Yogesh Joshi

Uh, myself, Yuki Tushy. I have about nearly 5 years of experience in, uh, data science, AI, ML, as well as, a Python development role. I have a versatile portfolio where I worked on many different types of projects, including machine learning, mostly deep learning and computer vision. My expertise lies within deep learning and computer vision, but as the trend has more diverted towards LLM and the large language models. I have also contributed myself to getting familiarized with Tuuk's technologies. Uh, about, uh, in a year ago, I have started working on NLMs. One of the projects that I am working on is based off of visual transformers as well as, uh, the latest, uh, video, which has been released, uh, with the code. So I have, uh, expertise in Python language as well. Apart from that, I have good understanding of core principles such as, uh, database managements, handling pipelines of data, handling deployment. Uh, I'm also eager to learn new technologies. So I have recently gained exposure to MLflow and other such deployable uh, technologies. I also have, uh, expertise in the latest upcoming post post PostgreSQL, this database.

The techniques depend on the and the size of data which we need to train. Uh, most preferably, I'll use continuous development and continuous integration techniques where I would, uh, have a base model prepared on the large dataset. And depending on the base model Accuracies and base model results performance benchmark on different sets of, uh, dataset. I'll Build up on that by hyperparameter tuning. I can do manual hyperparameter tuning depending on the experimentation as well as I can choose To opt for optioner, which is an automated hyperparameter tuning tool, there are also some different libraries such as the weights and biases and, uh, TensorBoard which can help in logging and Monitoring the different types of parameter used for experimentation.

Uh, there are different types of understanding of loss functions. Uh, primarily, uh, the Problem that we are, um, we are tackling all the algorithm that we are using, uh, depending on what Loss function has worked the best in the research paper. That is one of the mechanism that I'll use, uh, while approaching the deep learning and choosing the loss function, but in case there is some Other resources or other material which suggest better, I can opt for that particular loss question as well, But, primarily, uh, the research paper, the published conferences suggest the loss function which has been performing better in this case, and my approach would be to take that particular loss function. In case I have, uh, to switch loss function, my ideology would be to have a, basically have a, uh, experimentation approach where I would consider opting for multiple, uh, similar type of loss functions. Example, if I'm choosing, uh, cross entropy loss function, then it might be Also beneficial to make sure that, uh, it works well on the classification type of data.

Assigning task is, I think, what is primarily important when we are working on cross functionalities Depending on any manner they might be. And as the latest trends and technologies have been evolving, I think There have been incorporations to implement multiple sources or multiple types of technologies to get an better output. Uh, one of those example, uh, might be a Postgres ML where now the, uh, you can have a model, Uh, model inferencing through a basic SQL script. So I think that is something which I take as a lesson, And I would prioritize the individual tasks, uh, then the team task because once the individual tasks have been fulfill and met, then only the dependencies towards the other team members task can be functionally fulfilled. So while implementation in any kind of implementation of AI feature in a cross functional development, I would prioritize on an individual task than the team manner because in the cross functional platform or cross functional Environment, individual, uh, ideologies or individual work is more important than the teamwork because The individual's work is dependent on the cross functional team's

Uh, there are multiple approaches, uh, to solve, uh, the to tackle this particular problem of implementing a with continuously training and NLP model with new incoming data. Uh, one of the approach Or the most efficient approach is to have a, basically, have a, uh, storage mechanism where we have incoming traffic then the storage mechanism is attached to a data pipeline. The data pipeline job is to clean the data, have the data properly established or properly ready for the model, then there will be a training pipeline. The training pipeline's job will be to take that data which has been cleaned by the data pipeline and train the model whichever it may be to perform the NLP task. Once the training task is done, then there will be 2 types of pipelines. 1 can be a deployment pipeline and another is an inference Now within the deployment pipeline, there is again we need to add a scheduler or a scheduling mechanism where it would periodically depending on the solution that, uh, the depending on the solution that is required it would train it retrain itself, but also the doesn't forget, uh, doesn't have an catastrophic forget mechanism where Totally forget the previous, uh, rates and ready and have an some kind of different output. So that is something we need to monitor. Uh, to tackle those problems, I think newer, uh, trends and newer technologies such as vector databases can be implemented to basically have a, uh

