Editable Neural Networks

ICLR 2020 8-22-2020

Motivation

it is crucially important to correct model mistakes quickly as they appear during training the neural network. In this work, we investigate the problem of neural network editing — how one can efficiently patch a mistake of the model on a particular sample, without influencing the model behavior on other samples. Namely, we propose Editable Training, a model-agnostic training technique that encourages fast editing of the trained model.

Editing Neural Networks

• $f(x,\theta)$ : a neural network

• $\mathcal{L}_{base}(\theta)$ : task-specific objective loss function

The goal is to change model's predictions on a subset of inputs, corresponding to misclassified objects, without affecting other inputs, by changing the model parameters $\theta$ .

An editor function could be used to formalized this: $\hat{\theta}=Edit(\theta,l_e)$ , with a constraint: $l_e(\hat{\theta})\leq0$

• $\hat{\theta}$ is the changed parameters

For example, multi-class classification.

• $l_e(\hat{\theta}) = \max_{y_i} (\log p(y_i|x, \hat{\theta}) - \log p(y_{ref}|x,\hat{\theta}))$ where $y_{ref}$ is the desired label.

if under the constraint: $l_e(\hat{\theta})\leq0$ ,

The constraint: $l_e(\hat{\theta})\leq0$ is satisfied iff $\arg \max_{y_i} \log p(y_i|x, \hat{\theta})=y_{ref}$

So the goal is how to design the editor neural network.

Reference:

• https://openreview.net/forum?id=HJedXaEtvS

• https://github.com/editable-ICLR2020/editable

• https://github.com/xtinkt/editable