The most useful ideas: use a small conv net with dropout & data augmentation (to reduce overfitting), simulated annealing to find hyper-parameters, and an ensemble of many nets to improve performance.
Show this thread
New notes on using ML to generalize from small amounts of training data: attacking MNIST with just 10 training examples for each digit. 93.81% accuracy: http://cognitivemedium.com/rmnist_anneal_ensemble …
-
-
-
On the addictive enjoyment (?) involved in training neural nets: http://cognitivemedium.com/rmnist_anneal_ensemble …pic.twitter.com/A6MtKR37ko
Show this thread
End of conversation
New conversation -
-
-
If you need a benchmark, this type of stuff is relatively easy to do with Gaussian mixture models. Use every example as a mean & modify the covariances to increase the accuracy.
-
I ran a half-dozen different baselines in the post linked at the top. None got much above 75%. I haven't done GMM, but I'd be surprised if they were much better.
- 2 more replies
New conversation -
-
-
This looks like an interesting side project!
-
Yup, it's fun, and a nice way to learn my way round pytorch!
- 1 more reply
New conversation -
-
-
great progress
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
There ought to be healthier ways!
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
related q: if you want to do binary classification of 0/non-0, and you only get 100 examples, would you rather have 50 0s & 50 non-0s, or 10 of each class, or some other ratio?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.