Jurgen Schmidhuber’s research while affiliated with Dalle Molle Institute for Artificial Intelligence Research and other places

Audiovisual speech recognition (AVSR) is a method to alleviate the adverse effect of noise in the acoustic signal. Leveraging recent developments in deep neural network-based speech recognition, we present an AVSR neural network architecture which is trained end-to-end, without the need to separately model the process of decision fusion as in conventional (e.g. HMM-based) systems. The fusion system outperforms single-modality recognition under all noise conditions. Investigation of the saliency of the input features shows that the neural network automatically adapts to different noise levels in the acoustic signal.

Slime mould of Physarum polycephalum is a large cell exhibiting rich spatial non-linear electrical characteristics. We exploit the electrical properties of the slime mould to implement logic gates using a flexible hardware platform designed for investigating the electrical properties of a substrate (Mecobo). We apply arbitrary electrical signals to ‘configure’ the slime mould, i.e. change shape of its body and, measure the slime mould’s electrical response. We show that it is possible to find configurations that allow the Physarum to act as any 2-input Boolean gate. The occurrence frequency of the gates discovered in the slime was analysed and compared to complexity hierarchies of logical gates obtained in other unconventional materials. The search for gates was performed by both sweeping across configurations in the real material as well as training a neural network-based model and searching the gates therein using gradient descent.

Slime mould of Physarum polycephalum is a large cell exhibiting rich spatial non-linear electrical characteristics. We exploit the electrical properties of the slime mould to implement logic gates using a flexible hardware platform designed for investigating the electrical properties of a substrate (MECOBO). We apply arbitrary electrical signals to `configure' the slime mould, i.e. change shape of its body and, measure the slime mould's electrical response. We show that it is possible to find configurations that allow the Physarum to act as any 2-input Boolean gate. The occurrence frequency of the gates discovered in the slime was analysed and compared to complexity hierarchies of logical gates obtained in other unconventional materials. The search for gates was performed by both sweeping across configurations in the real material as well as training a neural network-based model and searching the gates therein using gradient descent.

Neural dynamics offer a theoretical and computational framework, in which cognitive architectures may be developed, which are suitable both to model psychophysics of human behaviour and to control robotic behaviour. Recently, we have introduced reinforcement learning in this framework, which allows an agent to learn goal-directed sequences of behaviours based on a reward signal, perceived at the end of a sequence. Although stability of the dynamic neural fields and behavioural organisation allowed to demonstrate autonomous learning in the robotic system, learning of longer sequences was taking prohibitedly long time. Here, we combine the neural dynamic reinforcement learning with shaping, which consists in providing intermediate rewards and accelerates learning.We have implemented the new learning algorithm on a simulated Kuka YouBot robot and evaluated robustness and efficacy of learning in a pick-and-place task.

We consider the problem of detecting mitotic figures in breast cancer histology slides. We investigate whether the performance of state-of-the-art detection algorithms is comparable to the performance of humans, when they are compared under fair conditions: our test subjects were not previously exposed to the task, and were required to learn their own classification criteria solely by studying the same training set available to algorithms. We designed and implemented a standardized web-based test based on the publicly-available MITOS dataset, and compared results with the performance of the 6 top-scoring algorithms in the ICPR 2012 Mitosis Detection Contest. The problem is presented as a classification task on a balanced dataset. 45 different test subjects produced a total of 3009 classifications. The best individual (accuracy = 0.859 ± 0.012), is outperformed by the most accurate algorithm (accuracy = 0.873 ± 0.004). This suggests that state-of-the-art detection algorithms are likely limited by the size of the training set, rather than by lack of generalization ability.