人脑细胞做成芯片打Doom！20万活体神经元自己探路杀敌，学习效率碾压深度强化学习

Core Insights - The article discusses the development of a biological computing system where 200,000 human brain cells, referred to as "brain PU," learned to play the classic game Doom through reinforcement learning techniques [1][10] - This achievement builds on previous work where brain cells learned to play Pong, showcasing advancements in translating digital game environments into signals that neurons can understand [12][15] Group 1: Learning Process and Performance - The process of teaching brain cells to play Doom was completed in under a week by independent developer Sean Cole using Cortical Labs' cloud platform API, contrasting with the 18 months taken for Pong [6][7] - The biological system demonstrated superior sample efficiency compared to three mainstream reinforcement learning algorithms (DQN, A2C, and PPO) in terms of key performance metrics such as average hits per game and error rates [20][22] - The study revealed that the biological cultures showed significant improvements in performance over time, while traditional algorithms did not exhibit similar enhancements [21][23] Group 2: Experimental Design and Findings - The research involved recording neural spike activities across 1,024 channels during 285 game sessions, with a sampling frequency of 20 kHz, allowing for detailed analysis of neural dynamics [16][25] - The experiments were designed to compare the biological system's performance against reinforcement learning algorithms under controlled conditions, ensuring that both systems had the same training volume [17][18] - The findings indicated that even with sparse input information, the biological system outperformed the algorithms, challenging the notion that more data always leads to better performance [22][23] Group 3: Implications and Future Directions - The research team introduced the concept of "Synthetic Biological Intelligence" (SBI), marking the first formal comparison between biological systems and reinforcement learning systems [29] - The study suggests that biological systems may rely on more efficient learning processes, such as predictive coding and active inference, which differ from traditional backpropagation methods [30][31] - Future goals include enhancing the capabilities of the neurons to not only play Doom effectively but also tackle more complex tasks, such as controlling robotic arms [38][39]