2023年6月20日
Brain & Neuroscience Workshop 2023
地点： Virtual
Date: 9:00am – 12:00pm China Standard Time, June 20, 2023
| Time | Session | Title | Speaker |
|---|---|---|---|
| 9:00-9:05 | Opening & Introduction | Opening remarks | Dongsheng Li, MSR Asia |
| 9:05-10:35 | Invited Talks | Exploring robotic minds using the free energy principle
Abstract: The aim of this research is to investigate the mechanisms underlying embodied cognition, with a particular focus on the top-down and bottom-up interactions between an organism and its environment. We developed a variational RNN model based on the free energy principle and applied it to a series of robotics experiments, using predictive coding and active inference frameworks. Our results demonstrate that the estimated precision in the top-down prediction develops differently depending on a meta-parameter known as the meta-prior, w. A robot trained with a larger w tends to exhibit stronger top-down intention, while a robot trained with a smaller w exhibits weaker top-down intention and tends to follow the bottom-up sensation more closely. We discuss how the precision structure developed could affect embodied cognitive processes in interactions with physical objects and other agents in social cognition. Additionally, we speculate on the potential for online adaptation of the meta-prior and its impact on embodied cognition. |
Prof. Jun Tani, Okinawa Institute of Science and Technology (OIST) |
| Brain design principles of efficient neural information processing
Abstract: The evolution of biological brains has led to the development of energy-intensive networks with diverse intelligent functions. This study aims to examine the trade-off between energy consumption and brain complexity in biological evolution, and explore how the biological brain efficiently relocates its energy for computational functions and information communications to improve artificial intelligence design. Through re-examining previous studies, we uncovered underestimations of the size of the axon, number of synapses, neurotransmitter release probability in the human brain, and differences between excitatory and inhibitory neurons. We also relocated the glial energy of Ca2+ influx and maintaining resting potential to communication. Our calculations show that the ratio of computation to communication in the human and rat brain is considerably different from previous estimates, with a range of 1:1.6 to 1:3.6 and 1:1.54 to 1:0.88, respectively. The high human brain communication is due to costly white matter, indicating increased information exchange between brain regions to perform prodigious functions. Computers have a ratio of 1:9 to 1:209 due to high energy consumption during operations and data movement. These findings suggest that the biological brain is highly energy-efficient in communication and computation, with an energy cost per bit that is lower than that of a computer. These insights may help to enlighten the development of more energy-efficient neuromorphic chips and deep learning algorithms. |
Prof. Yuguo Yu, Fudan University | ||
| Brain-like AI does not think like brains
Abstract: It is increasingly hard to distinguish AI from humans based on their behavior. That said, human-like, brain-like AI does not necessarily think like a brain. One way to investigate this issue is to examine fundamental questions that are easy to resolve for the brain but not for AI, such as goal-directed learning, metacognition, predictive cognition, and intuitive physics. First, I will introduce a framework to build reinforcement learning (RL) models from human behavior data without underfitting and overfitting (Brain↦AI), making it possible to discover neural evidence of human metacognitive reinforcement learning. Second, I will present a new framework, neural task control, in which an auxiliary RL algorithm learns to create a synthetic experience that guides human RL at the behavioral and neural level (AI↦Brain). Since this control framework broadens human experience, it will guide us to develop new hypotheses about brain functions and ultimately initiate a recursive research process: ((Brain↦AI)↦Brain)↦… |
Prof. Sang Wan Lee, KAIST | ||
| 10:35-11:15 | MSR Asia Talks | Goals and habits in synergy: A variational Bayesian framework for behavior
Abstract: Behaving efficiently and flexibly is crucial for biological agents and embodied AI. It is well recognized that behavior is classified into two types: reward-maximizing habitual behavior, which is fast but inflexible; and goal-directed behavior, which is flexible but slow. These behaviors are typically considered managed by two distinct systems in the brain. Here, we propose a unified approach based on variational Bayesian theory, incorporating both behaviors into a single deep learning framework using a probabilistic latent variable called “intention”. Habitual behavior is generated using the goal-less, prior distribution of intention, while goal-directed behavior uses the posterior distribution of intention that conditions on the goal. This novel Bayesian framework enables skill sharing between the two behaviors, and by leveraging predictive coding, it enables seamless generalization from habitual to goal-directed behavior without requiring additional training. Our work suggests a fresh perspective for cognitive science and embodied AI, yielding greater integration between habitual and goal-directed behaviors. |
Dongqi Han, MSR Asia |
| Rapid Context Inference in a Thalamocortical Model using Recurrent Neural Network
Abstract: Cognitive flexibility is a fundamental ability that enables humans and animals to exhibit appropriate behaviors in various contexts. The thalamocortical interactions between the prefrontal cortex (PFC) and the mediodorsal thalamus (MD) have been identified as crucial for inferring temporal context, a critical component of cognitive flexibility. However, the neural mechanism responsible for context inference remains unknown. To address this issue, we propose a PFC-MD neural circuit model that utilizes a Hebbian plasticity rule to support rapid, online context inference. Specifically, the model MD thalamus can infer temporal contexts from prefrontal inputs within a few trials. This is achieved through the use of PFC-to-MD synaptic plasticity with pre-synaptic traces and adaptive thresholding, along with winner-take-all normalization in the MD. Furthermore, our model thalamus gates context-irrelevant neurons in the PFC, thus facilitating continual learning. We evaluate our model performance by having it sequentially learn various cognitive tasks. Incorporating an MD-like component alleviates catastrophic forgetting of previously learned contexts and demonstrates the transfer of knowledge to future contexts. |
Weilong Zheng, MSR Asia & Shanghai Jiao Tong University (SJTU) | ||
| 11:15-11:55 | Panel Discussion | Challenges and opportunities for synergies between neuroscience and AI | Moderator: Miran Lee, MSR Asia Panelists:
|
| 11:55-12:00 | Closing | Closing remarks | Dongsheng Li, MSR Asia |