Description of "training_batch"
Opened this issue · 4 comments
In the documentation for scheme2, we see the following:
max_num_snippets_per_training_batch: The maximum number of snippets to use for training the classifier in each batch. See below for more details on what constitutes a batch.
There isn't another use of the word "batch" in the documentation, so I'm trying to understand exactly what is meant. My understanding was that scheme_2 only has a single training batch up-front before subsequent classification. Can you provide some insight into what this parameter is doing?
My understanding from reading the code is that this is the number of snippets per channel (or neighborhood) to be used during the training of the classifier. Is this right?
If so, for a channel with one high-firing rate neuron and one low-firing rate neuron, wouldn't picking the number of snippets in this way result in the classifier being exposed to more of the high-firing rate neuron?
My understanding from reading the code is that this is the number of snippets per channel (or neighborhood) to be used during the training of the classifier. Is this right?
That sounds right. But I'd need to look carefully at the code to confirm... it gets confusing and this is not as fresh in my mind. :)
If so, for a channel with one high-firing rate neuron and one low-firing rate neuron, wouldn't picking the number of snippets in this way result in the classifier being exposed to more of the high-firing rate neuron?
Yes I think you're right. Maybe there's a better way that's isn't as biased.
OK, don't worry about it. I'll figure out from the source-code.
I think there's a straightforward way to deal with the high firing-rate bias. I'll make an issue if it shows any promise.
Sounds good, thanks.