it's the same data bunch methods for train and test great and like we discussed before they are normalized by image the image status because all transfer learning so for my training I have I remember now it's one hundred one hundred and three thousand items so these okay these are some of my these are some of the training images save see great variety and here actually I was thinking like texture is important but anyway but we also want to do texture add shape okay and then the learner is pretty much what's what we have done in flash TI it's a CNN learner with using a rest at 50 models and I saw a lot of people they share their kernels so basically I fought for the kernel and then I'm making changes to it so I can play with it the same same learning rate finders we have done before that part is the same one thing maybe do things that might be new to you is one is this - time augmentation so so be so test time augmentation is a really nice feature is the say okay you do you know normally you just do data augmentations to your training data right but here I take the image of let's say let's say one of these kinds might estimate I do some rotation some manipulations and feed the same image to the and then predict the output and if the model is robust enough you know whatever test time augmentation I did I should get the right output I just need the same output so first AI has this test time augmentation feature that already built so one thing I saw a lot of people talk about here is the metric for learning is focal loss and I'm not too familiar with this I don't know if somebody else as has worked with this before I can show you the code for it oh it exists I just have are they just use the Farsi I was I don't know if anyone else has worked with focal loss so they're saying like so one of the things I read is like basically with high class imbalance focal loss is a better choice than so have used cross-entropy loss when I have multi classes I know sorry it's just I just check one more thing that's just when I ran it it's just what-what I scored I'm actually not doing quite where I am pretty am like 150 out of 300 so this business is not working sorry that's that's what so this is this is how the loss function is defined I don't know if someone has worked with this before it's the first time I come across it so I I still don't know how to so the talk was about you know like shape and texture I still don't know how to train this one how to use that idea and this classification problem less I'm still trying to do and what other ideas I can do to improve performance but this is a good composition because it's it's part of its part of constructs called cvpr so there's no prize money it's just like it's like a learning learning competition so actually that's all I had to share for today someone else has sure and stuff for me I'll make it public and sure and I'll share the okay thanks for sharing actually that's thank you okay that's all I have do you have anything anyone else wants to present or talk about any project they have we still have some time for this if not we could talk about so last week we finished the the part lesson of the study group and today anything the agenda part the representation so usually we have some per week so it's just the it was the question I will how do we want to set up our makeups going forward so as burned meetups we have the Americas meet up once per week slide by the Sam then we have EMEA meetups also once per month and that's by I wish Christian and I usually talk about some some some news like couple minutes and then the main presentation is like a paper presentation something similar that we had today and then we have the study groups that was around the first the I part 1 part 2 then again part 1 and again card 2 and so on and now as we finished that part 2 last week is a question how we want to set up the study group going forward so if anyone any suggestions they'll be great to hear how you see this our group developing couple of ideas we could meet weekly bi-weekly once per month we could start in part one we can start again part two or we could do something between or change the different topic or something like that any idea welcome I don't know if you can then interest we'll start part two again you know because I've been listening to part 1 and part one seems so much easy you know part one is so much easier to part two and this so much material in part 2 but I don't know if there's enough interest for part two again yeah so the three the part two is going to be made available to public something around June that's what Jeremy was saying okay so we could we could potentially start that about June time again so people that didn't have a chance to participate in their life apart we can get some new people on board and then maybe start part two about June time possibly that's one option that leaves us with couple of weeks between now and June well before maybe something like that so is the question also what we want to do between now and June yeah what could change anything we could talk about anything we could yeah and the question about additional lesson I didn't see that one was supposed to be about the audio didn't see any information on the on the forums about when the audio lesson would be available let me see the forums but I don't see I would imagine that would be from Jeremy but let me see there is any post from journey posted by Jeremy I guess that will be part two by the latest an LP challenge project discussion how did you first the plans now that we only post a soul about some sort of kind of like plans was about the plans for the library so the planning to implement those improvements they identified during the part two course in our version 1.2 and that's most about the callbacks they're going to implement new callback system the new data block API and the augmentation the stateful optimizer and the NLP pre-processing so those topics going to end up in the in the version 1.2 and it's going to take for them couple of weeks to do it that's so far I didn't of the additional lectures but they wanted to do on audio processing and some other topics so I don't I don't know actually one is going to be so we just waiting for that and yeah so just the question so if anyone's got any idea or what would be interesting to do if anything on weekly basis or anything other than that we can continue that discussion on a slack like if you see my screen like with the you know like the old Conan's like old competitions you you can the nice thing is even the competition is over you can still submit this right but it's that's more like this was like object detection so they wanted