How to add a new category to the existing pre-trained model such that it can classify into a category more the existing pre-trained model

[click-2]

Hi ,

 I want to build a new model using the pre-trained model(let's say the pre-trained model can classify into 3 categories) ,such that it can classify  into the 4 categories.

Resources Available :

1. Pre-trained model.

2. Dataset of the new category we wanted to add.

Requirements and Constrained of the problem Statement  :

The new Model should be build using the data set belong to only one category (i.e., The new category we want to add into the existing pretrained model),no extra information on the other categories should be used.

The performance of the new model should be similar to the pre-trained .What i meant here is because you have trained the model with only one dataset the new Model will be more inclined towards the new category .If I provide an Inputs of the other category other than newly trained category the model should give almost the same results as the original pre-trained model.

Note: What I am looking for is a methodology to solve the above problem. If you feel like the above resources are not enough to solve this problem, please do add more resources you needed and let me know why you used those resources.

Thanks,

[Daπid]

On 21 March 2017 at 14:31, click-2 <2311...@gmail.com> wrote:
> The new Model should be build using the data set belong to only one category
> (i.e., The new category we want to add into the existing pretrained
> model),no extra information on the other categories should be used.

Why do you want to do that? Why the constraints?

[Tomasz Melcer]

On 21.03.2017 14:31, click-2 wrote:
> […]
> *Resources Available :*

> 1. Pre-trained model.
> 2. Dataset of the new category we wanted to add.
>

> *Requirements and Constrained of the problem Statement :*

>
> The new Model should be build using the data set belong to only one
> category (i.e., The new category we want to add into the existing
> pretrained model),no extra information on the other categories should be
> used.
>

> […]

>
> Note: What I am looking for is a methodology to solve the above problem.
> If you feel like the above resources are not enough to solve this
> problem, please do add more resources you needed and let me know why you
> used those resources.

This is a research question, not a Keras question. You will probably
have better chances at getting answers on a more research-oriented
forum. However…

My take on a problem like this would probably be to train a one-class
model (maybe a classical one, like a one-class SVM [1], maybe some kind
of density estimation with a cut-off [2], maybe some deep equivalent
[3]) on your fourth category data and use it to decide whether it's the
fourth class or one of the other classes. Please note however that
one-class classifiers are usually pretty difficult to train, and the
results usually mediocre compared to full multi-class approach.

[1]
http://rvlasveld.github.io/blog/2013/07/12/introduction-to-one-class-support-vector-machines/

[2] https://en.wikipedia.org/wiki/Density_estimation

[3]
https://www.quora.com/Do-you-have-any-implementation-of-one-class-classification-using-neural-network


--
Tomasz Melcer