Speaker

Small and efficient neural networks

The problems that we solve using neural networks can (very often) be divided into simpler sub-problems. Unfortunately, datasets used for learning on these sub-problems are usually sparse and noisy. In such cases, deep neural networks with a large number of adaptive parameters are unacceptable due to poor generalization. Alternative to this approach are the learning algorithms, which simultaneously learn both the architecture of the neural network ( the number of neurons and their connectivity), and the parameters - synaptic weights. This presentation shows the results of these algorithms in some real-world problems.