Project Background:
Symbolic reasoning using deep learning arises from the intersection of two traditionally distinct paradigms in artificial intelligence: symbolic reasoning and deep learning. Symbolic reasoning rooted in classical AI involves manipulating symbols according to logical rules to perform tasks such as deduction, inference, and logical reasoning. On the other hand, deep learning has gained prominence for its ability to automatically learn representations from data often surpassing traditional methods in tasks such as perception and pattern recognition. However, deep learning models cannot typically perform explicit symbolic reasoning, relying instead on statistical patterns learned from large datasets. The project aims to bridge this gap by integrating symbolic reasoning capabilities into deep learning architectures, enabling models to perform tasks that require logical inference and symbolic manipulation. By combining the strengths of both approaches, the project seeks to develop AI systems that can reason symbolically about complex problems, interpret and manipulate structured data and perform tasks requiring logical deduction, with potential applications in natural language understanding, robotics, and scientific discovery.
Problem Statement
- Integration of Symbolic Reasoning often lacks explicit capabilities, limiting their ability to perform tasks requiring logical inference and manipulation of symbolic representations.
- Interpretation of structured data, such as graphs, logical rules, and symbolic representations, struggles with.
- Lack of the ability to generalize and abstract knowledge in a way that enables reasoning across diverse domains and problem contexts.
- Often lack interpretability, making it challenging to understand their decision-making processes, especially in complex reasoning tasks.
- Incorporating prior domain knowledge and logical rules into deep learning architectures presents challenges in model design and training strategies.
Aim and Objectives
- To develop deep learning models with integrated symbolic reasoning capabilities for tasks requiring logical inference and manipulation of structured data.
- Integrate into deep learning architectures to enable logical inference and manipulation of symbolic representations.
- Develop techniques for interpreting and manipulating structured data such as graphs, logical rules, and symbolic representations within deep learning frameworks.
-
Enhance the generalization and abstraction capabilities to enable reasoning across diverse domains and problem contexts.
-
Improve the explainability and transparency of symbolic reasoning tasks, enabling a better understanding of their decision-making processes.
- Explore strategies for integrating prior domain knowledge and logical rules to enhance performance and adaptability to specific problem domains.
Contributions to Symbolic Reasoning using Deep Learning
- Developed deep learning models with integrated symbolic reasoning capabilities, enabling logical inference and manipulation of structured data within deep learning frameworks.
- Advanced techniques for interpreting and manipulating structured data such as graphs, logical rules, and symbolic representations, facilitating reasoning tasks in diverse problem domains.
- Improved deep learning model generalization and abstraction capabilities, enabling reasoning across diverse domains and problem contexts.
- Enhanced the explainability and transparency of deep learning models for symbolic reasoning tasks, facilitating a better understanding of their decision-making processes.
- Explored strategies for integrating prior domain knowledge and logical rules into deep learning architectures, enhancing performance and adaptability to specific problem domains.
Deep Learning Algorithms for Symbolic Reasoning using Deep Learning
- Neural-Symbolic Integration
-
Neural Turing Machines (NTMs)
-
Graph Neural Networks (GNNs)
-
Neural Theorem Provers
-
Differentiable Neural Computers (DNCs)
-
Neural-Symbolic Machines
-
Recursive Neural Networks (RNNs) for structured data
-
Attention Mechanisms
-
Graph Attention Networks (GATs)
Datasets for Symbolic Reasoning using Deep Learning
- AGENDA
-
ATIS (Airline Travel Information System)
-
bAbI Tasks
-
CLEVR (Compositional Language and Elementary Visual Reasoning)
-
COGS (Comprehensive, Objective, and General Standardized) dataset
-
DROP (DuaL-Objective Reasoning Dataset)
-
Geometry Questions
-
SCAN (Semantic Composition Through Abstract Reasoning) dataset
-
SocialIQA
-
Textbook Question Answering (TQA)
Software Tools and Technologies
Operating System: Ubuntu 18.04 LTS 64bit / Windows 10
Development Tools: Anaconda3, Spyder 5.0, Jupyter Notebook
Language Version: Python 3.9
Python Libraries:
1.Python ML Libraries:
- Scikit-Learn
- Numpy
- Pandas
- Matplotlib
- Seaborn
- Docker
- MLflow
2.Deep Learning Frameworks:
- Keras
- TensorFlow
- PyTorch
