EFRI BRAID: DenPro3D – Dendritic Processing of Spike Sequences in Biological and Artificial Brains

人工智能（AI）在过去十年中发展迅速，但在能源、美元和碳排放方面部署和运营人工智能的成本正在不可持续地增长。目前，用户通过云访问人工智能，牺牲了个性化和隐私。该项目利用来自大脑的灵感，包括神经科学家最近发现的树突学习，在神经科学理论和实验技术的指导下，对大脑的学习规则进行逆向工程。这些见解将在使用新兴的三维（3-D）制造技术的新型神经形态芯片中实现。这种方法的成功将使人工智能能够运行，而不是在云端运行兆瓦级的电力，而是在智能手机上运行兆瓦级的电力。因此，树突学习可以控制不可持续增长的成本，公平地分配生产力收益，个性化用户体验，并恢复隐私。该项目还旨在通过正在进行的暑期实习项目和新的推广活动（高中生校园演示日和服务不足的高中工程之夜），招募黑人青少年，特别是女孩，为代表性不足的学生增加机会。将计算单元平铺并在三维空间中堆叠，缩短了距离，减少了能源通信的使用，而通信现在主导着当今处理器的能源预算。但是堆叠减少了散热的表面积，限制了3-D处理器的串行操作，而不是并行操作。如果单位之间的通信稀疏，产生的热量就会减少。这可以通过将二幅信号（例如，数字总线上的高电压或低电压）的模式交换为一幅信号序列（例如，来自神经元集合的尖峰信号）来实现。但这需要合成连接，以精确地排列树突状设备上的突触输入，将以突触学习为中心的大脑的主流抽象转变为以树突学习为中心的大脑，或者项目团队所说的树突中心学习。为了实现这一目标，提出了三个目标：(1)确定脉冲序列如何将信息表示为跨皮质列的子序列，并对列如何学习生成子序列进行建模。(2)建立相邻树突段用于解码序列子序列的组合逻辑，并对树突段如何学习检测子序列进行建模。(3)在纳米级电子器件中模拟一段树突的序列选择性，将其集成到3- d中，并设计一个交换网络来实现目标1和目标2中制定的学习规则。实现这些目标将使在资源稀少的环境中学习具有极高的能源效率，从而对信息技术如何服务社会产生变革性影响，同时确保公平和获取。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Artificial Intelligence (AI) has progressed rapidly over the past decade, but the cost of deploying and operating AI in energy, dollars, and carbon emissions is growing unsustainably. Users currently access AI through the cloud, sacrificing personalization and privacy. This project uses inspiration from the brain, including learning with dendrites that neuroscientists have recently discovered, to reverse engineer the brain's learning rules guided by neuroscience theory and experimental techniques. These insights will be implemented in a novel neuromorphic chip using emerging three-dimensional (3-D) fabrication techniques. The success of such an approach would allow AI to run, not with megawatts in the cloud, but rather with watts on a smartphone. Thus, learning with dendrites could reign in unsustainably growing costs, distribute productivity gains equitably, personalize user experience, and restore privacy. This project also aims to increase opportunities for underrepresented students by recruiting black teenagers, particularly girls, through ongoing summer internship programs and new outreach efforts (Demo Day on campus for high-schoolers and Engineering Night at underserved high schools).Tiling computing units and stacking them in the third dimension shortens distances and cuts the energy communication uses, which now dominates the energy budget of today’s processors. But stacking reduces the surface area for dissipating heat, restricting a 3-D processor to serial, rather than parallel operation. Less heat would be produced if units communicate sparsely. This could be accomplished by exchanging patterns of binary-amplitude signals (e.g., high or low voltages on a digital bus) for sequences of unary-amplitude signals (e.g., spikes from an ensemble of neurons). But this would require synthesizing connections to precisely order synaptic inputs on a dendrite-like device, shifting the prevailing abstraction of a brain centered on learning with synapses to one centered on learning with dendrites or what the project team refers to as dendrocentric learning. To accomplish this goal, three objectives are proposed: (1) Identify how spiking sequences represent information as subsequences across cortical columns and model how a column learns to generate its subsequence. (2) Establish the combinatorial logic neighboring stretches of dendrite use to decode a sequence’s subsequences and model how a stretch of dendrite learns to detect a subsequence. (3) Emulate a stretch of dendrite’s sequence selectivity in a nanoscale electronic device, integrate it in 3-D, and design a switching network to implement the learning rules formulated in objectives 1 & 2. Achieving these objectives would enable learning in sparse environments with extreme energy efficiency leading to a transformational impact on how information technology serves society while ensuring equity and access.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.