Dataset Description

I am working on famous ABIDE Autism Datasets. The dataset is very big in a sense that it has more than 1000 subjects containing half of them as autisitic and other half as healthy controls.The Dataset is taken from 17 sites across the word and each site used a varying time dimension when recording the subjects fMRI.

My Question

I want to use this dataset for classification task but only issue is time varying subjects as features set are fixed to 200 so you can say that I have subjects dimensions like 150 x200, 75 x 200 , 300 x 200... so on. So what are advanced AI or deep learning techniques that I can use to fix this time dimension for every subjects or can anybody suggest some deep learning framework or model that I could use to fix these varying time dimensions across subjects?

My Effort

Approach 1

I have applied PCA to the time dimension and fixed them to 50 and tried other numbers also but it did not produce good accuracy for classification

Approach 2

I also tried to use only specific time points from every subject like take only 40 time points from every subject to fix the dimension but again it did not work as definitely filtering some time series data on every subject would loose crucial information.


In deep learning, it is very common to use Recurrent Neural Networks (RNNs) to handle time-series data with varying input sequence lengths. Check out the RNN Wikipedia for more detail.

  • $\begingroup$ I know RNN and similarly there are many other techniques that one can use but can you give me concrete reference or link where this kind of data like I asked in my question with varying time dimension was first fixed using RNN and used for classification thereof? $\endgroup$ – 李 慕 May 20 '19 at 3:27
  • $\begingroup$ Could you be a bit more specific on what you're looking for? Are you looking for code that you could use, a tutorial on how to use RNNs, a good ML framework/tool to use, etc.? $\endgroup$ – David Rein May 20 '19 at 22:00
  • $\begingroup$ No not at all I am not looking for a code. I am just looking if someone has good architecture or model using deep learning or machine learning that I could deploy by myself. $\endgroup$ – 李 慕 May 21 '19 at 2:57

One way to tackle this problem is to identify the longest sequence in the training dataset and than expand (0 padding) all the other inputs to match that size.

  • $\begingroup$ I have done this kind of things but did not get good accuracy. I mean I am in search of some machine/deep learning model or some signal processing that could squeeze the information from all the subjects into some fixed dimension. As I have brain fMRI data and zeroing some time points will have no sense as on each time point there should be some value. $\endgroup$ – 李 慕 May 20 '19 at 12:52
  • $\begingroup$ Maybe the models you're experimenting with are not fitted for your type of problem. If you want to encode the inputs into a fixed size vector you can use any neural network architecture as an encoder (fx a 1 layer feed forward neural network will do the trick). $\endgroup$ – razvanc92 May 20 '19 at 13:27

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.