The performance of machine learning models has improved dramatically on various real-world tasks due to recent technological developments. However, most of these models need a significant amount of computational power, which makes the training and implementation of these models a difficult task. To make machine and deep learning models function more smoothly and effectively, researchers have now sought to look for several hardware alternatives.
One of these approaches relies on integrating neural networks with memristors or other cutting-edge memory techniques. Memristors are electrical components that control how much current flows while keeping track of how much energy has already passed through it. The fact that they are non-volatile or preserve memory without requiring energy contributes majorly to their significance. When it comes to applications with certain levels of uncertainty, limited data access, and requiring explainable decision-making, neural networks do not always turn out to be the ideal option.
Bayesian reasoning can be applied in this situation. However, putting bayesian models into practice is exceedingly computationally expensive, and in contrast to neural networks, these models do not naturally convert to memristor-based designs. Working on this problem statement, several researchers from various French universities, including Université Paris-Saclay-CNRS, Université Grenoble-Alpes-CEA-LETI, HawAI.tech, Sorbonne University, and Université d’Aix-Marseille-CNRS collaborated to develop a Bayesian machine that uses memristors and is designed for extremely energy-efficient Bayesian reasoning. The paper highlighting their research is also published in the scientific journal Nature Electronics.
Artificial intelligence systems’ requirement for energy is mostly caused by the separation of memory and computational tasks. Since the models use a lot of training data, they demand a considerable chunk of memory, which is expensive to access in terms of energy. In contrast to neural networks, human brains are far more energy efficient because the memory and computing processes are linked as closely as possible. The researchers used this concept as a foundation for creating their architecture.
By employing stochastic computing and distributed memory, the machine’s design is created by formulating Bayes’ law in a way that makes its application intuitive. This makes the circuit significantly more energy-efficient than previous hardware solutions since it enables the circuit to operate with only local memory and minimal data movement. Furthermore, memristors can better mimic the information processing techniques used by the human brain because they can perform both computations and act as memory storage.
The prototype of the bayesian machine consists of 30,080 transistors and 2,048 hafnium oxide memristors. The researchers used a gesture recognition task to show that their prototype recognizes certain human motions using 5000 times less energy than a conventional microcontroller unit. Other characteristics of the Bayesian machine include quick on/off functionality, suitability for low supply voltages, and resistance to single-event upsets. These remarkable findings pave the way for Bayesian reasoning as an appealing strategy for reliable and energy-efficient models.
Bayesian reasoning can be used as an alternative AI strategy in situations when deep learning has limitations, like a lack of data availability. This is due to its adequate performance in limited data scenarios and its capacity to deliver entirely comprehensible results. The team strongly hopes that their memristor-based Bayesian machine will contribute significantly to the improvement of the efficiency of AI models in the future. They anticipate that other researchers may use the design as a springboard for creating other related technologies, such as safety-critical applications like medical sensors.
The researchers further explained that they are currently developing a considerably scaled-up version of the Bayesian machine. They are also working on applying the machine’s underlying concepts to other machine-learning techniques. The researchers also strive to overcome certain limitations they face while scaling up their method, using memristor-based solutions.
Check out the Paper. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our Reddit Page, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.
Khushboo Gupta is a consulting intern at MarktechPost. She is currently pursuing her B.Tech from the Indian Institute of Technology(IIT), Goa. She is passionate about the fields of Machine Learning, Natural Language Processing and Web Development. She enjoys learning more about the technical field by participating in several challenges.