The basic architecture of the convolutional neural network is adopted from the original work of Collobert et al.58 in NLP with some modifications (Supplemental Fig. 2G). To avoid feature shrinking, each vector of length l was first padded with zeros to equalize dimensions before feeding into the network.36, 37 For each convolutional layer, we use a convolution operation with fixed filter width, k, to produce a new feature by given function φ. Here, φ is a non-linear function that may involve a bias term. The filter is applied to the entire feature vector with all possibilities, creating a new feature map. We then apply a max pooling operation over the feature map and take the maximum value as output from this layer.58 The penultimate layer takes all features obtained from the previous layers and outputs a probability distribution over either label. The last dense layer then outputs a binary classification indicating the species. Note that we also employed a dropout on the penultimate layer to avoid the co-adaptation of units.40 We introduce a random variable from a Bernoulli distribution, with probability p being 1, which determines the proportion of dropping units.
