23b0045_AIC

To make the code work:

1. Install the nessary libraries imported in each question by executing pip install (library) in the terminal.
2. Install the datasets required for each question. Further instructions on downloading the dataset are given later.

Question 1:

Dataset used : I have used PASCAL VOC 2007 dataset consisting of 20 different classes, Download the dataset from https://www.kaggle.com/datasets/vijayabhaskar96/pascal-voc-2007-and-2012

Work Analysis:

1. I have used the YOLOv1 algorithm for detecting objects as well as their bounding boxes. The model consists of a VGG model pretrained on the 'imagenet' dataset followed by 24 convolutional layers and 2 fully connected layers similar to the model in YOLOv1 paper.
2. I tried training the model on vscode but the training time was too much based on the size of the model.
3. So i ran this notebook on kaggle to use their free GPU's.

References : Research Paper of YOLOv1 -- https://arxiv.org/abs/1506.02640 https://www.youtube.com/watch?v=IA3WxTTPXqQ

Question 2:

Dataset used : https://investor.fb.com/investor-news/press-release-details/2024/Meta-Reports-First-Quarter-2024-Results/default.aspx

API Keys : Get LLamaCloud API key from https://cloud.llamaindex.ai Get ChatGroq API key from https://wow.groq.com

References : https://aksdesai1998.medium.com/improving-rag-with-query-expansion-reranking-models-31d252856580#:~:text=FlashRank%20is%20an%20ultra%2Dlite,from%20flashrank%20import%20Ranker%2C%20RerankRequest

https://medium.com/@onkarmishra/using-langchain-for-question-answering-on-own-data-3af0a82789ed

https://www.youtube.com/watch?v=HkG06wBbTPM&t=796s

Question 3:

Dataset used : https://ai.stanford.edu/%7Eamaas/data/sentiment/ Execute the first code cell in P2.ipynb to download the data

Work Analysis :

1. Max Length	Embedding Dimension	Train_Accuracy	Val_Accuracy
   75	16	1	0.7
   75	8	1	0.7725
   100	8	1	0.7870
   100	12	1	0.7805
   125	8	1	0.7930
   150	8	1	0.7950
   150	10	1	0.7990

Final Review :

1. As we cannot directly feed text to a model, we need to convert the text to some mathematical form. This is where word embeddings come to play.
2. Word embeddings vectorize a word. Similar words have similar embeddings. This similarity is checked through cosine similarity. The main reason we choose word embeddings is that they can capture relationship between words during training.
3. Word embeddings are better than one-hot encoding as in a one-hot encoding with a dictionary size of aroud 10000, a particular word will be represented by 9999 0's and 1 1's which is very inefficient.

References : https://www.tensorflow.org/text/guide/word_embeddings