to put so the idea would be like just doing more coding and most of the kernels are there so you you know you the hard part is getting started like load your images getting everything organized and then run running stuff so mostly the I think there is the flash the I kernel or I used a lot of Kara's before so there is there is a tennis kernel so you can you can take the kernel basically just work it and make make your changes and run it and everybody basically and now the improve the GPU I think it used to be a test nvidia t-80 and now it's like better ones but this is more coding like so you could have like half an hour of learning the method review of the method and then just hands-on coding basically like we could for example pick a kernel and then everybody does it like either does can do it beforehand or do it during a session yeah but this this guy from Switzerland is really good his name is Kevin Mahdi he's a professor in Switzerland and he writes the likes but he writes in kara so this is like yeah and then oh and we can even if you take like a completed competition it's his kernel will be like you you will not get in the top 10% but if you just get started you know most of the times for me it's been just organizing this data getting it getting it to train and tests plate question on the chart about doing a session about how to start in kaggle so like if someone is new to cargo platform we could buy session on how to do it it's for sure and there is another suggestion that we could have a presentation on deploying models in two weeks time so okay we can we can get a agenda things like that yep so the moment seems like we have the 25th of May so maybe we can do little things like that so if anyone wanted to present anything we could use that time slot we have on Saturdays 9:00 a.m. Pacific which seemed to work for everyone and if we have any topic on the agenda then we do the meeting if we if there is anything to talk then we then we just don't do it so so let me try to create some sort of like calendar with with meetings with plants so far which at the moment will be the deploying deep learning models in two weeks and if anyone want to do the how to start encargo presentation they'd also be great I can I can do that you can do that okay sounds like we have two already yes that's great that's fantastic okay it seems like we will just continue with those weekly meetups as long as we have ideas for the presentation that's great that's right and then and then then I guess that's only a suggestion if we want to do the part two part two is available to the public and then get some more people on board with part two that we can do that as well so I know Daniel Daniel did like this quick overview of full stack deep learning but I thought it would be interesting to maybe I started watching those videos I said we'll be interesting to discuss maybe them in details and like have questions and answers on those videos but I don't know if like how many people would want to do that yeah that's something this this course is also my wish to do it seems like interesting course or yeah so we could do that as well but they have like 12 videos or something like that's right that's going to take some time and you need you need to like I think you need some sort of a deployment platform so either like I think they use Google DCP by default right yeah we could we could do that as well so that's that's a good idea so for that we would need to plan like three months or something if it's one if it's one video per week that's going to take a while that's okay they're much easier they're like I think those easier those videos are much easier then first day I at least I think it's all about actually like you know working with code and implementing it okay yeah so yeah that's right so someone was asking what what sort of the videos this is the full stack deep learning where they talk about they have actually nice they have videos on the presentation available and it happened much this year so it's it's quite recent and they have a github also okay with with problems and another one was Brian another directory with solutions oh that's great so it's fantastic this one slide when they show all the all the deep learning kind of ecosystem Wilco like that and and the thing that Jeremy teacher does is just like one one bit of the whole the big picture and there's actually a lot more to that it's how we deal with the data how we were straight how loud the data will keep the data how we get the data and then we I guess do some exploratory data analysis we do they talk about this oh I know I don't remember it but I got into a third or fourth lecture okay and then of course you create a model and train a model but then the next step of course if you test your mobile and then you deploy your model then you maintain your model I guess they also talk about testing so the the Pope of T of that course is quite quite nice so we could do that you could look into that in a week basis and it's going to take a while but that's fine so what we could do we could tweak Goods I could create some sort of calendar with the agenda items for each week and then we'll just ask people to volunteer to present anything and then we all will go from there other gas yeah instead deep run will be good yeah we just learned we just learned algorithms part but there's a whole like you said there's a whole ecosystem if there is more than that yes and that depends it depends what you want to do someone like Jeremy interesting kind of research of declaring novels making them fast and and so on but I guess more reward problems there's also the deployment part and so on so on and and I would imagine kind of like bigger organization there will be there will be multiple roles like in a data scientist roles when some people work on algorithms some people would work on deployments some people would work on testing and so on so on but at least I think it's good to know what are those tack to kind of get the deeper understanding of the full spectrum and I was I was listening to a podcast with a guide company called stitch geeks what they do they have some sort of business model where they ship a a box with some clothes to customers and the clothes are selected by machine learning model so they have customers they they the machine learning model fix some like a fashion clothes or whatever do some some clothes they send them to customers and then if they like them they keep them and they pay for those of course if they don't like those clothes they send them back so they have like a feedback that costs the money if then even if the model is wrong that caused the because people didn't like