Academia.eduAcademia.edu

Figure 3 – uploaded by Prof Eduardo R. Miranda

See full PDFdownloadDownload figure
Fig. 3. Activations from the output layer on in two different views As in step 5 of the algorithm, all the active units are decayed proportionally to the decay parameter d,

Figure 3 Activations from the output layer on in two different views As in step 5 of the algorithm, all the active units are decayed proportionally to the decay parameter d,

Related Figures (7)

Fig. 1. Kohonen Self Organising Feature Map (SOM) In Fig. 1 we can see a diagram of a SOM with 16 units and one input. The dimension of the input vector determines the dimension of the weights vector of each unit. To determine which weight vector is the closest one to the input unit, the euclidean distance is calculated:
Fig. 1. Kohonen Self Organising Feature Map (SOM) In Fig. 1 we can see a diagram of a SOM with 16 units and one input. The dimension of the input vector determines the dimension of the weights vector of each unit. To determine which weight vector is the closest one to the input unit, the euclidean distance is calculated:
Fig. 2. SARDNET - Self-organising activation, retention and decay network
Fig. 2. SARDNET - Self-organising activation, retention and decay network
Eq. 4 is the propagation function of the Perceptron and g(h) in Eq. 5 is the activation function computed by the units. In this case this function is a sigmoidal function, The Perceptron is a neuron-like learning network developed by Rosenblatt [23] which is a one layer feed-forward neural network with a set of inputs that are fully connected to an output layer. The outputs of Perceptrons are explicit functions of the inputs. Fig. 5 shows its architecture.
Eq. 4 is the propagation function of the Perceptron and g(h) in Eq. 5 is the activation function computed by the units. In this case this function is a sigmoidal function, The Perceptron is a neuron-like learning network developed by Rosenblatt [23] which is a one layer feed-forward neural network with a set of inputs that are fully connected to an output layer. The outputs of Perceptrons are explicit functions of the inputs. Fig. 5 shows its architecture.
Fig. 5. Interaction diagram for the imitation game proposed n The number of inputs of the Perceptron is the number of units of the SARD- NET. The number of output neurons is arbitrarily defined as being 3 to be able to visualise the results in a tridimensional grid. This output grid enables the categorisation of the input sequences.
Fig. 5. Interaction diagram for the imitation game proposed n The number of inputs of the Perceptron is the number of units of the SARD- NET. The number of output neurons is arbitrarily defined as being 3 to be able to visualise the results in a tridimensional grid. This output grid enables the categorisation of the input sequences.
Fig. 6. Sarnet weight evolution without change of neighbourhood: a) Weight initiali- sation; b) After 20 iterations; c)After 80 iterations; d) Difference between the weights’ sum in consecutive iterations The graphs from Fig. 6 show the evolution of the third component of the weights (Inter-onset Intervals). The first graph shows the initial value of the weights, as explained above, the second shows the organisation process after 20 iterations, and the third shows the weights stabilised after 80 iterations. Fig. 6 d) shows the difference between the sums of the weights in two consecutive iterations, this being a measure of the stabilisation of the weights.
Fig. 6. Sarnet weight evolution without change of neighbourhood: a) Weight initiali- sation; b) After 20 iterations; c)After 80 iterations; d) Difference between the weights’ sum in consecutive iterations The graphs from Fig. 6 show the evolution of the third component of the weights (Inter-onset Intervals). The first graph shows the initial value of the weights, as explained above, the second shows the organisation process after 20 iterations, and the third shows the weights stabilised after 80 iterations. Fig. 6 d) shows the difference between the sums of the weights in two consecutive iterations, this being a measure of the stabilisation of the weights.
Fig. 7. Sarnet weight evolution with change of neighbourhood: a) Weight initialisation; b) After 20 iterations; c)After 80 iterations; d) Difference between the weights’ sum in consecutive iterations
Fig. 7. Sarnet weight evolution with change of neighbourhood: a) Weight initialisation; b) After 20 iterations; c)After 80 iterations; d) Difference between the weights’ sum in consecutive iterations
Fig. 8. a) Perceptron error in learning process; b) Categorisation space
Fig. 8. a) Perceptron error in learning process; b) Categorisation space
Related topics:
Computer ScienceArtificial IntelligenceArchitectureComputer MusicComputer AnimationSound and Music ComputingArtificial LifeLanguage EvolutionComputational MusicologyMulti Agent System
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved