Researchers Debate: Is Neuroscience the Foundation of Future AGI Systems?

The paper, ‘Toward Next-Generation Artificial Intelligence:

Catalyzing the NeuroAI Revolution’, co-authored by 27 highly prominent AI researchers and neuroscientists, proposes a roadmap for the path towards building an Artificial General Intelligence (AGI). An AGI system, unlike the Narrow Intelligence systems (ANI) which are designed to perform specific tasks like play chess—or participate in a game show like Jeopardy!—will be exposed to unpredictable environments, one that the system is not particularly trained on, and asked to navigate through the same.

For the authors, such a human-like artificial intelligence system is achievable. They categorise what is ‘human-like’ as systems that excel in “vision, reward-based learning, interacting with the physical world, and language”. 

Doubling down on NeuroAI research 

However, some of the key advances in AI research currently, such as convolutional Artificial Neural Networks (ANNs) and Reinforcement learning (RL), have been limited as they are built upon decade-old findings in neuroscience. According to the authors, the latest developments in the neuroscience field offer a broader scope for NeuroAI research on the path to AGI.  

The foreseeable step right now is to make systems that consist of a few basic ingredients of intelligence—namely, “adaptability, flexibility, and the ability to make general inferences from sparse observations”. These ingredients are already available in some form in most basic sensorimotor circuits. The paper argues that the neuroscience of embodied interaction with the world observed in all animals can be monumental in bringing the dream of a ‘human-like’ AI much closer. 

The idea takes inspiration from the evolutionary capabilities of animals to adapt to different environments. If the neural-level circuits of animals are broken down into their constituents, an AI system capable of the same can be emulated.  

Neuroscience research in AI development: Is it needed?

Since its release, the paper has renewed the discussion surrounding the role of neuroscience research in the development of AI systems. There are disagreements over whether neuroscience has had a tangible impact on AI modelling. 

What do the critics say? 

In response to the paper’s claim that neuroscience should continue to drive AI progress, DeepMind research scientist David Pfau said that neuroscience never drove AI in the first place, further adding that, “there’s a difference between drawing some high level inspiration from classic work and directly drawing on the latest research”. 

Sam Gershman, Professor in the Department of Psychology and Center for Brain Science at Harvard University, also adds to the discussion by expressing doubt that neuroscience research can directly deliver algorithms that can be plugged into the system. He writes, “new engineering ideas come from thinking about the structure of problems, not reading the tea leaves of biology.” 

Gershman also poses an interesting question that pivots the debate to a specific direction: Consider the counterfactual world where engineers knew nothing about neuroscience. Do you think that we wouldn’t have convolutional networks or reinforcement learning? 

The question propels us to think if the two fields driven by different levels of curiosity—conceptual and empirical—need merging. 

Adding to the list of critics, Luigi Alcerbi, Assistant Professor of Machine & Human Intelligence at the University of Helsinki, provides a fairly level-headed take over the current discourse, saying: “The importance of neuroscience on AI/ML development in the past is hard to quantify, but it’s fairly uncontroversial to say that some inspiration and ideas did come from neuro—although much less than one would expect or might like to admit. In the present, it’s close to zero.”

Alcerbi concurs with Pfau’s comment, adding that the influences neuroscience has had are all limited to high-level analogies used to model AI systems and not towards detailed biological implementation. 

In a similar vein, Alberto Romero, Analyst at Cambrian AI, explains that artificial neurons are extremely simple and are based on the 80-year old model of the neuron, compared to the current sophisticated models of the human brain. 

Experts on Neuroscience’s Potential for AI Research

Against the criticisms, several other researchers have made claims for how neuroscience has shaped, or can shape, the developments in AI/ML systems. 

Yann LeCun, Chief AI Scientist at Meta and one of the authors of the paper, writes this in response to such claims: 

https://twitter.com/ylecun/status/1583872918634655744

Similarly, the doubts over the impact of neuroscience on the field of AI research were also addressed by Surya Ganguli, Research Scientist at Meta and one of the contributors of the paper. Ganguli directs readers’ attention to an article he had written in 2018 wherein concrete examples of productive collaboration between biological and artificial systems in the past 60 years were provided. 

Gary Marcus, Professor Emeritus at NYU, also shared some key ideas in neuroscience yet to be embraced by Machine Learning models that must be included in the paper: 

https://twitter.com/GaryMarcus/status/1583997487702224896

Final Thoughts

Overall, there hasn’t yet been a serious rebuttal by the critics to the examples outlined above. Although, Pfau did respond to LeCun’s comment suggesting that neuroscience studies on detailed structures of a neuron or a cell do not directly correlate/answer to the problems that AI researchers work upon. 

The discussion so far leads us to believe that it is not so much whether neuroscience has been influential or not, but to what degree can latest neuroscience research help in solving some key engineering problems faced by the current AI systems in their pursuit of a general AI. 

However, what we know for sure is that neuroscience and AI share the same foundation—since AGI is dreamt on the fascination for building ‘human-like’ intelligent systems—until they reach a point of divergence, and this point of divergence is currently unknown. As long as the hope for AGI is alive, neuroscience will be a leverage AI research will hold onto to establish the foundation of the future models.