A little light neuromorphic computing for Friday
Intel has revealed details about its Loihi 2 Neuromorphic Compute Tile, a completely different approach to computing. Loihi 2 does not process data like the familiar silicon that powers the device you are reading this on, nor does it process data like the deep learning machines that most HPCs focus on, nor does it resemble quantum processors beginning to evolve into more than just proof device concepts.
Loihi 2 is instead a new type of neural network on a chip, which seeks to mimic the neurons, axons, and dendrites found in the legs of most creatures you encounter during the day. In quantum computers, a series of qubits is specifically set up to solve a particular problem and uses the rules of quantum physics to instantly generate a result that is assigned the probability of being accurate.
Quantum computers are incredibly powerful at solving the problems they are designed for and can produce instantaneous output that a traditional computer could take thousands of hours to process. The problem is that there is no flexibility in the system, the cubes need to be configured separately for each problem that is entered into the system. There is also a cooling problem, even the LN2 setting is not enough to keep that type of system in a stable condition.
Traditional models of deep learning are also quite different from neuromorphic computing, the HPC devices we’ve seen from NVIDIA and others consist of a large network of artificial neurons that respond differently to inputs depending on how and where they are placed. In order to get more detailed results, the network needs to be increased, or as the name suggests, deeper, resulting in increased costs and heat production.
Another challenge to achieving decent results from a neural network is that it needs to be trained for the task assigned to it. For example, convincing him to properly process and identify objects in an image or video requires a long training session that, when completed, will allow him to efficiently process and identify the objects for which he is trained. What he won’t be able to do is identify things he hasn’t trained on, or even those things he was trained to recognize if they are partially obscured or if the object is at an angle where they are not trained.
That lack of flexibility inherent in standard neural networks is exactly what Intel is trying to avoid with its Loihi 2 Neuromorphic Compute Tile. This new architecture is designed with flexibility as the primary goal, and accuracy of results as the secondary goal. This does not mean that they will spit out completely wrong answers, but Intel is aiming for a result that is good enough.
The example given by Ars Technica shows how the performance of an autonomous robot will change over time, how the wear of parts increases and the friction of the joints. Currently, these changes are being ignored, the robot has been fixed because it cannot function due to large differences in performance, or it is necessary to regenerate and install a deep learning algorithm that controls the robot. With Loihi 2 controlling the robot, it could theoretically detect changes in the robot’s physical capabilities and generate new control algorithms that take changes into account, without the need for programmer intervention.
If you are more interested in theory and hardware than theory, a visit to ServeTheHome will be worthwhile.
Friendly communicator. Music maven. Explorer. Pop culture trailblazer. Social media practitioner.