https://github.com/jiamings/fast-weights

Implementation of the paper [Using Fast Weights to Attend to the Recent Past](https://arxiv.org/abs/1610.06258)
https://github.com/jiamings/fast-weights

accuracy deep-learning fast-weights tensorflow tensorflow-models

Last synced: about 1 month ago
JSON representation

Implementation of the paper [Using Fast Weights to Attend to the Recent Past](https://arxiv.org/abs/1610.06258)

• Host: GitHub
• URL: https://github.com/jiamings/fast-weights
• Owner: jiamings
• Created: 2016-10-29T06:19:37.000Z (almost 8 years ago)
• Default Branch: master
• Last Pushed: 2016-11-03T15:49:22.000Z (almost 8 years ago)
• Last Synced: 2024-05-03T06:21:17.472Z (5 months ago)
• Topics: accuracy, deep-learning, fast-weights, tensorflow, tensorflow-models
• Language: Python
• Size: 227 KB
• Stars: 171
• Watchers: 7
• Forks: 23
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
## Using Fast Weights to Attend to the Recent Past

Reproducing the associative model experiment on the paper

[Using Fast Weights to Attend to the Recent Past](https://arxiv.org/abs/1610.06258) by Jimmy Ba et al. (Incomplete)

### Prerequisites

Tensorflow (version >= 0.8)

### How to Run the Experiments

Generate a dataset

```

$ python generator.py

```

This script generates a file called `associative-retrieval.pkl`, which can be used for training.

Run the model

```

$ python fw.py

```

### Findings

The following is the accuracy and loss graph for R=20. **The experiments are barely tuned.**

![](fig/acc.png)

![](fig/loss.png)

**Layer Normalization is extremely crucial for the success of training.** 

- Otherwise, training will not converge when the inner step is larger than 1. 

- Even when inner step of 1, the performance without layer normalization is much worse. For R=20, only 0.4 accuracy can be achieved (which is same as the level of other models.)

- Even with Layer Normalization, using slow weights (ie. vanilla RNN) is much worse than using fast weights.

Further improvements:

- Complete fine-tuning

- Work on other tasks

### References

[Using Fast Weights to Attend to the Recent Past](https://arxiv.org/abs/1610.06258). Jimmy Ba,  Geoffrey Hinton, Volodymyr Mnih, Joel Z. Leibo, Catalin Ionescu.

[Layer Normalization](https://arxiv.org/abs/1607.06450). Jimmy Ba, Ryan Kiros, Geoffery Hinton.