AI is a compute-hungry beast, pushing companies to make better hardware and experiment with the existing system architectures to ease the load. In this bid, Nikhil Malhotra, the global head of Makers Lab at Tech Mahindra, coined the idea of ‘Dream AI’.
Deep Reinforced Engagement Model AI, or Dream AI, as Malhotra explains is an architecture combining symbolic AI with deep reinforcement learning, marking a shift from conventional models and addressing fundamental limitations of today’s AI. In his LinkedIn post, Malhotra pointed out three problems with current AI systems.
- 1. AI of today is just pattern recognition in a narrow domain. It’s not generalised.
- 2. Autoregressive LLMs are divergent. If they take on a state space, it’s very difficult to bring them back.
- 3. They do well to pass the Turing test but have no reasoning or context of what they say.
Enter Dream AI
Highlighting the core challenge of current systems, Malhotra explained to AIM that Dream AI builds on a neurosymbolic architecture, drawing from two foundational schools in AI—connectionist (or deep learning) and symbolic (logic and symbols).
Over time, deep learning architectures like Transformers have gained traction for their remarkable performance. However, as Malhotra explained, they do not inherently understand or reason; they excel at remembering sequences rather than forming world models.
For instance, a Transformer model might present varying responses to slight prompt changes, which can lead to hallucinations—a symptom of shallow context alignment.
This is where Dream AI steps in, creating a dual-loop system where symbolic reasoning informs world models while neural networks enable actions within those models. By incorporating symbolic AI, the framework empowers agents to simulate and act based on physical and logical rules, much like human cognition.
“Symbolism helps us build world models, and Transformers enable actions based on those models,” Malhotra said. Dream AI, thus, aspires to create agents that “dream” by simulating environments and learning with a nuanced understanding of context, instead of relying on vast datasets alone.
The Role of RL in Dream AI
Speaking at Cypher 2024, Malhotra shared that his key research goal is to make AI less compute-intensive.
He calls this the ‘min-max regret model’, which shifts away from traditional reward models and aims to empower AI to “dream” about its capabilities and understand its existence more profoundly. Malhotra said that this “dreaming model” allows AI to contemplate its own questions and aspirations.
“Can you dream about yourself? Once you develop your dream, now come back to life and start with the life that you have,” he said, highlighting the physical aspect of existence and how it relates to AI. Drawing parallels between human cognition and AI functionality, Malhotra said that just as humans subconsciously store information in the hippocampus, AI systems will use their ‘memory’ to inform decision-making processes.
Central to Dream AI is its use of reinforcement learning (RL) to harmonise symbolic reasoning with neural actions. This learning process allows AI agents to interact within simulated environments, learning optimal behaviours through feedback. For example, OpenAI decided to move away from RLHF and shift towards RL for improving the reasoning capabilities of o1.
In this hybrid architecture, reinforcement learning serves as the bridge between symbolic world models—built through simulation engines like the NVIDIA Omniverse—and the actions guided by deep neural networks. As Malhotra noted, “The AI doesn’t merely repeat learned patterns; it refines its decision-making process based on the impact of its actions within a realistic context.”
Likening it to cycling—an instinctual skill honed from childhood that remains stored in our subconscious—he said, “A lot of the data that you collect and a lot of your information still resides at the back of the hippocampus. As a result, you pull out that memory when you have to cycle.”
Efficient and Scalable AI Training
The dual learning approach makes Dream AI a powerful solution for dynamic, complex environments, enabling agents to learn contextually rather than reactively. This is particularly useful in applications where conventional AI struggles, such as real-world robotics, autonomous systems, and other domains where both physical laws and logical reasoning are essential.
By incorporating symbolic models, Dream AI aims to reduce the training burden on AI systems. Rather than relying on enormous datasets, Dream AI uses symbolic structures to “dream” or simulate scenarios, effectively minimising repetitive data input.
Symbolic AI provides a contextual foundation, accelerating learning and reducing the dependency on real-world data for every scenario. This process not only shortens the training time but also yields models that are resilient and adaptable, allowing for faster deployment and reducing costs significantly.
This is similar to what Amit Sheth, the chair and founding director of the AIISC, also told AIM. “The government is focused on AI for health, cybersecurity, and education as three of the top application areas,” Sheth said.
That’s what Sheth conveyed to the ministry and Prime Minister Narendra Modi regarding the areas of AI India should prioritise for future investment. While generative AI has caught everyone’s attention, he outlined the relevance of neurosymbolic AI, which he believes will drive the third phase of AI.
A Leap Beyond Traditional AI Architectures
Malhotra’s Dream AI architecture offers multiple advantages over traditional AI models. Its integration of symbolic reasoning with deep learning enables a level of adaptability and contextual awareness often missing in autoregressive models.
Traditional systems are confined to specific tasks, lacking a broader understanding of real-world contexts. Dream AI, however, allows agents to simulate world views, thereby aligning their actions with physical and logical principles. The dual-learning loop—symbolic reasoning paired with neural-driven action—fosters a self-reinforcing cycle of refinement and understanding.
By reducing reliance on extensive real-world data, Dream AI makes training both more efficient and scalable.
The post What is Nikhil Malhotra’s Dream AI? appeared first on Analytics India Magazine.
