Robot butlers
ATLAS is a robot created by Boston Dynamics, and designed for search and rescue - but if ever the robot revolution arrives we can understand why…
25 YEARS AWAY - Do you believe that when artificial intelligence becomes smarter than us it will solve all our problems, or wipe out humanity altogether? Either way, AI is a game-changer. But currently, we still can’t create a robot that’s capable of making a cup of tea in the kitchen and then bringing it upstairs (not for want of trying…).
True, while AI is already better than us at playing Go, chess and even US TV quiz show Jeopardy!, Google’s DeepMind and IBM’s Watson are applying their machine intelligence to useful tasks like medical research and aviation safety. But will we ever have AGI – Artificial General Intelligence – that can match all the types of thinking humans do, using language in the same imprecise, contextual way, and adapting to the unpredictability of the physical world and emotional people?
One man who thinks we will is Prof Juergen Schmidhuber, head of Swiss research laboratory IDSIA. His Long Short-Term Memory (LSTM) machine-learning technique is used in Google Voice, Amazon’s Alexa and Facebook translation, and probably in your own smartphone, too.
LSTM is a development of earlier neural nets (NNs). NNs are programs that can find patterns or optimal solutions to problems without being given explicit rules. “NNs are computationally limited in many ways and insufficient for AGI,” says Schmidhuber. He’s been working on advanced deep-learning NNs for over 25 years, and developed the LSTM approach to give his AI a more human-like processing ability. Unlike previous versions, it’s able to hold relevant information until it’s needed, and to ‘forget’ less useful data. In other words, it can prioritise useful information to ‘remember’, and learn by trial and error from its mistakes. It’s had impressive results in sorting images, finding patterns and winning computer games. “LSTM relates to traditional NNs like computers relate to mere calculators,” says Schmidhuber. “It’s become the dominant general purpose deep-learning algorithm, and is now on three billion smartphones.”
Schmidhuber’s former students went on to co-found Google’s DeepMind, and to work for many other big tech companies. Now he’s started his own company, Nnaisense, and he’s hoping to achieve human-level AGI by 2050.
“Forget the dishes, leave the laundry and don’t even bother with the guttering. C-3PO is on his way”
Quantum computer
AT LEAST 20 YEARS AWAY - As the laws of physics hamper the rush for smaller, cheaper and more powerful microchips, the elusive power of the qubit (quantum bit) grows more tantalising. Caltech scientists recently announced a breakthrough in using light to store data for quantum computing, capturing individual photons in memory modules the size of a red blood cell. It’s another step towards a quantum chip, but a quantum computer fit for the mass market still looks decades away.
Time machine
NEVER. OR SURELY THEY WOULD HAVE COME BACK TO TELL US? - Breakthrough! A physicist at the University of British Columbia has calculated that it is theoretically possible to travel back in history, using the curvature of space-time. By recreating the time dilation that happens near a black hole, says Dr Ben Tippett, we could fold time into a circle. Unfortunately, to do that we’d need a new material called ‘exotic matter’ to bend space-time, and we haven’t invented that yet. Not that it’s the only problem with time travel.
Invisibility cloak
AT LEAST 10 YEARS AWAY - Invisibility is simple: it’s just a matter of redirecting light so it passes right through, or around, the object you don’t want to see. This year, a team from TU Wien achieved this by irradiating an object with a light pattern tailored to its internal structure, enabling them to guide the light through the object “as if the object was not there at all”. So it’s possible in the lab, but we’re still a long way from Harry Potter’s magical invisibility cloak.
No comments:
Post a Comment