Ebbinghaus simulated: Just do it 200 times

A series of neural network simulations investigated whether the original Ebbinghaus´ German non-sense syllables would be activated and processed in their raw format, or needed association with a preceding neighboring item to be best remembered. The `predictive´ autoassociative network was unable to reduce the error rate below .50, and node activation in the Hinton diagram appeared to be random. Adding a hidden layer and recurrent loops did not improve network performance, it worsened it. In contrast, the `reproductive´ model successfully learned the syllables in 200 sweeps, with the error rate reduced to about .10. The network could also produce the syllables with zero input (rote learning), and the learning success stayed intact with input of further syllable lists (learning transfer). The Hinton diagram showed a clear co-activation for each input and output node. Hence, the chaining hypothesis could be confirmed as a sequential neural activation, but not as a sequential neural association. Much reduced learning in this network with added hidden layers showed that syllables were better learnt by heart in their raw format, without further transformation. In addition, the network was run with a graphic code encryption of input rather than an auditory code, which worked in the same way, although with attenuated activations. The Hinton diagrams suggested that the network production heuristic was `one after the next´. It is concluded that the unmediated nature of rote learning lends these kinds of memories an authentic, durable and reliable quality that makes useful raw data to be integrated into higher cognition. The difference to Ebbinghaus´ associationist account was that there appeared to be no strategy to create associations between items, but rather a clear focus on the creation of an organized script for item production.