White House published the information about AI which requests mentions about 'the most important research gaps in AI that must be addressed to advance this field and benefit the public'.
What are these exactly?
Artificial Intelligence Stack Exchange is a question and answer site for people interested in conceptual questions about life and challenges in a world where "cognitive" functions can be mimicked in purely digital environment. It only takes a minute to sign up.Sign up to join this community
One way of illustrating the deficiencies of many of our current approaches at once is to consider how well it is possible to represent (equivalently, learn) commonsense knowledge.
In this area, the Winograd Schema Challenge has been proposed by Levesque, in which each problem is given as input natural language text containing an ambiguous pronoun:
Babar wonders how he can get new clothing. Luckily, a very rich old man who has always been fond of little elephants understands right away that he is longing for a fine suit.
Here, the program is asked to decide if 'he' in "he is longing for a fine suit" refers to Babar or the old man. Several thousand such questions have been collated and proposed as a more quantifiable alternative to the Turing test.
Despite the fact that the input domain is natural language, success here is undeniably a pre-requisite for AGI and (as implied in my answer here) for being able to interact ethically with the human world.
According to IBM Research organization in the response to White House as part of preparing for the future of Artificial Intelligence, AI depends upon many long-term advances, not only from AI researchers, but from many interdisciplinary teams of experts from many disciplines, including the following challenges:
Machine learning and reasoning.
Currently AI systems use supervised learning using huge amount of dataset of labeled data for training. This is very different to how humans learn by creating concepts, relationship, common sense reasoning which gives ability to learn much without too much data. Therefore machine learning with common-sense reasoning capabilities should be researched further more.
Current AI-based systems have very limited ability for making decisions, therefore new techniques must be developed (e.g. modeling systemic risks, analyzing tradeoffs, detecting anomalies in context, analyzing data while preserving privacy).
Domain-specific AI systems.
The current AI-based system is lack of abilities to understand the variety of domains of human expertise (such as medicine, engineering, law and many more). The systems should be able to perform professional-level tasks such as designing problems, experiments, managing contradictions, negotiating, etc.
Data assurance and trust.
The current AI-based systems require huge amounts of data and their behaviour directly depends on the quality of this data which can be biased, incomplete or compromised. This can be expensive and time consuming especially where it is used for safety critical systems which potentially can be very dangerous.
Radically efficient computing infrastructure.
The current AI-based systems require unprecedented workloads and computing power which require development of new computing architectures (such as neuromorphic).
Interpretability and explanations.
For people to follow AI suggestions, they need to trust systems, and this is only when they are capable of knowing users' intents, priorities, reasoning and they can learn from their mistakes. These capabilities are required in many business domains and professionals
Value alignment and ethics.
Humans can share the common knowledge of how the world function, the machine cannot. They can fail by having unintended and unexpected behaviour only because humans did not specify the right goals for them or them omitted essential training details. The systems should be able to correct specification of the goals and avoid unintended and undesired consequences in the behaviour.
The AI-based systems should be able to work closely to humans in their professional and personal life, therefore they should have significant social capabilities, because they can impact on our emotions and our decision making capabilities. Also sophisticated natural language capabilities will need to be developed to allow a natural interaction and dialog between humans and machines.