AI Engineering Intern @FixitAI | Incoming Software Engineering Fellow @Headstart AI | MEC'24

github

Sony Thomas  • Peerlist

🔍 Unveiling the Power of Deep Learning: Character-Level Token Prediction Using MLP 🔍I’m excited to share my latest work on a cutting-edge Multi-Layer Perceptron (MLP) model designed to predict character-level tokens and form coherent words. By focusing on text generation, this project encapsulates the core challenges of tokenization and neural network optimization.<h3 class="rich-heading">Project Highlights:</h3><ul><li>Objective: Predict the next character in a sequence by analyzing a context window of 3 characters.</li><li>Vector Encoding: Each character was encoded into a 10-dimensional vector space, reducing complexity while retaining linguistic features.</li><li>Network Architecture: The model utilized a hidden layer with 200 neurons, ensuring both computational efficiency and learning capacity.</li><li>Training &amp; Results:<ul><li>Achieved a Cross-Entropy Loss of 2.12 on the training set.</li><li>Attained a Cross-Entropy Loss of 2.16 on the dev set, reflecting robust generalization.</li></ul></li></ul>This project allowed me to dive deep into the intricacies of sequence modeling, where I experimented with bigram modeling and tokenization strategies to enhance the quality of text prediction. The model was designed with scalability and optimization in mind, showcasing my ability to build sophisticated neural architectures from the ground up.<h3 class="rich-heading">Why This Matters:</h3>With language models at the forefront of AI, the ability to efficiently predict sequences at the character level opens up opportunities in NLP applications, from chatbots to text generation engines. This project demonstrates my skills in:<ul><li>Neural Network Design &amp; Training</li><li>Natural Language Processing (NLP)</li><li>Deep Learning Model Optimization</li></ul>Feel free to check out my open-source project and dive deeper into the technical details: <a target="_new" rel="noopener noreferrer nofollow" class="text-green-text rich-link break-all link-break" href="https://github.com/STSonyThomas/bigram-llm/blob/main/makemorep2.ipynb">GitHub Project Link</a>.I’m always keen to collaborate on AI/ML initiatives and explore new possibilities in language models!#DeepLearning #MachineLearning #NLP #AI #CharacterPrediction #MLP #NeuralNetworks #OpenSource #TechTalent #TextGeneration #Recruitment #AIResearch

Character Level Token Prediction

Link: https://github.com/STSonyThomas/bigram-llm

Bigram Statstical Model and Neural Network from Scratch

Link: https://github.com/STSonyThomas/neural_micrograd
Built a neural network from scratch aiding in visualization of how neural networks make use of *backpropagation* to train on a set of data. Created the backpropagation algorithm by self.

Planning on building PyTorch from scratch :) Let's see how that goes.

Neural Network From Scratch

A resume analyzer that parses your PDF document, gathers information about the job being applied to from the user and provides suggestions to bridge the gap in skill and job.

· Uses pyresparser from Resume Parser package to parse through the resume.
· Uses NLTK and Spacy for Natural Language Processing.
· Uses CountVectorizer to convert text to Vectors.
· Uses cosinesimilarity to compare given resume keywords against job descriptions keywords.
· Uses streamlit for user interface.