Really small example (multi-class) mixture of experts model, almost. Technically, belief_per_model function needs to assign probabilities based on a function.

expert_mixtures.py

from keras.models import Model
from keras.layers import Input, Lambda, Dense
from keras.utils import to_categorical
from numpy.random import randint
import numpy as np


def belief_per_model(x):
    x1, x2, x3, x4 = x
    return x1 * .2 + x2 * .3 + x3 * .4 + x4 * .1


def my_model():
    inputs = Input(shape=(20,))

    m1 = Dense(64)(inputs)
    m2 = Dense(64)(inputs)
    m3 = Dense(64)(inputs)
    m4 = Dense(64)(inputs)

    mixture = Lambda(
        belief_per_model, output_shape=(64,)
    )([m1, m2, m3, m4])

    output = Dense(
        10, activation='softmax'
    )(mixture)

    model = Model([inputs], output)

    model.compile(
        loss='categorical_crossentropy',
        optimizer="sgd"
    )
    return model


def test_data():
    x = np.random.random((1000, 20))
    y = to_categorical(
        randint(10, size=(1000, 1)),
        num_classes=10
    )
    return y, x


if __name__ == '__main__':
    y_train, x_train = test_data()
    model = my_model()
    model.fit([x_train], y_train)