TO: Prof. Jason W. Ellis
FROM: Huzaifa Anas
DATE: September 17
SUBJECT: 500-Word Summary of Hassabis et al “Neuroscience-Inspired Artificial Intelligence”
Hassabis et al in Neuron argues that the field of neuroscience and AI (artificial intelligence) have a symbiotic relationship, but it’s in jeopardy, because of decreasing communication and collaboration. The contention states neuroscience provides a productive source of inspiration for algorithms and architecture, which is “independent of and complementary to the mathematical and logic-based methods and ideas that have largely dominated traditional approaches to AI” and “neuroscience can provide validation of AI techniques that already exist.” (Hassabis et al, 2017, p. 1). Moreover, they believe the progress in AI will eventually pay dividends to neuroscience by being a good test field. Within this article, past breakthroughs are examined to support this argument, while looking at how continued collaboration and communication can benefit both fields.
Two of AI’s backbones originate from neuroscience, which’s deep learning and reinforcement learning. Deep learning has revolutionized AI through dramatic advances in its neural and capable networks of learning freely from unstructured or unlabeled data. Reinforcement learning, the second pillar of modern AI, is a powerful tool enabling AI researchers to create software agents that act in an environment maximizing some sort of reward. In the 1940s artificial neural networks were developed, which could compute logical functions and ultimately “learn incrementally via supervisory feedback (Rosenblatt, 1958) or efficiently encode environmental statistics in an unsupervised fashion” (Hasabis, 2017, p. 2). This is the foundation for deep learning. Soon after backpropagation algorithms were made, which allowed learning to occur in networks of multiple layers whose value was recognized in 1986 by cognitive and neuroscientists working on Parallel distributed processing or PDP, which better-represented human-like behavior than serial logical processing, which AI researchers were focusing on. PDP has been applied to machine translation through the idea that “words and sentences can be represented in a distributed fashion (i.e., as vectors)” (Hasabis, 2017, p. 2). Deep learning ultimately became a field independent of PDP. Reinforcement learning comes from animal learning research, which Pavlov and Skinner pioneered. Reinforcement learning is used in robotic control, skillful play in backgammon and go.
If someone looks closely, AI research is still heavily inspired and guided by Neuroscience through AI work on attention, while eventually pivoting towards efficient learning and more independent behavior like transfer learning and imagination. The goal of AI is to form human-like behavior, and it’s practical an accurate biological framework as a reference. Attention is a critical issue currently because not all information is equal and therefore unlike before where all information was treated equally in neuroscience now information is being given different values, which allows for more efficient computing power usage. For the future, we want to decrease the computing power and a large amount of data needed for AI as currently. Humans can learn from a few examples, which AI can’t, and researchers are trying to apply developmental psychology ideas here. For imagination and transfer, learning neuroscience is still pioneering this part, but in the future, it’ll hopefully provide practical insights for AI work. All things considered, both fields can provide feedback to each other by having neuroscience provide ideas, and AI proves as a testing ground for these ideas. This isn’t compulsory, but just an effective and logical symbiotic relationship.
Article Cited APA format
Hassabis, D., Kumaran, D., Summerfield, C., & Botvinick, M. (2017). Neuroscience-inspired artificial intelligence. Neuron, 95(2), 245-258.
I’m not sure if restructuring definitions is considered plagiarism.