COSC 460 – Neural Networks Gregory Caza 17 August 2007
Elman (1993) Elman, J. L. (1993). Learning and development in neural networks: the importance of starting small. Cognition 48: Modelling first language acquisition using a progressive training strategy.
Elman (1993) Simple Recurrent Network (SRN) context units remember the state of the hidden units at the last time step
Elman (1993) input was a binary-encoded word words are presented one at a time output was an encoded prediction of the next word in a sentence predictions are expected to depend on the network learning a grammatical structure
Elman (1993) developmental constraints may facilitate learning limited view provides a buffer from a complex, potentially overwhelming domain simple network = child complex domain = language
Elman (1993) Training was performed using three different schemata: 1.using all training data and a fully-developed network 2.with the training data organized and presented with increasing complexity 3.beginning with a limited memory that increased throughout training
Elman (1993) developmental simulation #1: incremental input training sentences were classified as simple or complex ratio of complex : simple increased over time
Elman (1993) developmental simulation #2: incremental memory context would be reset when memory limit was reached Epoch #Memory (words) or or or or no limit
Elman (1993) full set: learning did not successfully complete incremental input: low final error; good generalization incremental memory: low final error; good generalization
Elman (1993) can training with a subset construct a “foundation for future success”? filter out “stimuli which may either be irrelevant or require prior learning to be interpreted” solution space is constrained
Elman (1993) Questions –how many sentences/epochs were used in the failed case? –what were the quantitative differences between the incremental memory/input results? –were results reproducible with different training corpora?
Assad et al. (2002) Assad, C., Harmann, M. J., Paulin, M. G. (2002). Control of a simulated arm using a novel combination of cerebellar learning mechanisms. Neurocomputing 44-46: Control of a robot arm using dynamic state estimation.
Assad et al. (2002) explore the cerebellum’s role in dynamic state estimation during movement single-link robot arm, capable of single- plane movement and releasing a ball ANN used to control the release time of the throw, with the goal of hitting a target at a certain height
Assad et al. (2002) 6 Purkinje cells (PC) 6 climbing fibres (CF) 6 ascending branches (AB) 4280 parallel fibres (PF) inhibitory; 3680 excitatory
Assad et al. (2002) each excitatory PF received a radial basis function (RBF) of 2 state variables PF-PC connections were strengthened through ‘Hebbian-like’ learning after each trial, a binary error signal was generated based on throw accuracy if the ball hit the target window, PF-PC connections were strengthened through ‘Hebbian-like’ learning
Assad et al. (2002) the target window was initialized to be “quite large” if a hit was recorded, the window was shrunk if there was an error, the window was expanded
Assad et al. (2002) physiological experiments demonstrate LTD between PF and CF most cerebellar models ignore the AB input the network suggests a possible role for LTP in cerebellar learning through the AB
Assad et al. (2002) details, details! too complicated => laying groundwork for experiments Why does no learning take place when the target is missed? What about negative reinforcement?