Given the scenario that the model trained Nice. Q. The mitigation strategies are basically, there are multiple mitigation strategies to avoid biases in the data. One of which is to incorporate, uh, and representative data where it basically represents all type of data where Whichever problem we need to solve, one of one of the approaches by going in such a direction. Another way is, uh, that Demographically, we need the data to show all the regions and all the countries. That is also one of the approach that we can use to mitigate the skewness and the bias in the data. Uh, there are also some other, There are also some other imputation and augmentation techniques that can be used to have the data normalized. Yeah. That's it. I think these are the strategies that we can use. Apart from that, there are also some other strategies That can be researched and implemented depending on the data and the scenario that we are working on.

Self attention mechanism wouldn't take 3 inputs of x. I think it would be a single input. Uh, the basic transformable lock has the attention layer with forward feedforward neural network, so I think that is something which is, uh, missing. And also before the attention layer, there should be some additional layers such as embedding layer In case we are incorporating it within the transformer block, uh, so we can add, uh, the positional embedding vector layer also within the transformer block. I think that is something. The input shape for the self attention mechanism, I think that is something which is

Uh, absolute mean mechanism is somewhat often, uh, nonstandard for Loss mechanism in LLM. The most frequently used or most promiscuously used, uh, loss mechanism for Any kind of generative model is mostly the root mean squared errors. So that, uh, so that the steps so that the step that they are taking towards gradient is higher, and it reaches, uh, this it reaches its, uh, global minima faster. So I think that is something we can implement here.

There are multiple, uh, readily available architectures which we can use for text generation. Uh, one of which is using the But mechanism, which is bidirectional encoding bidirectional encoding and coding representation of transformers, uh, where basically it has only the decoder layer And, uh, stacking the decoder layers together can be helpful index summarization problems, uh, whereas The output layer can be a soft match layer.

Uh, for ingestion pipeline, there are Multiple cleaning and preprocessing steps that can be done, uh, beforehand to identify the anomaly. One of the steps while condition of data is to clean the data and have a, uh, preprocessing done on the data. This, uh, particular step can be useful in avoiding the anomalies. I think that is a particular step that can be done to avoid anomalies in the data. Apart from that, we can also have a cloud watch mechanism where it, uh, checks the data. It has the crawlers. We can use, uh, different types of crawlers even before having the data ingested into the pipeline To basically, uh, sort and, uh, eliminate any kind of

There are different types of benchmarks to validate output, uh, based on the LLM. So The sentiment analysis part, I think there are different benchmarks readily available. I don't, uh, remember, uh, on the top of my head, the benchmarks to validate the particular particular process. But, uh, apart from The readily available benchmark, we can also have a human in loop mechanism where the human itself can go through that result output and validate it if the if the output is somewhat acceptable

Um, depending on the use case, we can choose either TensorFlow or PyTorch. In both cases, either or or, uh, I think the, Uh, programmatically and logical consensus between TensorFlow and PyTorch are somewhat similar. So either or or either choosing TensorFlow flow or choosing PyTorch won't hamper the overall development, but for the simplicity's sake, I would prefer PyTorch as it is open source, Although TensorFlow is also open source, but it has its own limitations and works well with TPU TPU, uh, format arch architecture, whereas PyTorch, it can handle almost all types of architecture. The conversion of TensorFlow and PyTorch to optimized version, uh, to an quantized or optimized situation are somewhat similar. Both can be condensed into smaller forms. Both has batch fetching mechanisms which can be useful. Uh, both have their own, uh, limitations and enhancements, but I think either or are Fine. Fine. Uh, while choosing development, uh, transformer based texture mostly most of the examples of transformers have been made available by PyTorch. So I think PyTorch is more preferable Then TensorFlow as the back end for the transformers hugging v c p I uses, uh, PyTorch. But, uh, again, TensorFlow, if we are choosing the, uh, choosing to develop Google based architecture like Bard, then I think TensorFlow is much better than Python. So depending on which you choose for the alternative model. It can, uh, it can be beneficial to have a uh