Would love to hear what interesting use cases people plan to train their convolutional networks on to test their performance. Since they’ve gotten so good to do, what was it, ‘fine grain inter-class classification’ I thought we could give it some real challenging tasks.
I thought id train mine to tell the difference between my math and statistics subject notebooks. This would probaby give it a tuff time as the the symbols are quite similar, more than in between languages.
Share your ideas!
Great idea !
Did you think of a threshold for the number of images per class?
Honestly I’m fascinated by the unconventional images idea from right at the end (i missed the name but it was for fraud detection) and now i’m madly looking around whatever non image datasets I can find to try to figure out if i can turn them into pseudo-images…
Maybe 50 each for starters
Our neighbours cat comes and eats the food our our two cats if we have it on the porch. It would be a nice idea to have an enclosed area that only our cats could enter, with a raspberry pi controlled door.
May be, adversarial attack on images might be of an interest to you.
https://medium.com/onfido-tech/adversarial-attacks-and-defences-for-convolutional-neural-networks-66915ece52e7
Also, can you give an example of getting pseudo-images from non image datasets. That sounds interesting !
found it!
2:43:00 of the lecture, Fast.Ai student working at Splunk created anti-fraud software using image analysis of mouse movements. Damn!
Taking multiples of two as the batch size is recommended. It would be neat experiment to plot the change in accuracy vs decreasing no of input images. It might give you an idea of a threshold limit for different architectures (resnet34, resnet50) for your particular dataset.
That Splunk thing is really, really out-of-the-box-ly creative.