FIGURE 4.

Orientation decoding. (A) Retrieving encoded information. Sine-wave gratings (SWGs) with varying degrees of additive Gaussian noise were presented to the three models. The following question was then posed: how much information about the stimulus (in this case, the orientation) can be decoded from the population responses? The theoretical limit of the accuracy of such a decoder can be approximated by estimating the Fisher information (FI) in the responses. In addition, a linear decoder was also used to directly decode the population responses. This could be a downstream process which is linearly driven by the population activity, or a less-than-optimal “linear observer.” (B) Linear decoding. The responses of each model were used to train a linear-discriminant classifier. The ordinate shows the accuracy (probability of correct classification) for each level of added noise (abscissa). Results for ICA are shown in green, SC in blue, and STDP in pink. (C) Post-training threshold variation in STDP. The SWG stimuli were also used to test STDP models with different values of the threshold parameter. The threshold was either increased (by 25, 50, or 75%) or decreased (by 25, 50, or 75%) with respect to the training threshold (denoted by θ_o). The abscissa denotes the relative change in threshold, and the ordinate denotes the estimated FI. The colour of the lines denotes the input SNR, which ranged from −6 dB (blue) to 6 dB (orange).