AWS ML Heroes in 15: Harnessing Low-code/No-Code Machine Learning on AWS

Key Takeaways

foreign [Music] my name is rustam I'm advised machine learning hero and staff machine learning engineer at instrumental and today I want to tell you how to harness low code and no code machine learning tools on AWS so we entered a new era of AI AI now is a commodity so you don't need to build it from scratch to use it's very low barrier to use or integrate AI or machine learning into your product and you don't need to pay any upfront costs to use it that affects multiple areas at once for AI augmented Workforce it means that you can start using AI for solving day-to-day tasks or using AI as your personal assistant in the area of AI as a core product it means that you don't if you want to build a product around AI you don't need to have dedicated machine learning team and you can deploy something very early uh to get feedback from customers earlier and finally with AI as a support tool it's very easy to embed AI features into your existing product and reduce costs of manual processes so today I want to go through machine learning pipelining challenges and show you different steps then I want to share the framework for comparing different low code no code machine learning tools after that I will go into a final details about some low code no code address machine learning services and finally I will show you the demo of sagemaker canvas custom models so first let's start with machine learning pipeline here you can see the whole pipeline when if you want to deploy machine learning into production so you start with data acquisition labeling then you build the model itself after that you build that training pipeline where you train the model and prepare it for inference after that you deploy the model and organize infrastructure for it and finally you integrate your model into your production infrastructure so you can see here is that there are multiple steps here yeah complexity here is not just code and actually I I think uh low effort no effort machine learning better captures uh what's happening here that you don't it's not only optimizes the code you need to write it also optimizes the data you need to label the infrastructure you need to deploy and the value here is that people with domain knowledge but without ml expertise can start getting value from ml they can deploy a machine learning models into production earlier to either get customer feedback or get value to the users so if we take the pipeline which I shared with you before and put all steps online so from Gathering data set to integrating the model we can put a Spectra here so here you can see this stack and if particular step is marked as orange it means it's handled by AWS and if it's marked as purple it means it will be handled by you so on the left we have integrated ml meaning that everything is handled by AWS with integrated automl means that you need to bring your data and you need to label it but you don't need to worry about anything else about machine learning uh as a service family here you don't need to bring your data but you still need to integrate the final endpoint into your product and with managed pipeline you effectively have to manage everything except for labeling the thing here is that the integrated family a factory can be used by business analyst so to the level where you don't need to write even a single line of code and you can start using machine learning on your data the machine learning is a service or Auto Machine learning as a service can be integrated by back-end engineer you don't need a team of data scientists or machine learning Engineers to do that so in these cases effectively you can start getting value from machine learning very early with minimal resources so if we put all these types of services on one axis uh basically on the left side we we will have something like motion like the service which is very easy to start less customizable but you don't have any training cost and your paper request which is perfect if you have low volume or Peak load on the right side you have managed Pipeline with higher threshold to start at the same time it will be customizable for your task and it may have lower cost at scale because you can adapt your model for your infrastructure for a type of task so now that we have this framework we can put different AWS low code no code tools uh and categorize them So based on uh so basically you can see this framework on the top and you can see the types of tasks you need to solve like visual tasks where we need to work with images or videos text tasks where we need to let's say classify text Snippets or tabular tasks where we need to work with tables of data and here you can see that for example uh you may have integrated ml where you have Canvas Plus recognition or text track you may have machine learning as a service just recognition or you may have just pre-trained model like JumpStart don't worry that there are a lot of sources here I will go into details about each service later but you can see that now it's clear like for each type of service like what part of the stack it will handle and what part will be handled by you and on the top left corner you can see uh Canvas Plus custom models plus different machine learning as a service Family Tools so these were released on March 30th meaning that it's very fresh and you can start using machine learning on your image and Text data right away without any code necessary and I will show you how to do it uh during the demo so now let's look into each service which I mentioned before specifically so first is machine learning as a service tools family so effectively it's a machine learning endpoint or service which you can call uh you don't need to provision it separately and you pay per request which is awesome when you have low volume requests model can be retrained but it will need to be provisioned separately examples here are Amazon recognition and textract and Amazon comprehend next is sagemaker autopilot this is automl solution for tabular data or where you can train the model without writing code or choosing architecture so what it will do based on the data you provide it will try multiple models and it would choose the best performing one and you can easily integrate it with a sagemaker data Wrangler to have a custom data preprocessing also no code necessary so you can basically organize data preprocessing pipeline using UI tool uh next is sagemaker Canvas Plus ready to use or custom models uh this is no code machine learning interface where you can either use ready models ready to use models or create ml models on your own automatically so it's uh basically can be used for that exploration training models and it supports multiple data sources next is sagemaker jumpstart effectively this is a collection of pre-trained proprietary and publicly available models with easy way to deploy them this collection is continuously updated so for example now it has foundational model like stable diffusion and it has a simple one-click model deployment meaning you can easily deploy this model into your account and you will get a Jupiter notebook to test the integration and finally sagemaker ground truth that's a fully managed data labeling service so it's very flexible you can use it to label on your own or with your internal Workforce or you can use something like Amazon mechanical torque or third-party vendors for labeling another Advantage with it is that it has ml powered automation meaning it can train model based on your labels and show predictions to the labelers so they don't need to label from scratch so how to decide on the Tool uh which you need so if you need fully complete solution then integrated ml is your choice and you can use it to start getting value from your data right away if you have a very specific task then it's better to use integrated automl meaning you can train your own model and run predictions on your data if you want to make ml part of your application then you need to basically ask yourself like does machine learning as a service tool solve your problem if yes then you can just use it right away just integrate it if it doesn't perform well on your data then you can use automl as a service if that still doesn't work because it doesn't solve your type of task then pre-trained model is your choice and you would need to just deploy it without training if it's too expensive at your scale that means uh that you will probably need to organize the whole training pipeline so that you can optimize model for your infrastructure the thing here is that it doesn't necessarily mean that you have to commit to a particular solution right away it could be that in the beginning you want to start with machine learning as a service to get some early customer feedback then you can switch to another one for for auto analysis service so that it performs better on your task while once you have more data and finally once you hit large scale you can switch the managed pipeline because you want to optimize costs so now uh let's switch to sagemaker Canvas custom model demo so during this demo we will get into sagemaker canvas see how the uh how working with images works and how to train model and run predictions so here right away there are models you can use and this is what I mentioned like integrated ml so models which you don't need to train I prepared two data sets Wildcats train and Wildcats test which have two types of images snow leopard and cheetah so the good part here is that if you want to import data you can do it either from local data source or from S3 so if we go to my models basically here I already have a model I trained but effectively what happens during build stage you select model data which you want to use for training so cheetah and snow their part you just need to label it after it it will use part of this data for training and part for validation and it will show you model performance on this data and not only you can see predictions whichever made correctly or incorrectly you can also see the scoring uh and how well the model performed so you can see individual metrics for each label and finally you can run prediction job so either on individual image or make a batch prediction so for example I can select data set A Wildcats test and I can generate predictions it will take some time but I already have a previously run prediction on this data set so here we will see the predictions made by the model I trained and you can see that not only it will classify each image it will give me the confidence for each prediction and what I can do I can verify predictions where they have correct labels or not and I can add them to the train data set and once I'm I added them to the training data set now I can build a new model which would perform better because now it has more data so to summarize the representation uh low code no code machine learning tools can help you in different cases whether you're building your own product or you want to use machine learning for your day-to-day tasks thank you very much for your time and would we be happy to respond to your comments in the comment section [Music] thank you [Music]