A new wave of AI is upon us. By 2025, it is expected that human-centric Cognitive AI systems with higher machine intelligence will emerge. Machines will be able to understand language, integrate commonsense knowledge and reasoning, and adapt to new circumstances. These capabilities will unlock a new set of competencies, ushering in next-generation AI applications.
Cognitive AI integrates technologies such as speech recognition, computer vision, machine learning, natural language processing (NLP), video analytics, and robotics into a single architecture to offer new levels of functionality. Intel Labs is researching broader principles of Cognitive AI to reach human-level cognitive abilities, which typically require learning from multiple modalities.
There has been significant progress on individual modalities that proved difficult just a decade ago. Pure computer vision, NLP, and recommendation systems now comprise most workloads in the data center, client, and edge products. However, the current approach relies on an ever-growing model and training on all digitized data, which cannot scale in a viable, cost-effective manner.
Increasing Scale and Complexity While Reducing Cost
Increasing the scope of AI while reducing energy consumption and cost is a major challenge. Despite growing complexity, scalability is required to make use of rapidly increasing knowledge and the coming zettabyte era. More complex, modular apps are computationally unsustainable due to current monolithic, inflexible, and data-hungry deep-learning (DL) architectures. As a result, current systems provide diminishing returns, and improvement costs are too high.
Neural networks, the primary architecture used today, were originally designed for creating functions to discover patterns and identify statistical information. To increase the scope of AI and reduce energy consumption and cost, neural networks must be combined with knowledge repositories. A categorically different AI system that goes beyond statistical correlations and a single perception modality like images or NLP is required.
Intel Labs’ Cognitive AI research is working to achieve this new level of stratification and expediency with architecture that can compactly accrue relevant knowledge and apply commonsense explainable reasoning. Currently, there are many potential approaches to solving these problems yet there is no one solution that addresses them all.
The Third Wave of AI
The next wave of of human-centric AI will have new capabilities at the algorithmic level that will drive architectural innovation. Efficient, neural-symbolic systems will combine artificial neural networks with logical inference, data retrieval, and symbolic methods so that systems can reason and learn from the environment.
Machine intelligence will be inherently multi-source, multimodal, and more robust and customizable. With these knowledge constructs, AI can continuously acquire new information and organize its view of the world. Integrating knowledge systems offers a promising approach to achieving context awareness and other capabilities outlined in DARPA’s perspective on the Third Wave of AI.
Three Levels of Knowledge
Intel Labs proposes a rethinking of knowledge construction for higher cognition called Three Levels of Knowledge. The new blueprint for the next generation of AI uses instantaneous, standby, and retrieved external knowledge to work with many data sources, data types, tasks, and update rates for continuous adaptation.
The Three Levels of Knowledge addresses the ever-growing size and complexity of deep learning (DL) models with fully encapsulated information. This type of architectuere addresses the gaps between language models and knowledge management, and can use multimodal models to retrieve and process deep knowledge to extract a reasoned response.
Multimodal Cognitive AI
Systems capable of integrating multiple modalities into their reasoning process is the next frontier of Cognitive AI. By integrating multiple modalities containing structured deep knowledge, AI can grasp meaning and relationships between entities and improve commonsense understanding.
Multimodal systems enable applications including:
- Multimodal semantic search
- Video and image search using language
- Video or image generation based on language description
- Video summarization of virtual meetings
- AI multimodal assistants
- Autonomous robots
- Drones interacting in the world using all available modalities
- Conversational AI
The move from recognition-level perception to the cognition-level ability of multimodal AI systems will require novel architectural approaches beyond what has been achieved. Intel Labs has developed a state-of-the-art temporal reasoning system and is focused on discrete reasoning, which involves counting, sorting, and connecting concepts – tasks that are hard for current neural network-based AI systems to solve.
Next Steps
Intel Labs believes that the next phase is new symbolic AI, which will include explicit, structured multimodal, multifaceted knowledge bases. A neural network is capabple of handling some of this, but it will likely be a hybrid solution that utilizes a network and machine learning to extract information and check the coherence of the overall system.
Third-wave AI technologies like Intel Labs’ Three Levels of Knowledge is expected to emerge at a similar pace to deep learning following a 10-year trajectory. Widespread implementations are expected to reach material commercial use by the end of this decade. To reach this goal, Intel Labs is collaborating with thought leaders in academia and business to develop technical specifications for a unified Cognitive AI architecture.
For more information:
- From Language to Knowledge: Intel's path to cognitive AI
- The Rise of Cognitive AI
- Video Presentation: "Cognitive AI", Gadi Singer, Intel Labs
- No One Rung to Rule Them All: Addressing Scale and Expediency in Knowledge-Based A
- WAIC 2021 presentaiton: “Cognitive AI - Architecting the Next Generation of Machine Intelligence,” Gadi Singer
- Understanding of and by Deep Knowledge
- Seat of Knowledge: AI Systems with Deeply Structured Knowledge
- Conceptualization as a Basis for Cognition — Human and Machine
- Thrill-K: A Blueprint for The Next Generation of Machine Intelligence
- Seat of Knowledge: Information-Centric Classification in AI
- KD-VLP: Improving End-to-End Vision-and-Language Pretraining with Object Knowledge Distillation