As it stands, the current world of AI struggles with two major obstacles: generalization and explainability, both of which stem from the neural networks’ inability to conceptualize like or as good as humans. To a large extent, neural networks are excellent pattern matching algorithms. They’re good at solving tasks which use low-cognitive functions. However, where reasoning (high-cognitive functions) is concerned, AI algorithms are still out of reach.
The pain? A glass ceiling in the missions that AI can handle and black-box nature of AI, which cannot be trusted in sensitive domains since the logic of its work is not human-based. Even American military intelligence agency DARPA put forth a $2 billion campaign to tackle this problem, coined the ‘third wave of AI systems’, among projects globally, in an effort to spur research and development on neural network’s explainable qualities.
While most research is concentrated in trying to improve from lower to higher function nature, Israeli startup Decodea.ai is working in the opposite direction, starting from modeling sophisticated human reasoning.
Decodea.ai is developing algorithms that can mimic high cognitive human functions with far spanning benefits in promoting development and implementation of AI today. “It’s something we think will bring a huge leap in reinforcement learning capabilities, requiring less data to learn and less time to run,” explained Decodea.ai’s co-founder and CTO Ofer Shamai. “Moreover, what we achieved is to provide explanations of the sequential decision making in contextual terms.”
The startup was founded by Zeev Fine and Ofer Shamai, a former mathematics tutor of once religiously observant Fine (a 20 year long friendship crystallizing into a startup), originally dedicated to solving the enigma of chess between humans and machines.
“In chess, there’s a beautiful contrast between how machines and humans find the next best move” explained Fine. “When looking ahead in the tree of possibilities, you can brute force into the future moves possible to be made. But if you ask a computer why it recommends that move, it would be in the form of an arithmetical answer, like the number 1.8. It doesn’t explain anything in human terms.”
“While chess, from the beginning, was a comfortable domain for developing tools to formalize the human way of explaining, it became clear to us that Decodea’s formalizations of reasoning can take a huge step forward,” said Fine. Today, they’re building new neural network architectures that mimic human-like thinking processes. Since they will imitate reasoning and conceptualization in a much more human-like manner, the promise is to drastically improve reinforcement learning algorithms learning time, execution time and overall robustness.
Decodea’s technology also takes a unique approach among the explainable AI (XAI) algorithms. While other methods try to find traits in the input that trigger the neural network (for example, which parts of the input image triggered the neural network), Decodea.ai methods are indifferent to the way the algorithm works. “Decodea’s algorithm creates its own independent model of human thinking based on the results of the recommendation algorithm. The way this algorithm got to these results is irrelevant,” added Shamai.
In a similar vein, DeepMind company (acquired by Google) came out with AlphaZero – the impressive general playing algorithm which can learn to play various games (Chess, Go, or Shogi). However, they used enormously strong Google computers (5,000 1st-generation TPUs were used in the training process) to support their work. Another open-sourced project, Leela Chess Zero, uses the same algorithm but still after over a year, they’re struggling to reach the performance of the best handcrafted algorithms. Moreover, even though AlphaZero’s neural network is much more sophisticated than handcrafted evaluation functions, it is still statistical in nature and it needs to search millions of positions in order to play properly. “It is highly interesting to compare that to human grandmasters’ way of thinking” explained Shamai. “They calculate, at most, dozens of moves by using a flexible way of thinking which adapts the general principles to a concrete situation, in contrast to statistically based and much more elaborate neural networks. This flexibility of human thinking is what Decodea’s new algorithms are all about. We’re testing our new capabilities by developing a chess playing algorithm – like AlphaZero – but will learn much faster and require much less positions to scan before reaching its recommended move.”
According to Fine, the Decodea technology can provide immense value to many agent optimization problems, across multiple sectors, such as military and robotics applications, industry 4.0, flexible and adaptable manufacturing robotics. It’s also highly applicable to e-sports playing algorithm development, as well as autonomous driving systems with limited compute and energy resources. “From the business side we are in the process of focusing on one chosen vertical.”
The startup’s currently a participant of The Xcelerator, an incubation program partnership between TAU Ventures and Israel’s Security Agency (Shabak).
“Decodea’s approach might be ambitious but it’s also more natural,” said Shamai. “It’s going after how an expert thinks directly without having the burden to also understand the black-box statistical and non-human information. It is our compass that guides us directly to what needs to be changed in current Neural Networks.” Decodea’s approach boils down to tackling it by relying on how people think, a method with signs of promise.