Figure 8. Correspondence between the temporal prediction model’s ability to predict future auditory input and the similarity of its units’ responses to those of real A1 neurons.
Performance of model as a function of number of hidden units and L1 regularization strength on the weights as measured by (a), prediction error (mean squared error) on the validation set at the end of training and (b), similarity between model units and real A1 neurons. The similarity between the real and model units is measured by averaging the Kolmogorov-Smirnov distance between each of the real and model distributions for the span of temporal and frequency tuning of the excitatory and inhibitory RF subfields (e.g. the distributions in Figure 6d–e and Figure 6g–h). Figure 8—figure supplement 1 shows the same analysis, performed for the sparse coding model, which does not produce a similar correspondence.
Figure 8—figure supplement 1. Correspondence between sparse coding model’s ability to reproduce its input and the similarity of its units’ responses to those of real A1 neurons.
Performance of model as a function of number of units and the L1 regularization strength on the activities as measured by (a), reconstruction error (mean squared error) on the validation set at the end of training and (b), similarity between model units and real A1 neurons. The similarity between the real and model units is measured by averaging the Kolmogorov-Smirnov distance between each of the real and model distributions for the span of temporal and frequency tuning of the excitatory and inhibitory RF subfields (e.g. the distributions in Figure 3d–e and Figure 3g–h).
Figure 8—figure supplement 2. Interactive figure exploring the relationship between the strength of L1 regularization on the network weights and the structure of the RFs the network produces when the network is trained on auditory inputs.
The interactive version of this figure can be found at https://yossing.github.io/temporal_prediction_model/figures/interactive_supplementary_figures.html. The left hand panel shows the performance of the network with the hyperparameter settings specified on the x and y axes. The x axis signifies the strength of L1 regularization placed on the weights of the network during training. The y axis signifies the number of hidden units in the network. The colour represents the predictive capacity of the model as measured by the prediction error (mean squared error) on the validation set at the end of training.
How to interact with the figure: Hover over a point in the left hand panel to show the corresponding spectrotemporal receptive fields of the network in the right hand panel. Using the settings near the right hand panel, zoom, pan and reset the image to explore the shapes of the spectrotemporal receptive fields. Many hidden units’ weight matrices decayed to near zero during training. Inactive units were excluded from analysis and are not shown.
Figure 8—figure supplement 3. Interactive figure exploring the relationship between the strength of L1 regularization on the network weights and the structure of the RFs the network produces when the network is trained on visual inputs.
The interactive version of this figure can be found at: https://yossing.github.io/temporal_prediction_model/figures/interactive_supplementary_figures.html. The left hand panel shows the performance of the network with the hyperparameter settings specified on the x and y axes. The x axis signifies the strength of L1 regularization placed on the weights of the network during training. The y axis signifies the number of hidden units in the network. The colour represents the predictive capacity of the model as measured by the prediction error (mean squared error) on the validation set at the end of training.
How to interact with the figure: Hover over a point in the left panel to show the corresponding spatial receptive fields of the network in the right panel. Using the settings on the right of the right hand panel, zoom, pan and reset the image to explore the shapes of the spatial receptive fields. Change the slider labelled 'time step' to change the time-step of the spatial receptive fields shown in the right hand panel. Some hidden units’ weight matrices decayed to near zero during training. Inactive units were excluded from analysis and are not shown.