what was provided to them so they send them back and they have to they have to keep them in warehouse whatever so that's what I do and they what they suggest they do well they have like a full stack deep-learning guys so they don't have those although it's quite a big company but they don't have those specific roles like someone only that's models someone who does deployment they have like a full stack guys and I believe that this kind of boot for and it's also works was good for them because people don't need to wait so someone who is deploying the models don't need to wait for someone who's creating the model so the one person can do everything so I think that was quite that absent message to this guy asking before like what certain maybe training they have for those people how would a dragon to become like do all of that but I didn't get any response but I think this will stuck the planning's it's about something like that experience like we've been most like my boss has been asking me to like look into Amazon sage maker most of most of what we are doing in you know medical imaging and we're mostly using Amazon and right now so for example I start an instance I will have to you know like copy the data there have died I have like a particular format called I comms I have to convert it to PNG so that these fastly I or different models can use it but Shh make that is supposed to like something to help keep running developers also for for training and then deployment but maybe this may be full stack over it yes H my guess got us is got quite a lot of I looked at that and say it's they have they have a lot of ready for you so like you don't necessarily need to know how to create different models if you can use existing one we've been we've been training too so we can start a Jupiter notebook and you know on our instance and then they because you we have the weight right like the weights so the question and where to save the weights so I can deploy it they'll make it easy to save weights and also I think the sentiment seismic and they also have nice idea that they have you can so you can start your notebook on CPU only instance which she doesn't cost a lot of money and then free from that cpu instance you can start using some sage America oh you can start like a parallel not parallel instances only for threading and with the GPU you can start multiple GPUs if you have train a lot of images but your mind and then once you once you finish training that GPU virtual machine is going to stop so you don't have to pay for that young faithfully for that if you need it so so I think that's also quite interesting what they have there okay excellent is there anything else from anyone so that's great so let's let me create that calendar with the some dates and then I'll I'll put that on a slack and then people can put their names and we can start with the full stack and the and with the cargo so to star in cargo session and then we have the 25th of my deployment session it's nice excellent okay thanks everyone for joining today thanks Michael for presentationwhen scream and today I will talk about this paper called a imagenet train CNN's are biased towards texture and increasing the shape bias improves accuracy and robustness so this was presented at at a major conference it's called I CLR I don't remember the full form of I Sierra and this is just presented last week this is one of my favorite papers from this conference and if you have any questions to stop me at any time more interactive it is the more fun the discussion will be so just my acknowledgments I basically present used his talk and the first author is Robert Garros and I use screenshots from his talk and also from his blog and and from these two courses if you are interested in cnn's at the Stanford 231 n is the best convolutional neural network courses I think now we have intuition CNN's they have different versions on this course and all the material is online the videos are on YouTube and we have it for 2016 17 18 and this is just an older more traditional course where it's doing texture and handles okay so the title of this talk is you know we know CNN's are doing really well so for example in object recognition we have in the image net competition their rate is below 5% so one of the things this paper is trying to understand is what is the CNN learning why is it doing so well is it looking at the shape of objects or is it learning focusing on the texture so just to define image texture here you see three examples so you have some so image texture gives us information of pixels or intensity when an image and you can have say artificial texture like say this image of bricks or you have you found find this in the nature to like say the skin the four of a cat or the floor of a dog or the skin of any animal or basically that's that's the general idea about image texture and if you just do a google search on so before the deep learning error there were lot of image processing methods to find image texture features so there's both they're looking at how frequently say image intensity is up here how pure periodic things are so you know this is this is a very periodic pattern or it could be just some random texture so there's a whole bunch of literature on image temperature features okay so so far you know like if you if you ask the question like how do cnn's recognize objects a the standard in explanation we have got from various sources it's also layer one you have edges clear to you build more complex shapes so in this example author this is from the author stock so he says layer three you have like a door and as you build up there you get you get more complex objects so in the air Phi has the house and we have something similar in in different questions too so this is showing the features in different layers of vgg from cs2 31 and of course that the ottoman has one more example where maybe it's more easier for me to describe this ideas that the earlier layers have some basic shapes and then cure layer see is getting to a wheel and then layers I might to the complex object and this is from a paper in 2015 there's also some so people are trying to think that or is CNN's recognizing objects as how humans recognize objects but nobody knows that for sure so if so you see these three images right do you do you see any any object in these three images you can type or see it in the chat yeah so this is just a texturized version of okay okay great great guesses so this looks like you know like just something but I'll show you the original image from which these are created so these are created from the dog image and when I saw this example I was quite surprised like