The paper “Towards the next generation of artificial intelligence:
Catalyzing the NeuroAI Revolution, co-authored by 27 top AI researchers and neuroscientists, proposes a roadmap on the road to building artificial general intelligence (AGI). An AGI system, as opposed to the narrow intelligence systems (ANI) which are designed to perform specific tasks such as playing chess or participating in a game show Danger!– will be exposed to unpredictable environments for which the system is not particularly trained and will be asked to navigate through them.
According to the authors, such a human-like artificial intelligence system is feasible. They categorize what are “human-like” as systems characterized by “vision, reward-based learning, interaction with the physical world, and language.”
Doubling NeuroAI research
Some of the most important advances in AI research, such as However, new technologies such as convolutional artificial neural networks (ANNs) and reinforcement learning (RL) have been limited as they build on decades of neuroscience. According to the authors, the latest developments in the field of neuroscience offer a broader scope for NeuroAI research towards AGI.
The current foreseeable step is to develop systems that consist of a few basic components of intelligence – namely “adaptability, flexibility and the ability to draw general conclusions from sparse observations”. These ingredients are already present 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 “human-like” AI much closer.
The idea is inspired by the evolutionary ability of animals to adapt to different environments. If the neural circuits of animals are broken down into their component parts, an AI system capable of doing this can be emulated.
Neuroscience Research in AI Development: Is It Needed?
Since its publication, the paper has renewed the discussion about the role of neuroscientific research in the development of AI systems. There is disagreement as to whether neuroscience has had a tangible impact on AI modelling.
What are the critics saying?
In response to the paper’s assertion that neuroscience should continue to drive AI advances, the DeepMind researcher said David Peacock said that neuroscience has never advanced AI at all, adding that “there is a difference between taking high-level inspiration from classical work and going straight to the latest research.”
Sam Deutschman, a professor in the Department of Psychology and Center for Brain Science at Harvard University, also contributes to the discussion by expressing doubts that neuroscience research can directly provide algorithms that can be plugged into the system. He writes, “New technical ideas come from thinking about the structure of problems rather than reading the tea leaves of biology.”
Gershman also poses an interesting question that steers the debate in a specific direction: Consider the counterfactual world where engineers knew nothing about neuroscience. Do you think we wouldn’t have convolutional networks or reinforcement learning?
The question prompts us to consider whether the two areas driven by different levels of curiosity—conceptual and empirical—need to be merged.
Adding to the list of critics, Luigi Alcerbi, Assistant Professor of Machine and Human Intelligence at the University of Helsinki, takes the current discourse quite soberly and says: “The importance of neuroscience for AI/ML development in the past is difficult to quantify, but it is quite uncontroversial to say, that some inspiration and ideas came from Neuro – albeit much less than one might expect or like to admit. In the present it is close to zero.”
Alcerbi agrees with Pfau’s comment, adding that the influences of neuroscience are all limited to high-level analogies used to model AI systems, rather than a detailed biological implementation.
Similarly, Cambrian AI analyst Alberto Romero explains that artificial neurons are extremely simple and based on the 80-year-old model of the neuron compared to the current sophisticated models of the human brain.
Experts on the potential of neuroscience for AI research
Contrary to criticism, several other researchers have claimed how neuroscience has shaped or can shape developments in AI/ML systems.
Yann LeCun, senior AI scientist at Meta and one of the paper’s authors, writes this in response to such claims:
Similarly, the doubts about the impact of neuroscience on the field of AI research were also raised by Surya Ganguli, research scientist at Meta and one of the paper’s contributors. Ganguli draws readers’ attention to an article he wrote in 2018 that gave concrete examples of productive collaboration between biological and artificial systems over the past 60 years.
NYU Professor Emeritus Gary Marcus also shared some key ideas in neuroscience that have yet to be adopted by machine learning models to be included in the paper:
Overall, the critics have not yet seriously countered the examples outlined above. Although peacock did reply to LeCun’s comment suggesting that neuroscientific studies of detailed structures of a neuron or cell do not directly correlate/answer to the problems AI researchers are working on.
The discussion so far leads us to believe that what matters is not so much whether or not neuroscience has been influential, but to what extent the latest neuroscience research can help solve some important technical problems that current AI systems face in their quest after a general AI are faced .
What we do know for sure, however, is that neuroscience and AI share the same foundation – as AGI is dreamed of by the fascination of building “human-like” intelligent systems – until they reach a point of divergence, and that point of divergence is currently unknown. As long as the hope for AGI is alive, neuroscience will be a lever that AI research will hold onto to lay the foundation for future models.