The next big step for Artificial Intelligence will require Intelligence

In a recent analysis published in the journal Neuron, Demis Hassabis, founder of DeepMind (the company behind AlphaGo purchased by Google for $650 million in 2014), and three coauthors argue that only a human intelligence better understanding will allow to push the boundaries of what artificial intelligence can achieve.

The huge ongoing progresses are essentially based on mathematics, data availability and computing power. This concerns: Deep Learning, Reinforcement Learning, Attention (Convolution Neural Network), Episodic (Deep Q-Network) and Working Memory (LSTM, GRU), Continual Learning (Multiple Timescale Learning)

But, according to them, building AI that can really perform general tasks, rather than niche ones, will need to be inspired by the way the human brain works. Concretely, their plan consists in providing AI with:

• the intuitive understanding of the physical world (i.e. an axiomatic set of basic concepts like space, number, objectness etc...)
• efficient learning (or the ability to rapidly learn about new concepts from only a handful of examples)
• transfer learning (or how to generalize knowledge gained in one context to novel, previously unseen domains)
• imagination and planning (inductive inference, select actions based on forecasts of long-term future outcomes through simulation-based planning)
• virtual brain analytics (apply the same set of tools in Neuroscience and AI systems to better understand what happens during the training process)