Conventional AI makes use of neural networks which are mostly realized in normal microprocessors, such as used in PC computers, that is on von Neumann architecture. It could be said that, in most implementations, AI is actually just an algorithm that is being executed on a normal digital computer.
But we know that the human brain does not operate upon digital (logic) circuits. If it did, doing math calculations would be one of the easiest human activities and we know it is in fact one of the hardest. Instead, brain works using and counting neuronal pulses, with each neuron having functions like: counting up or down, comparing two numbers, resetting the counter, generating an output pulse, etc.
Stochastic (aka random-pulse) computer, uses time-wise random pulses, much alike those found in human brain. In the past decade it has been shown that it can be VERY efficiently implemented into digital hardware, using only a few gates per mathematical operation, allowing for more functionality to be packed in a limited supply of hardware. It also possesses a high immunity to hardware failure and noise, much higher than digital computers for the same task. Finally, it operates without a clock: all computations are data driven, so at any moment it produces an optimum output given the input data.
Stochastic computing is a very promissing venue for AI, but hasn't yet been fully exploited. So the question for this discussion is: could it be implemented for improvement of robotics and AI in general and how?
Starting literature:
A. Alaghi et al. "The Promise and Challenge of Stochastic Computing," DOI: https://doi.org/10.1109/TCAD.2017.2778107
M. Stipčević, "Biomimetic random pulse computation or why Humans play basketball better than Robots?" DOI: https://doi.org/10.3390/biomimetics8080594
Yes, stochastic computing (SC) can potentially improve robotics and AI in several ways due to its unique approach to computation. It holds significant promise for improving the efficiency, robustness, and capability of robotics and AI systems, especially in environments where resources are constrained and robustness is critical. However, it requires careful integration and may not be suitable for all applications.
Stochastic computing is an emerging approach in artificial intelligence (AI) that leverages probabilistic representations and operations to perform computations. Unlike traditional binary computing, which uses deterministic operations on precise values, stochastic computing manipulates probabilities and random variables. This method can offer significant advantages in terms of power efficiency, fault tolerance, and parallel processing capabilities, making it particularly suitable for certain AI applications.
Deepak Sharma Indeed. And, as a crucial difference to neural networks and spiky neurons, it does explicitly perform math operations, which could make it useful n different task. I beleieve that a future computer will have differant sub processors (eg, digital logic, NNs spiky NNs, quantum computer, random-pulse computer...) that are optimal for different types of tasks. No single paradigm can do it all optimally.
Nowadays, artificial intelligence (AI) is mostly dominated by neural networks. Stochastic computing (SC) has lagged behind developments for decades but is rapidly catching up with modern computing, especially when precision is not of the utmost importance in the AI model inferencing use case. However, many people don't like the idea of randomness despite being recognized by the recent Abel Prize and Turing Award. You can say that the SC uses random pulses but it is mostly pseudo-random, possessing random behaviors but a predictable one. Interestingly, pseudo-random has higher accuracy than true random in the SC. That's why even though the working principle is based on randomness, the outputs of the SC system are pretty much deterministic provided with the same input.
Yang Yang Lee Very interesting remarks, thanks! The links of awards you mentioned are here:
https://www.newscientist.com/article/2425914-mathematician-wins-turing-award-for-harnessing-randomness/
https://www.smithsonianmag.com/smart-news/mathematician-who-made-sense-of-the-universes-randomness-wins-maths-top-prize-180984020/
Regarding precision, it is a wrong lane of research that pseudorandom pulse trains with tailored correlation is more precise. This is not true in general and that program of research entered a dead alley. Also it is a wrong concept that finbite pulse trains are to be used. On the contraty, true randomnness not only the sole conceivable source of randomness in o living systems but also yields a BETTER PRECISION: M. Stipčević, M. Batelić, “Entropy considerations in improved circuits for a biologically-inspired random pulse computer”, Scientific Reports, 12, 115 (2022) DOI: 10.1038/s41598-021-04177-9
In a nutshell: if randomness is important then better randomnes equals better result! It is a no-brainer.
Mario Stipčević you can google them. Although not directly related to the SC, the importance of randomness in mathematics and computer science is being recognized. Since randomness is part of SC, I believe it would gain more media attention in the future as an additional option other than analog or neuromorphic computing, especially in the AI workload.
By the way, regarding how SC could be implemented, there are only 2 options as far as I know, i.e., ASIC or FPGA. Binary computing can simulate the SC operation but it is only meant for case studies.
Yang Yang Lee Thanks for the second reply. Not seeing it, I upgraded my previous answer which you may wish to re-visit it.
Anyways, we do agree that SC, or to be more precise random pulse train computation, is an important re-emerging field. I hope more researchers will be drawn into it.
Yes, we use the FPGA approach, and we did use ASIC for RFF realization on a chip in std 55-nm CMOS.
BTW, there is no such thing as randomness in mathematics. But you know, saying this today, even if backed up with science, is as saying that there are not 4 elements but 92 in old Greece. They could listen, but would not be able understand for the next 2000 years...
- Badges
- Science topic
- Similar topics
- Psychology