if I see these three images I can never say there is a dog but the CNN is able to say that oh there is a dog in all these three images which was quite surprising okay so this is so this is the main question authors try to answer okay you have an image of a cat you have the texture which is the local features and you have the shape so what is the CNN trying how did it recognize the object and there are some papers that said that all CNN is focusing on on the texture and so there was like constituting papers so one one set of people like know CNN focusing on shape and the other was like CNN is focusing on texture so this whole object of objective of this authors paper is can we figure out which is it is a shape or texture okay so in any question so far okay so so what the author did is this this is also nice way like I like the scientific contribution to he's like okay I want to answer this question what is the deep learning experimental setup so he took images from basically imagenet so he took the cat shape and the elephant texture and then just combined it into into this new image basically using standard style transfer methods okay it's a method from 2016 so here's the first result from his work is basically when they they take the original image the texture image the cat image or the new modified image you see the CNN is completely confusion it thinks it's an elephant so they did this on thousands of images pretty much like every image in image and now they also do like what it's called as they also did human experiments so when we created these thousands of images they showed it to a standard CNN and they also had like a bunch of students a bunch of people in the lab look at the images and try and say whether it's a cat or an elephant so what dish means the red is the results for the humans so no matter what you did the human observers had a strong shape biased so they can look at it and say oh this this is a cat and on the right hand side you see what happens to like last night 51 of the most popular CNN architectures is that it has a very strong textured bias so it's it's getting it's getting misleaded by by texture and so the question might be okay this is only a test at 50 how about like google net or alex net or all the other architectures so they all one of the things one of the findings reported is that all of them have have this big texture bias so they are more focused on the local texture rather than learning the shape and the authors say maybe maybe it is because it's easier to learn the local texture than to learn the shape which might be more complex okay so that was their first conclusion was that they mostly recognized cnn's mostly recognized textures and broad shapes so now the second the second question is can we can we make cnn's or have a bigger gift shape a bigger importance and how would you design that experiment what would your training set be so the idea is pretty I think it's a clever idea is you develop something called there's a new data set big made it's called stylize imagenet so this is the original image on the left and what they did is they they replace the texture with random stuff right so you are basically getting rid of the texture but in all these images you can see most of them at least I can say that there is a dog present so they've created this stylized image net with random textures and this data set is available online any any question so far yeah by shape by shape we mean the shape of the object yeah so then it's an elephant or car or let's say all cats have some specific shapes yeah kind of a Leslie yeah yeah it's like you're hallucinating so then it's if you retrain less than 15 with these with this new data set you see your shape bias you start getting a little bit closer to you get performance closer to humans but it is also interesting like for some shapes it's it's still naughty before like like an aeroplane it's still not as good as the red was the human absorb performance and orange is when you train it on the stylized image net and the black is original this lot so one of the questions of the chat is exactly I think yeah that that's exactly one way to think about it this is not data augmentation we don't want the model to learn texture we wanted to learn the shape and I think in the talk the author says like yeah maybe it's easier for the model to learn the texture so let's remove the texture right so we don't want the model to focus on it's exactly what you said okay so then so the two things is okay we think this is kind of a clever idea right okay I want this it says we see a way of how to make the CNN learn something you are interested in so maybe here its shape but maybe you have some other thing in your data set that you are or in your problem you're interested so far the cnn's mostly recognized textures not shapes and then if you drain on a suitable data set you can make them develop better have a shape bias instead so we are able to do that but then the question is is there any benefit of this is this really this is useful in any Cell so what they did is they the first go is the traditional CNN and you have the image net top one performance so if you just took the first output from image net what was your accuracy and if you took image net top Phi is basically just I should took the top five what is the chance and say your input image was a dog and what is the chance you you are correct if you get a dog in the top fire disposal so it's pretty hard to it improve image net so they still get like a point six improvement for the top one so this is for object recognition and there is another task called object detection where you are not trying to say it's when you're trying to put you're trying to find where the object is in the image so you're trying to put a box around around the object right so you use some methods called phone star or CNN or faster CNN or retina net those are some of the methods for object detection so you're trying to say oh is this is there a person in the image and you try to put a box around it and you see a significant improvement for object detection from 70 and for object detection the metric is different it's map ma P so they show there is since there was an improvement in both recognition and object detection so one of if you have any questions you can just speak up right I don't see anything new in the chat so now we have the question why do we there's another interesting question why do we have a shape bias so here you know so you start on the left hand side you see this image of a person on a bicycle and as you move through you're just adding more noise and as you so first four images I can still see the bicycle maybe