https://github.com/lucidrains/esbn-pytorch

Usable implementation of Emerging Symbol Binding Network (ESBN), in Pytorch
https://github.com/lucidrains/esbn-pytorch

abstract-reasoning artificial-intelligence deep-learning memory neuro-symbolic-learning

Last synced: 3 months ago
JSON representation

Usable implementation of Emerging Symbol Binding Network (ESBN), in Pytorch

• Host: GitHub
• URL: https://github.com/lucidrains/esbn-pytorch
• Owner: lucidrains
• License: mit
• Created: 2021-01-01T22:16:08.000Z (about 4 years ago)
• Default Branch: main
• Last Pushed: 2021-01-06T00:37:20.000Z (about 4 years ago)
• Last Synced: 2024-10-15T00:16:45.085Z (3 months ago)
• Topics: abstract-reasoning, artificial-intelligence, deep-learning, memory, neuro-symbolic-learning
• Language: Python
• Homepage:
• Size: 64.5 KB
• Stars: 23
• Watchers: 5
• Forks: 2
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        


## Emerging Symbol Binding Network (ESBN) - Pytorch

Usable implementation of Emerging Symbol Binding Network (ESBN), in Pytorch. They propose to have the main recurrent neural network interact with the input image representations only through a set of memory key / values.

The input image representation are cast as memory values, and are explicitly bound to memory keys that are generated by the network. The network generates the memory keys after getting a sum of all previous memory keys weighted by the similarity of the incoming representation to the set of memory values in storage.

This decoupling / indirection of sensory to abstract processing allows the network to outperform all previous approaches, including transformers.

## Usage

```python

import torch

from esbn_pytorch import ESBN

model = ESBN(

    value_dim = 64,

    key_dim = 64,

    hidden_dim = 512,

    output_dim = 4

)

images = torch.randn(3, 2, 3, 32, 32) # (n, b, c, h, w)

model(images) # (3, 2, 4) # (n, b, o)

```

## Citations

```bibtex

@misc{webb2020emergent,

    title={Emergent Symbols through Binding in External Memory}, 

    author={Taylor W. Webb and Ishan Sinha and Jonathan D. Cohen},

    year={2020},

    eprint={2012.14601},

    archivePrefix={arXiv},

    primaryClass={cs.AI}

}

```