50+ days of Machine Learning

 Hello :) Today is Day 147! A quick summary of today: attended day 1 of Microsoft Azure hackathon The learning objectives of the hackathon: Understand GenAI concepts and how they can be used to build chat applications Learn how to configure the Azure OpenAI service and use Azure AI Search to build private OpenAI with your own data Gain practical experience in implementing a chat app architecture with Terraform Explore features of chat apps, such as multilingual queries and advanced chat response settings Apply GenAI to real-world scenarios, including increased user interactions and dynamic document processing Operationalize AI-enabled applications with enterprise-level monitoring and logging Scale AI-enabled applications with enterprise-level load balancing We are assinged to a table and the people on a table work together as a team. It is all about using Azure and its services. There are in total 6 challenges, 3 for Day 1 and the others for Day 2. Below are some pics/notes from th...

 Hello :) Today is Day 146! A quick summary of today: did 2nd homework from MLOps zoomcamp on experiment tracking covered some material on devops on Azure MLOps zoomcamp Module 2 homework on experiment tracking The homework link is here . The homework followed the below outline: Install mlflow Preprocess NYC taxi data Train a model with autolog Launch a tracking server locally Tune model hyperparams Promote the best model My code for it is on my github . The used model was a random forest regressor using rmse as the main evaluation metric. The homework was more focused on running mlflow commands rather than the preprocessing/training (which were done in Module 1 of the course).  Below is a graphical comparison of some of the models I ran I had to set up logging and mlflow db for a sample run Run hyperparam optimization  And finally choosing and registering the best model For a bit I faced an issue with the params passed to the model because they were missing (initially no...

 Hello :) Today is Day 145! A quick summary of today: covered  course 1 for the hackaton: Azure Kubernetes and the cloud I don't think I shared before but on Wednesday I got confirmation that I can attend  Microsoft Azura x NVIDIA hackaton in Seoul  held on Monday and Tuesday 27th and 28th of May. It will be an amazing opportunity to visit the Seoul Microsoft office, learn from Microsoft professionals and in general participate in a hackaton. On the website they mention that if we want to prepare, they recommend covering two courses: the one I did today on 1.  building and modernizing apps  and 2.  improving productivity with github and Azure for devs  - which I will do tomorrow.  Before the material, there are some nice articles that I will review further But as for the taught material, below is the outline Notes I took: K8r (Kubernetes) secrets: Store sensitive information (like passwords) safely not in plain text Encoded in base64 which ma...

 Hello :) Today is Day 144! A quick summary of today: started reading papers on using graphs for taxi demand and origin/destination predictions One of the papers my professor wrote is on  Measuring taxi ridesharing effects and its spatiotemporal pattern in Seoul, Korea  and because I am learning about GNNs from Stanford's XCS224: ML with Graphs  I remembered that one of the most popular use cases of graphs is on traffic prediction.  I read 2 papers: GNN-based Passenger Request Prediction Origin-Destination Matrix Prediction via Graph Convolution: a New Perspective of Passenger Demand Modeling I actually found the 2nd paper from the 1st paper because it was referenced. There are these 4 models used as comparison vs the proposed GNN-based model As for the models, the 1st paper kind of goes off from the 2nd one on the idea to separate a city into grids.  and construct an OD (Origin-Destination/Adjacency) matrix based on that with requests from one cell to anot...

 Hello :) Today is Day 143! A quick summary of today: did forward and backprop of a toy neural net by hand and confirmed with python set up my station at the student lab at uni The idea of doing backprop manually again came back yesterday night. But this time I wanted to do it 'as in the old days'. In one of the lectures of CS109 Professor Chris Piech said he used to do forward/backprop manually back before autograd, so I wanted to experience that as well (in a toy example).  Below are my handwritten notes of the forward, backprop and param update for the 1st iteration of a toy neural net.  My tablet battery died towards the end so I wrote it on the board. I confirmed all the numbers with python - the used notebook is here . As for my 1st day at the lab I liked it. It was mostly waiting for downloads and setting up cuda/torch/python/etc because it was a new PC. There was a seminar where 2 of the students talked a bit about what they are doing. At the end of the day, I had...

 Hello :) Today is Day 142! A quick summary of today: did my 4th assignment for the XCS224W:ML with Graphs course interviewed for a lab assisstant at my university's Big Data department Unfortunately, as with the previous assignment s, I am not allowed to share anything from them. I can only share about the topic which was heterogeneous graphs (which I learned about on Day 103 ). It took a bit of review, coding and debugging torch/cuda version disparity, but ~ I got the max score in the end. As for the interview ~ yesterday I sent an email to the head of the department of the Big Data department at my uni - Professor Minju Park, last night he responded and invited me today at 12 at his office. We had a chat, answered questions, ambitions about research, projects, and at the end ... he said there is 1 space left in his lab and that I can join. The lab is all undergrads and myself (the only non-korean) and I am excited to study, learn, share about anything related to AI - from math ...

 Hello :) Today is Day 141! A quick summary of today: read a bit more of A Primer For The Mathematics Of Financial Engineering by Dan Stefanica applied (sent an email) about becoming a TA for one of the undergraduate big data/ML/stats courses in my uni I like the Primer book because it is pure math - just Lemmas and Theorems and proving them (+ application in finance). Below are my notes from Ch 4 - lognormal random variables I saw that a deadline for the 4th assignment from Stanford's XCS224W: ML with Graphs is due on the 28th so I will likely start doing it in the afternoon after a quiz I have at uni.  Also I saw that the official 2nd homework on model management and experiment tracking from the MLOps zoomcamp is up, but let's see if I can do it tomorrow. As for the 'application' to be a TA. I believe teaching is the highest form of knowing something. Being able to explain a concept to someone who does not know it and them understanding it is one of the biggest cert...

 Hello :) Today is Day 140! A quick summary of today: Ch 1 - calculus review and plain vanilla options Ch 2 - improper integrals, numerical integration, interest rates and bonds Ch 3 - probability, Black-Scholes formula, Greeks and hedging  From the books I found yesterday I thought that starting from this one is the best because of its pure math + real application format. Each chapter consists of some math/stats and also how they are used in the real world. Ch 1 - calculus review and plain vanilla options Ch 2 - improper integrals, numerical integration, interest rates and bonds Ch 3 - probability, Black-Scholes formula, Greeks and hedging  On another note ~ today I realised that calling this whole blog project self-study into AI might not be the best. Because math, probability and stats are at the root of AI. So maybe some kind of renaming would be appropriate.  That is all for today! See you tomorrow :)