by this this image there's just all noise the last two images so since they're trying to see okay why do we have a shape why do humans have a shape bias and they think they don't have any evidence but they say you know distortion will destroy texture but even when you add noise you might still be able to see the shape so if you were trying to develop a more robust system maybe it's better to have give shape more importance so again they do like another experiment with lesson 850 and so on the y-axis is the classification accuracy and the x-axis is the amount of noise that's added so initially Nimet and 0.9 knowing noise strength is I think it's from zero to one point nine my strength is very noisy and you see the rest net fall starts falling really quickly right so by 0.2 noise you are below 20 percent classification like this so let's say you collected images right you have your training data set that you have connected and I don't know you are doing driving right we're trying to do automated driving and you collect real images from from a live stream or a video and they might not be as clean or as you know as we well created as your training data set so you might start seeing a big drop in your desk perform so if you if we just focus on texture when they try the same things with like a bunch of people and they see that we have pretty good so even with 0.2 percent noise humans much better than the rest at 50 the question is can we can we make the rest let go more towards this red curve so this is the the rest net the yellow is the rest net trained on their stylized image net they are moving towards the human but it's still a little far away there's still a gap so one of so the link I shed it has it has the talk and it has its presentation and it also has like question or an answer session yeah that's so one question of the chat is resonate 50 texture biases yeah is it 50 that's right it's just at fifty trained on imagenet so if you add a little bit of noise or you know your test images change change a little bit if you can see one of the question somebody asked him was like we still have this gap right so his answer was right now we only changed the training set if we only made this stylized image net but we haven't if you maybe we could do some things with the architecture maybe we need a new architecture for improving shape bias and or maybe there are some other knobs like that you can trained because this is the first time you're seeing like a paper like this where you know you're comparing doing all these experiments in the lab and showing the same images with a large set of images to humans and the CNN yeah so you know that's a good question what type of changes to the CNN eye yeah maybe different corners maybe different number of layers I'm not sure we can look at the different architectures like if you see there's one architecture called inception and there are a lot of things in there that I would have never thought about I think that that's still an open topic like what type of architecture would would work here just that'sthat's interesting too like he was they think like it's it's learning like local texture like you know it's learning let me see yeah they think it's it's somehow focusing on in a menace in the CNN saying can't it's focusing on like say the local texture on the fur and yeah maybe if you show them enough cats you can learn that like okay cats have this kind of texture and then I think there's actually like maybe after the talk I'll show you the paper and then as like at least ten different papers trying to figure this out like is it shape or texture the CNN God okay that's benefits of shape bias that was yeah if Distortion yeah that was the last thing we discusses if you add noise humans are still pretty good but we just met shifty trained on the new data set it gets a little better this is one part still I didn't fully understand the paper he says like they didn't give it like images with lot of noise but it was still able to be more robustness be more robust to different types of noise and fresh image here was wasn't too helpful to me so this is something I'll try to figure out right exactly that there might be more information on shape okay so they were that was the three main conclusion from the paper that because before this paper we did not know so we know like CNN's I'm working really well but for object recognition but we didn't have like the clear evidence like how how is it learning you just we were just putting the images in and we get the output so this is the first time we see like result like this that it's mostly textures or not shape and then you can make you can make it learn shapes if you have a data set they have something about noise robustness right and yeah that's a good way to think about is just like you can think like maybe shape is a better discriminator between objects for for this task the people there was some interesting go back yes I'm just trying to share the people to show you some things about so do you see a PDF right so yeah that that's the people it's actually surprising it's like 22 pages for a conference paper which is just pretty nice so this is some more motivation you know like you even though we had the first successful CNN in 2012 a lot of people trying to figure out what what is doing so this guy right a little he thought like that CNN's have shaped like bias like children but in this paper it shows it's not true and these are this is basically the background like why the author is very interested in this topic and so there's one group that's what it was shaped before and this other group here which thinks it's was a texture but they kind of answer this question this was another nice thing I liked in this paper if you see okay so you see you see finger to write the show to make sure so the column each column is like how the CNN does and the top row is the performance of different networks and bottom row is the type of input so even though yes so from original to risk in CNN's are pretty good especially rest at 50 so the one way to read it this is 2012 2013 2014 and less than 50 is 2015 or 16 but if you start giving edges or thing else you start seeing drops in the performance see what else and if good is all available online for this yeah and this was this isn't this graphs a very this figure is very interesting too so like we're seeing it like in in production you might not get good quality data right maybe maybe something is you have your camera has some noise and then your test images have noise or maybe you know the contrast changes right of your image is very possible or there is some other other type of noise so each ABCD is a different