Tuesday, May 30, 2023 @ 8:00-9:00am PT / 17:00-18:00 CET you are invited to attend an INRC Forum talk from Jason Eshraghian & Ruijie Zhu, University of California, Santa Cruz.
Scaling up SNNs with SpikeGPT
Abstract: If we had a dollar for every time we heard "It will never scale!", then neuromorphic engineers would be billionaires. This presentation will be centered on SpikeGPT, the first large-scale language model (LLM) using spiking neural nets (SNNs), and possibly the largest SNN that has been trained using error backpropagation.
The need for lightweight language models is more pressing than ever, especially now that we are becoming increasingly reliant on them from word processors and search engines, to code troubleshooting and academic grant writing. Our dependence on a single LLM means that every user is potentially pooling sensitive data into a singular database, which leads to significant security risks if breached.
SpikeGPT was built to move towards addressing the privacy and energy consumption challenges we presently run into using Transformer blocks. Our approach decomposes self-attention down into a recurrent form that is compatible with spiking neurons, along with dynamical weight matrices where the dynamics are learnable, rather than the parameters as with conventional deep learning.
We will provide an overview of what SpikeGPT does, how it works, and what it took to train it successfully. We will also provide a demo on how users can download pre-trained models available on HuggingFace so that listeners are able to experiment with them.
Bio(s):
Dr. Jason Eshraghian is an assistant professor of Electrical and Computer Engineering at UC Santa Cruz. He is the developer of snnTorch, a widely adopted Python library used to train and model brain-inspired spiking neural networks. He was awarded the IEEE TCAS-I Darlington'23, IEEE TVLSI'19, and IEEE AICAS'19 best paper awards, and the best live demonstration award at IEEE ICECS'20. He was the recipient of the Fulbright Future Fellowship (Australian-America Fulbright Commission), the Forrest Research Fellowship (Forrest Research Foundation), and the Endeavour Fellowship (Australian Government). He leads the UCSC Neuromorphic Computing Group which focuses on porting principles from neuroscience into building more effective learning algorithms in software and hardware. Dr. Eshraghian is the Secretary of the IEEE Neural Systems and Applications Committee and an Associate Editor with APL Machine Learning.
Ruijie Zhu is commencing his Ph.D. in Electrical and Computer Engineering at UC Santa Cruz in the Fall of 2023. He recently completed his Bachelor Degree in Computer Science at the University of Electronic Science and Technology of China, where he became a regular contributor to open-source neuromorphic projects, including snnTorch, SpikingJelly, and led the development of SpikeGPT, the first spiking neural network generative language model. He was elected as the chair of the 2020 Students Open-Source Conference (SOSConf), which attracted over 3,000 online participants. His research focus is on enabling the development of large-scale spiking neural networks.
Recording & Slides:
For details on past and future INRC talks, see the full INRC Forum Summer 2023 Schedule (accessible only to INRC Affiliates and Engaged Members).
If you are interested in becoming a member, here is the information about joining the INRC.
Intel vLab systems for the Neuromorphic Research Cloud and Loihi 2 systems will be unavailable periodically due to regular maintenance from today, 23 May 2023 until Thursday 26 May 2023. If you encounter any disruption to your vLab access via SSH or ability to run models on Loihi 2 systems, please try again after the maintenance period. Please contact nrc_support@intel-research.net for questions.
Tuesday, May 23, 2023 @ 8:00-9:00am PT / 17:00-18:00 CET you are invited to attend an INRC Forum talk from Wallace "Ed" Lawson, Naval Research Lab.
Sigma-Delta Networks for Robot Arm Control
Abstract: Our autonomous robot, Bight, can be a reliable teammate that is capable of assisting in performing routine maintenance tasks on a Naval vessel. In this talk, we consider the task of maintaining the electrical panel. A vital first step is putting the robot into the correct position to view all of the parts of the electrical panel. The robot can get close, but the arm of the robot will need to move to where it can see everything. A second related task is using Bight to fetch the tools needed to perform this task. Here, we propose to solve this using a sigma-delta spiking network that is trained using deep Q learning. Our approach is able to successfully solve this problem at varying distances. While we show how this works on these specific tasks, we believe this is general enough to work with a variety of related tasks.
