Visualization of activation for LSTM

2 次查看（过去 30 天）

Pierre Harouimi 2018-12-3

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/433513-visualization-of-activation-for-lstm

编辑： Stuart Whipp 2018-12-10

Like deepDreamImage for CNN layers visualization, is there any approach to check/analyze activations of a LSTM layer into MATLAB ?

My workaround for now is to change my TS to image and then check activations, but it is not perfect.

2 个评论
显示无隐藏无

Stuart Whipp 2018-12-10

myActivations.m

Hi,

I recently posted this for yielding activations from an LSTM network.

Simply copy into a .m file (or save attachment) and call as per the usual activation function (specifying layer as a number rather than by name). The basic premise is to slice the end of your network, and place a regression output just after the layer of interest. This will show activations of specified layer as they would be passed to a subsequent layer during 'normal operation

LSTM activation

This should work for all SeriesNetworks, I don't think it would work for a DAGnetwork but could amend it to work surely.

Cheers

Stuart

Stuart Whipp 2018-12-10

编辑：Stuart Whipp 2018-12-10

Incidentally, LSTM doesn't work with DAG anyway. I think this is exactly what you're looking for in terms of activations from LSTM/BiLSTM. It's not the same thing as 'deepDreamImage' of course, but the usual 'activations' function which I'm sure you've realised only works for non recurrent networks.

As I understand, DeepDream is like saying 'what change to the inputs would make the activation more pronounced upon this layer/channel'. In this manner, we can show the net a picture of a cat and ask what change to said input would make the network assume the answer was say, chicken? Pareidolia

Wikipedia actually gives a good overview. It's essentially running a layer backwards to reverse engineer a desired output. It results in trippy images where small artefacts are overprocessed to produce what look like chicken features on our original cat image.

请先登录，再进行评论。

请先登录，再回答此问题。