Embark on a cutting-edge journey at the intersection of vision and language, where our innovative project seamlessly merges pre-trained Vision-Language Models with real-world knowledge, addressing modality bias and object hallucination challenges while offering adaptability to diverse cross-domain data.
Cross-Domain Vision-Language Understanding:
Explore the seamless integration of a pre-trained Vision-Language Model (VLM) with real-world knowledge from a Large Language Model (LLM) to generate textual descriptions for images across diverse domains, eliminating the need for fine-tuning on specific image-text pairs.
Mitigating Modality Bias:
Tackle the challenges associated with modality bias by developing a model that effectively bridges the gap between visual and textual representations, ensuring a more unbiased and accurate portrayal of image content in generated descriptions.
Object Hallucination Resilience::
Address the issue of object hallucination in image descriptions by implementing robust mechanisms that enhance the fidelity of generated textual outputs, promoting a more reliable and contextually accurate understanding of the visual content.
Adaptable Image-Text Pairing:
Achieve a high degree of adaptability to cross-domain data by designing a solution that does not rely on specific image-text pairs for fine-tuning, allowing the model to generalize effectively across a wide range of visual and textual inputs.
Next-Level Vision-Language Synergy:
Showcase a project that leverages the strengths of both Vision-Language and Large Language Models to create a synergistic approach, providing a comprehensive understanding of images with rich contextual descriptions, reflecting real-world knowledge without sacrificing
Decision Optimization in Cricket Gameplay
In the dynamic world of cricket, making the right decisions on each ball can be the difference between victory and defeat. This project was dedicated to enhancing decision-making in cricket gameplay scenarios by leveraging principles from Markov Decision Processes (MDPs). Here’s a summary of our key accomplishments:
Contextualizing Cricket Gameplay:
The project aimed to provide an optimal action in various in-game situations. To achieve this, the focus was on modeling the game setting as an MDP.
MDP Conversion:
The project transforms the intricate dynamics of a cricket game into a formal Markov Decision Process. This conversion process allows to mathematically represent the transitions between game states, enabling more structured decision-making.
Algorithm Implementation:
One of the core contributions of the project was the implementation of various MDP solution algorithms, including Value Iteration and HPI (Howard's Policy Iteration), from scratch. These algorithms played a pivotal role in determining the optimal actions in different cricket gameplay scenarios.
Decision Optimization:
The work ultimately led to the optimization of decision-making in cricket. By applying MDP techniques and the custom-built algorithms, the algorithm recommended optimal actions based on the given probabilities of different outcomes for specific actions in the game.
Regret Minimization in Multi-Armed Bandit Problems
In the realm of decision-making under uncertainty, multi-armed bandit problems serve as a quintessential challenge. Our project was dedicated to devising effective strategies for minimizing regret – the difference between the reward accumulated by an algorithm and the maximum achievable reward. Here are the key highlights of the work
Algorithm Diversity:
The project confirms the theoretic logarithmic upper bounds on regret for different algorithms including UCB (Upper Confidence Bound), KLUCB (Kullback-Leibler Upper Confidence Bound), and Thompson Sampling. These algorithms formed the backbone of our approach, enabling us to explore various avenues for regret minimization.
Batch Thompson Sampling:
An innovative aspect of the project involved the implementation of Thompson Sampling in a batch setting. This adaptation was aimed at minimizing regret when samples are taken in batches.
Addressing Special Cases:
The project addressed a unique scenario in bandit problems where the number of arms matches the horizon. In such cases, we provided a straightforward yet effective approach to minimize regret, ensuring that even under these specific constraints, our strategies remained optimized.
Refractive Index Prediction with Terahertz Time Domain Spectroscopy
The project focused on developing an advanced optimization algorithm for accurately predicting the refractive index of unknown materials using Terahertz Time Domain Spectroscopy (THz-TDS). Key highlights include:
Cutting-Edge Algorithms:
Harnessing the power of optimization techniques such as Dual Annealing, SHGO, and Newton Raphson to ensure precise predictions.
Python Library:
Creating a comprehensive Python library that encapsulates our optimization algorithms, serving as an educational and research tool.
Material Characterization:
The provided solution enhances material characterization in fields like optics and electronics by providing accurate refractive index predictions.
Educational Impact:
The project offers a valuable resource for students and researchers interested in THz-TDS and refractive index prediction.
Interdisciplinary Collaboration:
The project combined expertise in optics, spectroscopy, and optimization to develop a holistic solution.