1. Static network graph is separated from invocation context.
   a) Static graph captures layers, connections between them
      and shapes of the units of data.
   b) Invocation context specifies the batch size and stores
      all data associated with an invocation (data, gradients).
2. Batch size is now explicit in the context instead of being
   implicitly extracted by layers from incoming data.
3. Separation into Layer and ILayer is now gone, everything is
   now handled in layer implementations (with "leaf" layers focusing
   on data manipulations while container layers focusing on
   network composition).

This is still a very early prototype not intended for mergin:
1. Solver architecture not changed and just crudely hacked to support
   new network architecture.
2. Shared weights not supported.
3. Serialization not supported.

1. Static network graph is separated from invocation context.
   a) Static graph captures layers, connections between them
      and shapes of the units of data.
   b) Invocation context specifies the batch size and stores
      all data associated with an invocation (data, gradients).
2. Batch size is now explicit in the context instead of being
   implicitly extracted by layers from incoming data.
3. Separation into Layer and ILayer is now gone, everything is
   now handled in layer implementations (with "leaf" layers focusing
   on data manipulations while container layers focusing on
   network composition).
4. Solvers replaced by a more linear architecture of a top-level
   Trainer and different Optimizers (although only SGD with momentum
   is currently supported since both RMSprop and Adam require
   squaring backend support).

This is still a very early prototype not intended for mergin:
1. Shared weights not supported.
2. Serialization not supported.
3. Not all layers are migrated.

1. Static network graph is separated from invocation context.
   a) Static graph captures layers, connections between them
      and shapes of the units of data.
   b) Invocation context specifies the batch size and stores
      all data associated with an invocation (data, gradients).
2. Batch size is now explicit in the context instead of being
   implicitly extracted by layers from incoming data.
3. Separation into Layer and ILayer is now gone, everything is
   now handled in layer implementations (with "leaf" layers focusing
   on data manipulations while container layers focusing on
   network composition).
4. Solvers replaced by a more linear architecture of a top-level
   Trainer and different Optimizers (SGD with momentum and Adam
   are currently supported).

This is still a very early prototype not intended for mergin:
1. Shared weights not supported.
2. Serialization not supported.
3. Not all layers are migrated.