An integrated, ultrafast, optical neuromorphic processor that:
- Realises an arbitrary Turing machine that can represent any arbitrary defined function of a multidimensional space of coordinates varying with time.
- It has the same expressive power of state of the art fully trained deep neural networks.
- As in a reservoir computing platforms, learns the required function by leaving a large part of the network in the initial state and untrained.
- Learns a required function in a single step by a suitable technique that is mathematically stable and has higher predictability than any current available platform.
- Contains 100M equivalent neurons, with all-to-all connections, in extremely compact sizes.
- Dissipates ultra-low powers.
US patent App. No. 62/963,747. No further disclosure of information is possible at this stage.
Recent Comments