type of noise and the red is basically the humans how the humans do with the different type of noise and again the yellow is with this shape eyes right and the black is the original resonate 50 so what you see is like the original ResNet 15 very very susceptible to different types of noise and they're trying to show how can we be more robust and so I just showed you the noise in that slide I just showed no plot but for uniform noise but then this is for contrast and I think it's it's very hard to get like oral presentation at this conference so you can see how much work they did to get get a oral talk they were huge pattex number of trials see if there's anything interesting this is the same plot I was showing in the presentation but now I try to I was trying to see if I can use this idea for my for a classification task but I have not been successful and I can I can show you what I did so far it's actually a kaggle competition so in this competition the goal was to Claire classify images from like from a museum let me show you some images and then basically what I'm doing is I I didn't write the kernels from scratch it's if you have done cattle before you will know there may be you know whenever there's a competition people will share it share some kernels for say like exploratory data analysis and in this case this is a Museum of Art in New York and they have a lot of objects that have been categorized by a subject matter expert and there are many different mabel's it's a very unbalanced data set so the reason I pick this to try it is there are a lot of kernels so kernels are somebody is written like starter kernel like either with frosty eye or just exploratory data analysis cotton so this let me show you one of them I try and show you what I did and I haven't been able to implement the idea but I can show you what I'm trying to do so there are different images of artifacts right so there are paintings jars objects some some objects are like this they have different shapes they're weird shapes so some objects have like weird height here with and so here in in this one they are trying to so we are trying to tag apply tags to these images and then this exploratory data analysis nice I also use it to like learn more Python stuff because I'm more of a MATLAB MATLAB person so I'm trying to learn how to do these good plots with with Python or so they're using Seaborn so you can quickly look at the code and see how to make these plots okay so this is mostly for the data statistics so they are trying to look at what is the mean value in each Channel and so on so for example you see there is some object with some writing and the image is mostly green this weird and this data set I think the training data said almost it has more than hundred thousand images let me show you the classification code it's basically using fastly ice or something we have discussed it in the previous session different bad size the size of the images are 250 I'm mostly running on on the Kaggle curtains the reason I'm doing Kaggle is because they they give me they give you a free GPU so for example I have run about six or seven experiments and I can run up to let me see what was the see there so 32,000 seconds so that is how many hours it's almost 10 hours so 1 hour is 3600 so 10 hours is 3600 so I have you know this it's only one GPU but I just let it run overnight so these are some of the different experiments I did with different faster parameters okay so what I'm trying to do is show you I'm just using resonate 15 similar to what Jeremy said and using transformations for data augmentation it's the same data bunch methods for train and test great and like we discussed before they are normalized by image the image status because all transfer learning so for my training I have I remember now it's one hundred one hundred and three thousand items so these okay these are some of my these are some of the training images save see great variety and here actually I was thinking like texture is important but anyway but we also want to do texture add shape okay and then the learner is pretty much what's what we have done in flash TI it's a CNN learner with using a rest at 50 models and I saw a lot of people they share their kernels so basically I fought for the kernel and then I'm making changes to it so I can play with it the same same learning rate finders we have done before that part is the same one thing maybe do things that might be new to you is one is this - time augmentation so so be so test time augmentation is a really nice feature is the say okay you do you know normally you just do data augmentations to your training data right but here I take the image of let's say let's say one of these kinds might estimate I do some rotation some manipulations and feed the same image to the and then predict the output and if the model is robust enough you know whatever test time augmentation I did I should get the right output I just need the same output so first AI has this test time augmentation feature that already built so one thing I saw a lot of people talk about here is the metric for learning is focal loss and I'm not too familiar with this I don't know if somebody else as has worked with this before I can show you the code for it oh it exists I just have are they just use the Farsi I was I don't know if anyone else has worked with focal loss so they're saying like so one of the things I read is like basically with high class imbalance focal loss is a better choice than so have used cross-entropy loss when I have multi classes I know sorry it's just I just check one more thing that's just when I ran it it's just what-what I scored I'm actually not doing quite where I am pretty am like 150 out of 300 so this business is not working sorry that's that's what so this is this is how the loss function is defined I don't know if someone has worked with this before it's the first time I come across it so I I still don't know how to so the talk was about you know like shape and texture I still don't know how to train this one how to use that idea and this classification problem less I'm still trying to do and what other ideas I can do to improve performance but this is a good composition because it's it's part of its part of constructs called cvpr so there's no prize money it's just like it's like a learning learning competition so actually that's all I had to share for today someone else has sure and stuff for me I'll make it public and sure and I'll share the okay thanks for sharing actually that's thank you okay that's all I have do you have anything anyone else wants to present or talk about any project they have we still have some time for