Anubhav Kamal

Yeah. So I am I'm Anwar Samil. I have graduated from Indian Institute of Technology, director I have a master's degree in mathematics and computing. I've worked with an organization called Ceramosic. It is a start up. Of building semiconductors. We were working on a propriety hardware called neural processing unit or an NPU. Currently, I'm working with Samsung Semiconductor India Research. I'm in AI computing department there. So I've been working on a propriety hardware called processing in memory. With a lot of AI and ML driven projects in between, I've been working with in collaboration with foundry team to work on machine learning related tasks and solution of the real problem that they were facing. That they were facing related to their PDK problem. They were trying to find an analytical solution 2 to do their PDC, and we help them And this is in the domain of action enabled AI. I use a technique called, symbolic regression and a Python package called PISR. And, I've been working with AIML and like, for past 3 years, First with Ceramorphic, in the deep learning compiler. And now with Samsung Semiconductor India Research on voice, processing in memory and other projects.

And then balance it in it.

When integrating cell phone model into production, which has been in the at least in model service. model service. I'm not aware of the term with modules. So when I integrate things, PyTorch module into production is challenging task in itself. So we were working with Python's lite and integrated and tried to, run Python's lite on ARM devices on, like, mobile phones and stuff. So we use, Samsung S 24 Ultra to deploy Python Lite and, those were, like, the problem that we were facing was the compatibility issue with the architecture and building pipelines. Right? And to, with certain compiler and stuff. So while building things with ARM devices, we generally use, C line plus or LLVM compilers. And, the version mismatch can happen with Python. And those are the major issues, along with some issues in Linux that you can see. So I'm not aware with model serving and stuff.

So we do this computation first. So once you have dimension reduction okay. So dimension reduction can be applied in various form. 1 of the, popular technique, it's called PCA. The full form is, component principal component analysis. So it is a linear algebra based technique where you find the importance of each of the features and either combine those features with given weights. So it's, or you if the, like, importance of feature is very low, you actually remove them. So that is 1 of the technique, and that is the most popular technique that we use because it does not automatically, make a feature useless, but it can also combine the feature into 1. that is 1 way. Other is to perform various feature engineering on the whole dataset and, see if, like, the correlation matrix and, like, find how much, data is correlated with the how much each teacher is correlated with the output. And, so there is a simple correlation formula depending on each, random variable. And, that can be organized into a matrix and giving a number between minus and 1 and 1. Minus 1 being inversely correlated. That is, like, absent like, if, the output is going is increasing, then your feature would be decreasing. That is absolute, negative correlation. And 1 is, positive correlation and 0 is no correlation. So some a number around 0 is something we do not want into our dataset, and that's how we can reduce the dimension of the data as well.

I think we're just lost in dollars. So for class and moments, A class time balance can be, address with certain, I'm not sure about this patient. No. Yeah.

When setting up supervised learning pipeline, like, it is deep consumerizing, which are in GMS directions. So, while setting up supervisor machine learning pipeline, there have few consideration to keep in mind. Something like you always make sure that the split of the dataset is random enough to give it, like, whole holistic view of the entire dataset, what we are working on. The second is, when working with the feature engineering, you always, normalize the data and then continue the feature engineering. You always make sure, you always make sure that the output and input the output is not need to be supervised. Output is not need to be normalized with the input is. And, you know, you and you always select right feature engineering tools and right feature engineering methods to go about the feature engineering. 1 of them can be data visualization with different tools. So selecting a couple of feature engine features and plot it against the output data. That can give you a sense of how each feature behaves along with the data. There can be, nonlinear, relationship between output and the features. but there should be some correlation. The second would be looking into the correlation matrix and selecting the feature with high correlation shows. And, yeah.

Let me start at 1 end. It's been the issue and how often I get experience. So the optimizer here has been used as, opt in dot SCD. And then we have in files module dot. We have not specified what kind of optimized code yet.

In the 2nd system, it's in the 1 division. It requires and it's saying why it's not much. Much outside. It's accepted. It's a when we need this and accept to generate a short. So division where she was handled, that is really the major concern when you're dividing to number. Right now, in the second print section, we are dividing 10 with a string 2. That is also a major issue because, that is fine. And then we are dividing 2 numbers, not 1 number within, string. So we need to either convert to into a number or this is for an, for an error in Python.

So, I have deployed a few modules on mobile USB or SD license. basically, on ARM CPU. And to look into the latency issue, we generally benchmark server or benchmark any application we are working on. We benchmark using, some of the benchmarking tools, SimplePerks, is a great benchmarking tool. You can directly launch it from ADB. And, what you can also you can get all the cache information, all the memory information, and all the, all the computation information as well. And how you can decide whether to, you know, work on, like, working on the computation part or there is problem with the memory management in the system or the cache is under or overutilized and mostly underutilized. So, in certain cases, like, you make according to the need, you make certain changes.

So selecting a 3 day model. So it should be trained on a very similar dataset that I've I wanted to work on. it should, the data the input dataset should be, of similar or same size that I want, it will work on the filter and the number of, number of image layer that we want. So, the RGB layer, so, channel base should be the same as, we want to we wanted to make we want our, testing image to be. And, it should be tunable enough. The it's the source code should be available enough that we can make enough changes that we can debug the code and fine tune the model for our needs. And also, like, frame the gradient and stuff in between if required.

We want a CW client for the 7.99, so that's 1008, because Okay.