Tags

Constelation

Sequence model support, with jupyter notebook example, although based in
sensitive data. This release leads up to but does not include the full Fashion-MNIST example
with classification networks. This release holds many of the required
changes for both forward and backrpop with a slightly modified
decryption method as an explicit node.

Full forward bugfixes

Forward pass and certain nodes had some bugfixes when utilised together.
For instance Loss functions will now need a signal input OR y_hat + y

Complete Initial Selection of Nodes

Iteration:
- Additional neuron nodes such that a complete MNIST graph is now
  possible
- Added unittests for all new and any residual nodes that were not yet
  tested
- Added experimental custom marshmallow field to serialise and
  deserialise numpy attributes
- Expanded on MNIST example and created a new interactive pyvis graph
  for it
- Expanded documentation in several places as well as adding a whole new
  category for traversers since there will likeley be various algorithms
  toward stimulating and harvesting the neurons activations.

CNN as CC operations

CNN has been re-added but split as individual operations that can be
chained together. This marks the completion of most individual nodes.
Now we just need to work on the network traversal to handle some of our
newer but more complex cases.

Overhauled nodes and networks

Nodes now take centre stage, in a network abstraction. Inheritance and
complexity has been slashed. Now we have a very simple architecture
while still having minimal boilerplate/ redundant code.

This is completeley incompatible with 0.1.1 but this is allowed for
rapid development untill version 1.0

Dark Docs
=========

Documentation is now eye-friendly!

Original ReArray

This current version represents the nodeless implementation where FHE
was used in nodeless CNN ANN and activation functions. The FHE
implementation is based on inbuilt MS-SEAL, abstracted as rearray. The
intention is to in future split FHE backends into seperate modules/
plugins, and from now on all layers, activations etc will be nodes in
our custom graphs.