Rakesh Kumar Mallik

Hi, I'm Rakesh. Currently, I work as a senior machine learning engineer at Reliance Jio Platform in India. I have been working in various machine learning models, especially in different domains, such as production, supply chain, and HR. And now at Reliance, I'm working on industry safety optimization where we use deployed models in factories to detect people, restrict zone entry, detect falls, and this understanding will help design a robust industrial setting. So, I have worked with different technologies. I was introduced to machine learning, especially deep learning, since my MTech life at the Indian Institute of Science almost 5 years ago. I have worked in TensorFlow, PyTorch, and also a little bit of Keras and other different technologies. I am very good with Python and PyTorch. I deploy end-to-end machine learning models, train from scratch, design optimal object detection systems, train data, prepare data, deploy models, and finally evaluate the models and generate reports end to end.

So, the complexity is a subjective word. In order to release the complexity of the model, we need to understand if the model is supposed to do what it's supposed to do. For example, if you give me a task where it detects a person just, or if you give me a task of classifying a person, then instead of going to a very huge and heavy model like object detection, segmentation, key point, and whatever kind of model, these are very heavy parameter-wise. I can simply go to a ResNet and do a classification. So, to understand the model, to redo the complexity, we need to understand what the model is doing. And given that, okay, we understand now. This is the model, and we're already deploying a model that the industry is doing. Now to reduce the complexity, that so many things we can do. 1, we can reduce the parameters of the model. We can see what are the different layers which are without, also the model can perform as good as the actual model. We can go and check other different cost functions, the loss function that's being used. Is it too expensive? We can have a loss which can solve our problem and which can also do the work with less cost. So, last, the computation. We can review the computation this way. We can look at it this way: if we want to reduce the complexity of the model, we can also do a better preprocessing where, like, we take care of some of the aspects where the model doesn't have to do a lot of preprocessing and post-processing. So, everything we don't need to give to the model. For example, let's say we do object detection where we predict some bounding box. Now we can have a simple NMS algorithm which is non-max suppression, which actually gives me a non-interacting bounding box with some threshold confidence. So, preprocessing, post-processing, that will also help to kind of take away some of the complexity that we do in the modeling. So, these are the modeling aspects I believe that can help to decrease the complexity of the deep learning model while preserving the predictive powers. Thank you.

To assess an adversary attack, we need to understand what kind of data and what kind of distribution the model was trained on. We try training a model on a particular set of datasets and with a train. However, if you infer an unknown dataset, then the model will not perform well. To assess, we need to keep monitoring the model's performance. We need to keep monitoring the model's performance. Let's say we have a model deployed with 80% or 85% accuracy, and we need to monitor the model's performance so that it doesn't fall below a certain point drastically. For example, if it's 30% tomorrow, that means something has totally gone wrong. Now, how to mitigate this? To mitigate this, as I said, we need to train the model on very generic, robust features, and we need to keep monitoring the model's performance to understand if there's a chance of total distribution or trend that prevents the model from performing a given task.

I do not have much experience in LLM, but I do have experience with GPT-based pre-training on a given book corpus and trying to query them and deploy them. So I believe, to deploy your LLM-based model, we need to first understand what problem we are solving. And we need to select the model. What kind of model do we want? Do we want a very generic, open-ended model that has been trained on various topics and can query on various things? Second, we need to understand the size parameters. How large is the model? How many tokens has it been trained on, and how many parameters does it contain? Then we need to understand the inference time, how long it takes to infer. We also need to have constraints on the model's response. We cannot let the model respond to anything and everything without any limit. If we want, it should be limited. And finally, deployment. So the ML pipeline should be such that we detect, in advance, any lag or downtime in the model-deployed server. So, there are different tools that can actually detect if the model is down. This kind of analysis will help safely deploy a model and reduce downtime.

There is always a trade between costs and accuracy. Computational cost versus accuracy. Yes. So many times, we especially in financial prediction where accuracy is a critical matter. So, in order to, but that doesn't mean we should have a very huge competition that will slow the product, which will slow down the total prediction. A high competition can lead to a high-accuracy model, but a high competition will lead to a slower model. To do this, we need to have a proper balance between speed and accuracy. We need to find the sweet spot that is acceptable by the business, so that it will not hamper our prediction while maintaining a threshold of minimal decrease in performance that is tolerable. For example, let's say we are predicting the stock price. To predict the stock price, there are different ways to do it. We can have time series analysis like classical models. We can have LLM. We can also train a deep neural network with a large number of parameters. So, having an edge neural network theoretically says it can solve an infinite complexity problem. In theory, a neural network can solve very complex prediction problems. But if you want to deploy that model, the compression cost is very high, and also the interest time will be more. On the other hand, if you have a simple regression like random forests or time series analysis, for example, random forests, which are highly overfitted models that learn based on trends and then make predictions quickly, but where we can have a simple equivalent model like time series, which will understand the trend of a few seasons and try to predict based on that. And that can solve our problem. It is imperative to understand the sweet spot, as I said, between complexity and performance. We can have, if it's required, a very fast response, then we can go as simple as a regression model or a random forest model or a decision tree model to give a prediction. But if we need to achieve 99% accuracy, then we can go as fast as any of the latest models.

Well, okay. So if we are having a kind of overloaded model, it definitely means we're using more number of parameters than what is required. And it's all that needs to consume memory will need to consume it by all the parameters to do the inference. So we can do is we can quantize the model and we can reduce the number of neurons based on the training and performance variation so that we can get as much performance without actually reducing it, maybe with some threshold of 5 to 10% of lag in the performance, but it can definitely decrease the number of computations we do. We can decrease the number of computations, and we can. That will help in quantizing the model using TensorFlow. These are the already available tools which we can use nowadays to decrease the model and quantize on top of that to make it more faster. Yes.

I'm a Python person. I've literally never coded in JavaScript, but I'll give it a try. Okay. So since it's just an array of strings, there's a chance it may contain null values or empty strings. So filing a tech to the upper class, it will fail. Like, I hope, that makes sense because we are trying to convert every item in the string array to upper case. If it's empty or null value, then it will fail.

So it looks like you've fully connected, dense network that we have been getting is an input layer and one is the output layer. And the best size kind of looks suspicious to me. We're giving 32x32 data points and having only 10 inputs in our neural network, which is something that's going to create a problem.

I'm going to. The simple rule to avoid overfitting is to have more data. And to have more data, and that can be achieved in different ways. We can apply augmentation to generate more data. So, if the model isn't fixed, we can choose a smaller model so that we can solve a small problem. One thing is we can generate more data, and the second thing will be to have a model that's appropriate for our task and our market prediction.

So the continuous model and innovation, the con interview, some kind of test. You're not with somebody? Some test. What is it? Some kind of recording. Some people are there. No. It's a kind of screening, video screening. Who is that in the? We'll allow you. Well, the AI field is continuously evolving, and since it's continuously evolving, like, one way to keep track of it is to keep reading. And, I suggest focusing on one area rather than focusing on all areas, but attending more and more conferences. And, focusing on one particular stream while having a broader view of the other streams, that will help.

Both. I'm on the laptop, Bruce. The phone. Last person. Extract an analysis up inside from them. So to extract analysis, yes. See, when you generate from the data from LLM, there will be noise. There will be garbage. Like, no, as you know, if LLM doesn't have actual contact, we'll have to hallucinate. They have to provide some hypothetical data. So if we are using NSA data for decision making, then we should have a bounce and check so that it doesn't go away from reality. That's something we should do.