Final Weights

The final weights are generated as a sequentially stored values in binary format, in layer order. Within each layer, the final weights are stored in order of node index starting from the 0th to the N-1 for N nodes if the layer. The value immediately following the N-1th node is the weight assigned to the bias element.

The model specifications allow you to set some node weights to fixed values, such as 0.0 or 0.5. These sentinel values inform us that those node weights do not participate in the error reduction process. This allows a far more sophisticated design of the NN model. The final weights retains these values at the appropriate offset within the weights file. This design is expected to help you to run the model on subsequent data inputs smoothly.

The final weights file however, is available only as an encrypted file, with GCM codebook encryption using the key that was used to provide us the input file. The key must be 128, or 192 or 256 bits long, and the file is encrypted using AES methodology. No other form is considered safe enough for transfer of the weights.

The weights file is posted back to the cloud storage bucket where the source data was provided once the run is complete. Please be aware that we cannot maintain copies of the weights data at our end since we cannot risk keeping the encryption key value. Without the key value and the model specifications file, the weights data is anonymized as a binary file with unknown contents.