A case for Context Awareness in AI
Does applied AI have the necessary insights to tackle even the slightest (unlearned or unseen) change in context of the world surrounding it?
In discussions AI often equals deep-learning models. Current deep learning methods heavily depend on the presumption of Independent and Identically Distributed data (I.I.D) to learn from, something which has serious implications for the robustness and transferability of models. Despite very good results on classification tasks, regression and pattern encoding, current deep learning methods are failing to tackle the hard open problem of generalization and abstraction across problems. Both are prerequisites for general learning and explanation capabilities.
There is great optimism that deep learning algorithms, as a specific type of neural network, will be able to close in on “real AI” when it only is further developed and scaled up enough (ref. Bengio 2018[1]). Others feel that current AI approaches are merely a smart encoding of a general distribution into a deep learning networks´ parameters (“ein geschicktes zählen”[2]) and regard it as insufficient to act independently within the real world. So where are the real intelligent behaviours as in the ability to recognize problems and plan for solving them, understanding of the physics and logics, causality and analogy. All in order to understand, predict, influence and act on it?
AI is not ML: learning patterns with deep-learning is not enough to understand and act in the real world.
What would be needed is a better understanding by machines of their context, as in the surrounding world and it´s inner workings. Only than machines can capture, interpret and act upon previously unseen situations. This will require:
- Understanding of logical constructs as causality (as opposed to correlation). If it rains, you put on a raincoat. But putting on a raincoat does not stop the rain. Current ML struggles learning causality. Being able to represent and model causality will to a large extend facilitate better explanations and understanding of decisions made by ML models.
- Ability to tackle counterfactuals: “if a crane has no counterweight, it would topple over”
- Transferability of learned “knowledge” across/between domains (current transfer learning only works on small tasks with large domain overlap between them = similar tasks in similar domains)
- Withstand adversal attacks. Only small random changes made in the input data (deliberately or not) can make results of connectionist models highly unreliable. Abstraction mechanisms might be a solution to this issue.
- Reasoning on possible outcomes, finding problematic outcomes and a) plan for avoiding them while reaching the goal or b) if that is not possible find alternative goals and try to reach those. -> Problem Solving capabilities.
In DPIR wave 1[3] we have made the case for extending the context in which AI models are operating, using a specific type of models which can benefit from domain knowledge in the form of knowledge graphs. From the above it follows that knowledge alone probably will not be enough. Higher level abstraction and reasoning capabilities are also needed. Current approaches aim at combining connectionist approaches with logical theory:
Joshua Tenenbaum, professor of computational cognitive science at MIT, posits that we must combine the achievements from symbolic AI, probabilistic and causal models, and neural networks to solve the challenges of deep learning.
- Some connectionists feel that abstraction capability will follow automatically from scaling up networks, adding computing power and use more data. But it seems that deep learning models cannot abstract or generalize more than learning general distributions. The output will at the most be a better encoding but still not deliver symbolic abstraction, causality or showing reasoning capabilities.
- Symbolic AI advocates concepts as abstracted symbols, logic and reasoning. Symbolic methods allow for learning and understanding human made social constructs like law, jurisprudence, country, state, religion, culture. Could connectionist methods be “symbolized” as to provide capabilities as mentioned above?
- Several innovative directions can be found in trying to merge methods into hybrid approaches consisting of multiple layers or capabilities:
3a. Intuition layer: let deep learning algorithms take care of the “low level” modeling of intuition or tacit skills shown by people having performed tasks over a long time (like a good welder who can hardly explain how she makes the perfect weld after years of experience).
3b. Rationality layer: the skill-based learning where explicit learning by conveying rules and symbols to a “learner” plays a role. (e.g. a child told by her mother not to get to close to the edge. A single example, not even experienced, might be enough to learn for life). Assimilating such explicit knowledge can steer and guide execution cycles which, “through acting” can create “tacit skills” within a different execution domain as the original layer.
3c. Logical layer: logics to represent causality, analogy and providing explanations.
3d. Planning and Problem-solving layer. A problem is understood, a final goal defined and the problem divided in sub-domains/problems which lead to a chain of ordered tasks to be executed, monitored (with intuition and rationality) and adapted.
In general: machine learning models that incorporate or learn structural knowledge of an environment have been shown to be more efficient and generalize better[4]. Some great examples of applications are not difficult to find, with the Neuro-Symbolic AI by MIT-IBM Watson lab as a good demonstration on how hybrid approaches (like NSQA in this case) can be utilized for learning in the connectionist way while preserving and utilizing the benefits of full-order logics in enhanced query answering in knowledge intensive domains like medicine. The NSQA system allows for complex query-answering, learns along, understand relations and causality while being able to explain results[5].
Latest developments in applied AI shows that we get far by learning from observations and empirical data, but that there is a need for contextual knowledge in order to make applied AI models trustable and robust in changing environments.
Takeaways –
Hybrid approaches are needed to model and use causality, counterfactual thinking, problem solving and structural knowledge of context.
Neural-symbolic processing: combine the benefits of connectionist and symbolic approaches to solve issues of trust, proof and explainability.
Contextual knowledge: AI needs modeling more of the world to be able to understand the physics and logics, causality, analogy in the surrounding world.