Bio:
Dr. Wallace "Ed" Lawson is a research scientist at the Navy Center for Applied Research in Artificial Intelligence at the Naval Research Lab in Washington, DC where has worked since 2004. Dr. Lawson received his PhD in computer science from George Mason University in 2011, studying motion analysis applied to biometrics. In his research, he is interested in using machine learning to solve perception problems in robots. He has also authored publications related to human-robot interaction, biometric person authentication, and surveillance.
Recording:
For the recording, see the full INRC Forum Summer 2023 Schedule (accessible only to INRC Affiliates and Engaged Members).
If you are interested in becoming a member, here is the information about joining the INRC.
Tuesday, May 9, 2023 @ 8:00-9:00am PT / 17:00-18:00 CET you are invited to attend an INRC Forum talk from Bradley Theilman, Sandia National Laboratories.
Stochastic Neuromorphic Circuits for Solving MAXCUT
Abstract: Finding the maximum cut of a graph
(MAXCUT) is a classic optimization problem that has
motivated parallel algorithm development. In this talk, I will present two neuromorphic circuits that transform a source of randomness into computationally useful correlations for approximating solutions to graph MAXCUT. Neuromorphic computing has been successfully applied to various graph algorithms, by exploiting the analogy between a graph and the connectivity of a neural circuit. However, the physical constraints of neuromorphic hardware make translating an arbitrary graph into the neuromorphic domain challenging. Neuromorphic computing is also beginning to explore stochastic devices as efficient sources of randomness for large-scale stochastic algorithms. Graph MAXCUT is a well-known NP-complete discrete optimization problem with the best-known approximate solutions being stochastic algorithms, such as the Goemans-Williamson algorithm. I will show how to combine large-scale sources of intrinsic randomness with neuromorphic principles to implement two classes of stochastic approximations to graph MAXCUT in neuromorphic hardware. These approaches have architectural advantages over other neuromorphic graph algorithms and benefit from the theoretical performance guarantees of their algorithmic inspirations. I will show results from simulations of these circuits as well as results from an implementation of one of these circuits on Intel’s Loihi neuromorphic system. This work opens a new direction for stochastic neuromorphic circuits applied to discrete optimization.
Bio:
Bradley Theilman is a postdoctoral appointee at Sandia National Laboratories. His research focuses on applying neuroscientific principles to neuromorphic computing. He earned a Ph.D. in computational neuroscience in 2021 from UC San Diego, where he worked on topological approaches to understanding neural population activity in the auditory regions of songbird brains in the laboratory of Dr. Tim Gentner.
Recording:
For details on past and future INRC talks, see the full INRC Forum Summer 2023 Schedule (accessible only to INRC Affiliates and Engaged Members).
If you are interested in becoming a member, here is the information about joining the INRC.
Summary:
This demo centers on a key challenge for the space technology industry: scheduling a large number of Earth observation requests to a constellation of satellites. The problem can become impossible to solve at large scales because we have a finite amount of time to apply an algorithm that slows down proportional to the number of satellites and the square of the number of requests. For commercial satellite companies with dozens of vehicles and thousands of customer requests, current algorithms will not find the best solution in time.
NCL has released a new software library, Lava Optimization, which includes tools and models for solving NP-Hard computational problems like the satellite scheduling problem. Using Lava Optimization, we can map the satellite problem to a neuromorphic solver called QUBO that runs on Intel Loihi 2. The solver uses brain-inspired principles of parallel computation to consider far more potential schedules in less time and using less energy. This will enable larger satellite constellations to serve more customers than is possible today.
The neuromorphic solver can also be applied to find optimal solutions to a wide range of NP-Hard problems across many industries, such as routing a fleet of delivery vehicles through real-time traffic, dynamically assigning tasks to systems in a data center, and selecting the perfect portfolio in rapidly changing market conditions.