Cognitive computing, the AI ​​that aims to equal the human

Where is research in cognitive computing? What are the first results? Can it claim to lead to a strong artificial intelligence?

Cognitive computing is often touted as one of the biggest challenges in computing today. It aims to provide a digital system with reasoning capabilities close to those of a human. Translation: create a strong AI. To achieve this, cognitive computing mainly combines two technologies: deep learning, particularly applied to automatic language processing (NLP), and symbolic artificial intelligence.

“In deep learning, symbolic AI makes it possible to integrate the notion of knowledge into the activation function of neurons”, explains to JDN Vincent Perrin, partner ecosystem technical leader at IBM. Knowledge that de facto makes part of the results of the model explainable and controllable, which is far from being true in the case of a purely probabilistic neural network. “For certain applications, such as questions and answers, this method also improves the level of confidence that users will have insofar as the results are based on a logic of causality (and not only on a probabilistic approach, editor’s note)”, continues Vincent Perrin. By forcing the network to follow business rules, it can also help to limit certain biases inherent in deep learning.

Learning with ever less data

Touted as one of the main levers of cognitive computing, neuro-symbolic AI surpasses traditional deep learning in areas such as image and video recognition and analysis. This type of model, dubbed logical neural networks (LNN), has also been shown to achieve a much higher level of accuracy than a classic deep neural network with lower volumes of training data. “Hence a faster training phase that consumes less computing capacity,” adds Vincent Perrin (read IBM’s blog post on the subject).

Still, the purpose of cognitive computing is to move towards a strong AI. “We are still very far from it. General artificial intelligence remains confined to theoretical research”, recalls Didier Gaultier. For the AI ​​director of Business & Decision (Orange group), researchers in strong AI should break away from neurobiology. “This analogy can certainly give ideas at the start, but our knowledge in this field is very insufficient to allow us to model the brain”, argues the data scientist. “It has been discovered that each of our cells probably stores the millions of years of evolution of the human body. This is a level of complexity that is not within our reach today.”

“Let’s take the example of Alexa or Siri assistants. These are multi-contextual AIs that implement cascading neural networks”

Are there other solutions to achieve strong artificial intelligence? Several avenues of development are emerging. Initially, neuro-symbolic AI could be applied to large NLP-oriented neural networks (Bert, GPT-3, LaMDA, etc.). “The idea is to achieve language processing models that can be generalized to a large number of use cases, while being safer in terms of results. But also to more frugal AI in terms of machine resources”, explains Vincent Perrin .

Then, LNNs could evolve into models capable of generating knowledge themselves. “The idea is to get new business rules out of the network which would then be reused as new functions for activating neurons”, says Vincent Perrin.

Neuroscience: a dead end?

Last but not least, strong AI could also find its lifeline in mathematics. “For example, we are at the very beginning of fundamental research in complex number neural networks, or Hilbertian spaces”, recalls Didier Gaultier, at Business & Decision. “These environments can potentially make it possible to model a much greater volume of information with less computational capacity.”

Another avenue mentioned by Didier Gaultier, within the scope of applied research this time: combining learning models. “Take the example of the assistants Alexa or Siri. These are multi-contextual AIs that implement cascading neural networks: the first to activate the speaker, the second to determine the question, the third to respond, etc.” Even if the challenge of building a strong AI is still far from over, cognitive computing is no longer science fiction.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *