Eliminate features that are highly correlated to the output labels.

Combine highly co-dependent features into one representative feature.

Instead of feeding in each feature individually, average their values in batches of 3.

Remove the features that have null values for more than 50% of the training records.