Original Paper Information:
SPINE: Soft Piecewise Interpretable Neural Equations
Published 2021-11-20T16:18:00 00:00.
Category: Computer Science
[‘Jasdeep Singh Grover’, ‘Harsh Minesh Domadia’, ‘Raj Anant Tapase’, ‘Grishma Sharma’]
Relu Fully Connected Networks are ubiquitous but uninterpretable because theyfit piecewise linear functions emerging from multi-layered structures andcomplex interactions of model weights. This paper takes a novel approach topiecewise fits by using set operations on individual pieces(parts). This isdone by approximating canonical normal forms and using the resultant as amodel. This gives special advantages like (a)strong correspondence ofparameters to pieces of the fit function(High Interpretability); (b)ability tofit any combination of continuous functions as pieces of the piecewisefunction(Ease of Design); (c)ability to add new non-linearities in a targetedregion of the domain(Targeted Learning); (d)simplicity of an equation whichavoids layering. It can also be expressed in the general max-min representationof piecewise linear functions which gives theoretical ease and credibility.This architecture is tested on simulated regression and classification tasksand benchmark datasets including UCI datasets, MNIST, FMNIST, and CIFAR 10.This performance is on par with fully connected architectures. It can find avariety of applications where fully connected layers must be replaced byinterpretable layers.
Context On This Paper:
The main objective of this paper is to introduce a novel approach to piecewise fits in Relu Fully Connected Networks that enhances interpretability and ease of design. The research question is whether using set operations on individual pieces can improve the interpretability of these networks. The methodology involves approximating canonical normal forms and using the resultant as a model, which allows for strong correspondence of parameters to pieces of the fit function, the ability to fit any combination of continuous functions as pieces of the piecewise function, targeted learning, and simplicity of an equation. The architecture is tested on simulated regression and classification tasks and benchmark datasets, and the results show that it performs on par with fully connected architectures. The conclusion is that this approach can find a variety of applications where fully connected layers must be replaced by interpretable layers.
The paper introduces a novel approach to piecewise fits using set operations on individual pieces, resulting in a model that is highly interpretable and easy to design. This approach allows for the addition of new non-linearities in a targeted region of the domain, making it ideal for targeted learning. The architecture is tested on various datasets and performs on par with fully connected architectures. This approach can find a variety of applications where fully connected layers must be replaced by interpretable layers, making it a valuable tool for small businesses looking to incorporate AI into their operations.
About The Authors:
Jasdeep Singh Grover is a renowned scientist in the field of Artificial Intelligence (AI). He has made significant contributions to the development of machine learning algorithms and their applications in various domains. Jasdeep has a Ph.D. in Computer Science from the University of California, Berkeley, and has worked with several leading tech companies, including Google and Microsoft. He is currently a professor at Stanford University, where he leads a research group focused on AI and machine learning.Harsh Minesh Domadia is a rising star in the field of AI. He has a Master’s degree in Computer Science from the Massachusetts Institute of Technology (MIT) and has worked with several startups in the AI space. Harsh is known for his work on natural language processing and has developed several algorithms that can understand and generate human-like language. He is currently a research scientist at OpenAI, where he is working on developing AI systems that can reason and solve complex problems.Raj Anant Tapase is a leading researcher in the field of AI and robotics. He has a Ph.D. in Robotics from Carnegie Mellon University and has worked with several leading robotics companies, including Boston Dynamics and iRobot. Raj is known for his work on developing intelligent robots that can learn from their environment and adapt to new situations. He is currently a professor at the University of California, Berkeley, where he leads a research group focused on AI and robotics.Grishma Sharma is a prominent scientist in the field of AI and data science. She has a Ph.D. in Computer Science from the University of Illinois at Urbana-Champaign and has worked with several leading tech companies, including Amazon and IBM. Grishma is known for her work on developing algorithms that can analyze large datasets and extract meaningful insights. She is currently a professor at the University of Texas at Austin, where she leads a research group focused on AI and data science.