this if not we could talk about so last week we finished the the part lesson of the study group and today anything the agenda part the representation so usually we have some per week so it's just the it was the question I will how do we want to set up our makeups going forward so as burned meetups we have the Americas meet up once per week slide by the Sam then we have EMEA meetups also once per month and that's by I wish Christian and I usually talk about some some some news like couple minutes and then the main presentation is like a paper presentation something similar that we had today and then we have the study groups that was around the first the I part 1 part 2 then again part 1 and again card 2 and so on and now as we finished that part 2 last week is a question how we want to set up the study group going forward so if anyone any suggestions they'll be great to hear how you see this our group developing couple of ideas we could meet weekly bi-weekly once per month we could start in part one we can start again part two or we could do something between or change the different topic or something like that any idea welcome I don't know if you can then interest we'll start part two again you know because I've been listening to part 1 and part one seems so much easy you know part one is so much easier to part two and this so much material in part 2 but I don't know if there's enough interest for part two again yeah so the three the part two is going to be made available to public something around June that's what Jeremy was saying okay so we could we could potentially start that about June time again so people that didn't have a chance to participate in their life apart we can get some new people on board and then maybe start part two about June time possibly that's one option that leaves us with couple of weeks between now and June well before maybe something like that so is the question also what we want to do between now and June yeah what could change anything we could talk about anything we could yeah and the question about additional lesson I didn't see that one was supposed to be about the audio didn't see any information on the on the forums about when the audio lesson would be available let me see the forums but I don't see I would imagine that would be from Jeremy but let me see there is any post from journey posted by Jeremy I guess that will be part two by the latest an LP challenge project discussion how did you first the plans now that we only post a soul about some sort of kind of like plans was about the plans for the library so the planning to implement those improvements they identified during the part two course in our version 1.2 and that's most about the callbacks they're going to implement new callback system the new data block API and the augmentation the stateful optimizer and the NLP pre-processing so those topics going to end up in the in the version 1.2 and it's going to take for them couple of weeks to do it that's so far I didn't of the additional lectures but they wanted to do on audio processing and some other topics so I don't I don't know actually one is going to be so we just waiting for that and yeah so just the question so if anyone's got any idea or what would be interesting to do if anything on weekly basis or anything other than that we can continue that discussion on a slack like if you see my screen like with the you know like the old Conan's like old competitions you you can the nice thing is even the competition is over you can still submit this right but it's that's more like this was like object detection so they wanted to put so the idea would be like just doing more coding and most of the kernels are there so you you know you the hard part is getting started like load your images getting everything organized and then run running stuff so mostly the I think there is the flash the I kernel or I used a lot of Kara's before so there is there is a tennis kernel so you can you can take the kernel basically just work it and make make your changes and run it and everybody basically and now the improve the GPU I think it used to be a test nvidia t-80 and now it's like better ones but this is more coding like so you could have like half an hour of learning the method review of the method and then just hands-on coding basically like we could for example pick a kernel and then everybody does it like either does can do it beforehand or do it during a session yeah but this this guy from Switzerland is really good his name is Kevin Mahdi he's a professor in Switzerland and he writes the likes but he writes in kara so this is like yeah and then oh and we can even if you take like a completed competition it's his kernel will be like you you will not get in the top 10% but if you just get started you know most of the times for me it's been just organizing this data getting it getting it to train and tests plate question on the chart about doing a session about how to start in kaggle so like if someone is new to cargo platform we could buy session on how to do it it's for sure and there is another suggestion that we could have a presentation on deploying models in two weeks time so okay we can we can get a agenda things like that yep so the moment seems like we have the 25th of May so maybe we can do little things like that so if anyone wanted to present anything we could use that time slot we have on Saturdays 9:00 a.m. Pacific which seemed to work for everyone and if we have any topic on the agenda then we do the meeting if we if there is anything to talk then we then we just don't do it so so let me try to create some sort of like calendar with with meetings with plants so far which at the moment will be the deploying deep learning models in two weeks and if anyone want to do the how to start encargo presentation they'd also be great I can I can do that you can do that okay sounds like we have two already yes that's great that's fantastic okay it seems like we will just continue with those weekly meetups as long as we have ideas for the presentation that's great that's right and then and then then I guess that's only a suggestion if we want to do the part two part two is available to the public and then get some more people on board with part two that we can do that as well so I know Daniel Daniel did like this quick overview of full stack deep learning but I thought it would be interesting to maybe I started watching those videos I said we'll be interesting to discuss maybe them in details and like have questions and answers on those videos but I don't know if like how many people would want to do that yeah that's something this this course is also my wish to do it seems like