Authors:

Lorenzo Cacciapuoti

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

The most widespread artificial intelligence (AI) paradigm is the deep learning neural network. This type of AI arranges its neurons in consecutive, feed-forward layers and learns using the backpropagation algorithm. While proven effective, this type of AI has some significant drawbacks, notably the large network size required, its low data efficiency, and its tendency for catastrophic forgetting. To solve these issues, we propose the Interconnected Dendritic Network (IDN), a new type of neural network that takes close inspiration from biological neural networks of cortical pyramidal neurons. In an IDN, neurons are arranged in a multidimensional space, forming homogeneous connections across layers, allowing for recurrent information processing. Furthermore, neurons in an IDN receive their inputs through dendrites (hierarchically divided into branches of increasing order) that activate non-linearly while also considering inhibitory (depressing) inputs. This grants individual neurons greater pattern recognition abilities than that of current point-source models utilized in traditional neural networks. IDNs learn using a Hebbian learning rule regulating the variation of connection weights as well as the formation and destruction of connections. To quantify the efficacy of learning, the performance of IDNs on an image classification task (using the MNIST dataset) will be compared to that of traditional neural networks to determine the effectiveness of this type of model. The model will also be tested on its ability to retain information about multiple tasks at once to assess whether it also falls victim to catastrophic forgetting. These tests will serve to quantify the viability of IDNs for AI applications.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Lorenzo Cacciapuoti