Deep Learning Limitations

Will computers ever get smarter than us humans? Believe it or not, this question is a hot topic again after being dismissed as impossible in the past. One of the early pioneers of deep learning technologies has been hard at work and there have indeed been some incredible breakthroughs recently.

Judea Pearl, winner of the 2011 Turing Award, is known as one of the fathers of Bayesian networks. He has always regarded Bayesian networks as having tremendous possibilities for AI. But now he says that, "All the impressive achievements of deep learning amount to just curve fitting."

The achievements of deep learning, so far, have failed to live up to their promise. Today's intelligent machines can only produce statistical correlations. They do not understand causality. To deliver true intelligence, the machines need to move beyond just finding correlations. They need to understand causality too. Only then can we say we have real AI. Pearl's book, "The Book of Why", shows that Judea Pearl has been committed to changing the direction of artificial intelligence research for many years, that it has been his sincere desire to bring talk of causality back into the conversation.

In 2017, Judea Pearl pointed out that Neural Information Processing Systems (Neural IPS) machine learning were only running statistical or blind models, which have certain theoretical limitations. But the development of graph networks is expected to break through these limitations, give machine learning new engines, and give us intelligent machines, machines that can actually understand causality.

Graph Networks

All of these new developments have been part of a tremendous response to Google recently opening their DeepMind Graph Nets (GN) library to the public. This new "graph network" has demonstrated a structure that can be applied generally and with tremendous expressiveness. This gives its enormous advantages in terms of relational reasoning and generalized permutations. A graph network has significant advantages when it comes to causal reasoning. The graph model represents a way of expressing actual knowledge, and as such, it brings us one step closer to automatic reasoning. Its structure includes graph models, structural equations, counterfactuals, and interventional logic. Counterfactuals help clarify problems, and structural equations can represent clear semantic relationships, which take model-driven reasoning in a more promising direction. This new deep learning model is closer to how the human brain works.

AI and the Missing Piece of the Puzzle That Is Our Human Brain

With deep learning, only the weights can be adjusted, but on a graph network learning occurs on each node. On a graph network, feedforward control is used at every step, at the edge, and at every node. Information can be modified at any point in the process. This architecture makes is possible for the entire network to learn.

The network depends on the underlying graph database and graph-computing platform to store queries and perform analysis. And the HUAWEI CLOUD Graph Engine Service (GES) is one of the best of these types of platforms out there. This next-generation, high-performance graph computing platform, with its superb computing performance, large-scale analysis capabilities, and high compatibility is exactly what is needed for enterprise-level intelligent analysis.

The graph engine performs association and relationship analysis on a massive sea of data, organizing the results into a "knowledge graph", and this knowledge graph can depict relationships between real things. It constructs a relational network, which, in turn, lays the foundation for causal reasoning. User behavior in real world scenarios can be reproduced and analyzed by mining and analyzing relational networks. And this leads to real world applications like social or purchasing recommendations or the analysis of financial fraud.

Semantic search, one of the hottest technologies being developed today, is one of the most natural applications for artificial intelligence. Understanding semantics has always been one of the main goals for this technology, and the development of the graph network is the last piece of the puzzle, a piece that gives AI the ability to finally really understand the world.

With a knowledge map supporting it, a graph engine can drive a search engine that can analyze semantic relationships between images and text, and produce search results that take the semantic context into account. A graph engine can treat search as a conversation.

In 2018, HUAWEI CLOUD Graph Engine Service become the official image computing platform of Apache TinkerPop. The self-developed kernel EYWA has earned widespread recognition from customers in many different fields. It is used in applications providing shopping or social recommendations, project analysis, enterprise insights, or knowledge mapping. It has found applications in financial risk management, enterprise IT, and relationship mining. In the field of artificial intelligence, the image engine can be found at the core of multiple AI technologies like Internet of Vehicles (IoV), intelligent customer service, and deep learning.

HUAWEI CLOUD Graph Engine Service earned the "2018 Leading Science and Technology Achievement Award" for achievements in "New Technologies" and "Black Technology".