interesting course or yeah so we could do that as well but they have like 12 videos or something like that's right that's going to take some time and you need you need to like I think you need some sort of a deployment platform so either like I think they use Google DCP by default right yeah we could we could do that as well so that's that's a good idea so for that we would need to plan like three months or something if it's one if it's one video per week that's going to take a while that's okay they're much easier they're like I think those easier those videos are much easier then first day I at least I think it's all about actually like you know working with code and implementing it okay yeah so yeah that's right so someone was asking what what sort of the videos this is the full stack deep learning where they talk about they have actually nice they have videos on the presentation available and it happened much this year so it's it's quite recent and they have a github also okay with with problems and another one was Brian another directory with solutions oh that's great so it's fantastic this one slide when they show all the all the deep learning kind of ecosystem Wilco like that and and the thing that Jeremy teacher does is just like one one bit of the whole the big picture and there's actually a lot more to that it's how we deal with the data how we were straight how loud the data will keep the data how we get the data and then we I guess do some exploratory data analysis we do they talk about this oh I know I don't remember it but I got into a third or fourth lecture okay and then of course you create a model and train a model but then the next step of course if you test your mobile and then you deploy your model then you maintain your model I guess they also talk about testing so the the Pope of T of that course is quite quite nice so we could do that you could look into that in a week basis and it's going to take a while but that's fine so what we could do we could tweak Goods I could create some sort of calendar with the agenda items for each week and then we'll just ask people to volunteer to present anything and then we all will go from there other gas yeah instead deep run will be good yeah we just learned we just learned algorithms part but there's a whole like you said there's a whole ecosystem if there is more than that yes and that depends it depends what you want to do someone like Jeremy interesting kind of research of declaring novels making them fast and and so on but I guess more reward problems there's also the deployment part and so on so on and and I would imagine kind of like bigger organization there will be there will be multiple roles like in a data scientist roles when some people work on algorithms some people would work on deployments some people would work on testing and so on so on but at least I think it's good to know what are those tack to kind of get the deeper understanding of the full spectrum and I was I was listening to a podcast with a guide company called stitch geeks what they do they have some sort of business model where they ship a a box with some clothes to customers and the clothes are selected by machine learning model so they have customers they they the machine learning model fix some like a fashion clothes or whatever do some some clothes they send them to customers and then if they like them they keep them and they pay for those of course if they don't like those clothes they send them back so they have like a feedback that costs the money if then even if the model is wrong that caused the because people didn't like what was provided to them so they send them back and they have to they have to keep them in warehouse whatever so that's what I do and they what they suggest they do well they have like a full stack deep-learning guys so they don't have those although it's quite a big company but they don't have those specific roles like someone only that's models someone who does deployment they have like a full stack guys and I believe that this kind of boot for and it's also works was good for them because people don't need to wait so someone who is deploying the models don't need to wait for someone who's creating the model so the one person can do everything so I think that was quite that absent message to this guy asking before like what certain maybe training they have for those people how would a dragon to become like do all of that but I didn't get any response but I think this will stuck the planning's it's about something like that experience like we've been most like my boss has been asking me to like look into Amazon sage maker most of most of what we are doing in you know medical imaging and we're mostly using Amazon and right now so for example I start an instance I will have to you know like copy the data there have died I have like a particular format called I comms I have to convert it to PNG so that these fastly I or different models can use it but Shh make that is supposed to like something to help keep running developers also for for training and then deployment but maybe this may be full stack over it yes H my guess got us is got quite a lot of I looked at that and say it's they have they have a lot of ready for you so like you don't necessarily need to know how to create different models if you can use existing one we've been we've been training too so we can start a Jupiter notebook and you know on our instance and then they because you we have the weight right like the weights so the question and where to save the weights so I can deploy it they'll make it easy to save weights and also I think the sentiment seismic and they also have nice idea that they have you can so you can start your notebook on CPU only instance which she doesn't cost a lot of money and then free from that cpu instance you can start using some sage America oh you can start like a parallel not parallel instances only for threading and with the GPU you can start multiple GPUs if you have train a lot of images but your mind and then once you once you finish training that GPU virtual machine is going to stop so you don't have to pay for that young faithfully for that if you need it so so I think that's also quite interesting what they have there okay excellent is there anything else from anyone so that's great so let's let me create that calendar with the some dates and then I'll I'll put that on a slack and then people can put their names and we can start with the full stack and the and with the cargo so to star in cargo session and then we have the 25th of my deployment session it's nice excellent okay thanks everyone for joining today thanks Michael for presentation\n"