Name

dnn_probe - Probe the ability of DNN activation to predict behavior.

Synopsis

dnn_probe [-h] -anal Analysis -act Activation [-layer Layer [Layer …]]
          [-chn Channel [Channel…]] [-dmask DnnMask] [-iteraxis Axis]
          -stim Stimulus -beh Behavior -model Model [-cv FolderNumber] -out Output

Arguments

Required Arguments

Argument	Discription
anal	choices=(‘uv’, ‘mv’) ‘uv’: univariate analysis ‘mv’: multivariate analysis
act	Path of a .act.h5 file which contains activation information to extract features.
stim	a .stim.csv file which contains stimulus information.
beh	Specify behaviors as the groud truth
model	Method of analysis model. Enter one parameter ‘Model’, choose from: ‘glm’: general linear model ‘lasso’: lasso regression ‘svc’: support vector machine ‘lrc’: logistic regression ‘corr’: pearson correlation
out	An output directory

Optional Arguments

Argument	Discription
layer	Name of the target layer(s). Default is all.
chn	Index of target channel(s). Default is all. Channel index starts from 1.
dmask	Path of a .dmask.csv file in which detailed information of neuron(s) of interest in DNN is specified. Argument layer/chn and dmask are mutually exclusive. Provide only one of them if needed.
iteraxis	Axis for model iteration. Default for ‘uv’ analysis is to summarize the maximal prediction score for the whole layer. Default for ‘mv’ analysis is to do analysis using all units in layer. If -anal is ‘uv’: ‘channel’: Summarize the maximal prediction score for each channel. ‘row_col’: Summarize the maximal prediction score for each location (row_idx, col_idx). If -anal is ‘mv’: ‘channel’: Do mva using all units in each channel. ‘row_col’: Do mva using all units in each location (row_idx, col_idx).
cv	Cross validation fold number.

Outputs

An output directory containing the analysis result files. For each layer, analysis result is stored as a .csv file in a subfold named after the layer.

Examples

Train a logistic regression model for each layer in Test.act.h5 to decode its representation to category labels in Test.stim.csv. The accuracy of the model is evaluated with a 10-fold cross validation, and the outputs are saved in Test_lrc_label_cv10.

dnn_probe -anal mv -act Test.act.h5 -stim Test.stim.csv -beh label -model lrc -cv 10 -out Test_lrc_label_cv10