Grounding Transfer
My Work
Version 1.
- Experiment consist of 4 net Label1Net, Label2net, Label3Net and Instance Net.
- Label1Net is fully connected with Label2Net which is fully conected with Label2Net and InstanceNet.
- Connections between Label1net and Label2net and between InstanceNet and Label2net are plastic and are learned.
- Label1Net has 8 CAs (pyramid+vertical, vertical, horizontal, red, green, moveleftleg, move rightleg and moveleftleg + moverightleg).
- These CAs get committed to corresponding CAs of Label2Net through compensatory learning.
- Learning rate is low (0.0001) for Label2Net and for (0.0001) for Label2Net.
- Learning run for 15 epochs where each epoch consists of 2000 CantStep. Where each CA gets 250 cycles.
- System is tested for further 5 epochs.
- And than Lables are learned by compensatory learning from Instancenet (learning rate of 0.2) to Label2net for 3 epochs and than tested for 2 epochs.
- If in Label2Net more than 1 CA is firing at the same time they get committed to Label3Net to form super Label CA.