New possibilities in artificial intelligence Huge volumes of data are available today on many of the most pressing challenges facing society, yet current computing architectures are relatively inefficient at handling this data. We spoke to Dr Abu Sebastian about the work of the Projestor project in developing a new memory device concept which could open up new possibilities in Artificial Intelligence. The typical Artificial Intelligence (AI) system is currently run on computers which consume kilowatts or even megawatts of power. By contrast the human brain consumes something like 20 watts, so there is a big disparity in the amount of power required. “A big factor behind this disparity is that the architecture with which the brain computes is fundamentally different from how conventional computers work. For example, in the brain there is no notion of processing happening in one place, and memory being stored in another – they are co-located,” explains IBM researcher Dr Abu Sebastian. As the Principal Investigator of the Projestor project, Dr Sebastian is now exploring a new computing paradigm. “In the project we are trying to come up with new physical computing systems which are, in a sense, inspired by the architecture of the brain,” he explains. “We are looking at having co-located memory and processing for example, and trying to bring down this disparity between conventional computing systems and the brain. We call it in-memory computing.” New computing architecture This research is built on a recognition of the limitations of the existing computing architectures. Current computing systems are based on the von Neumann architecture, where storage is in one location and the processing engine elsewhere; every time a computation is performed, data is shuttled between the two units. “This is very energy-intensive,” outlines Dr Sebastian. Researchers in the project aim to develop new solutions, new architectures, where data doesn’t need to be moved between memory and processing, greatly improving energy efficiency. “Conventional memory can be viewed as a place where data is dumped, and there isn’t any intelligence associated with that. It’s just a place where you store stuff,” says Dr Sebastian. “We’re trying to design a new kind of computational memory chip, where the memory is actually an active participant in the computation process. It’s not enough to view memory as a place where you simply store information.”
A computational memory chip based on phase change memory devices.
The human brain is an importance source of inspiration in this respect, as it’s among the best available examples of a cognitive computer, with very closely entwined memory and processing units. While
using memory itself, through the synapses. So, how about building memory chips which can compute?” he outlines. “We take the fact that you are processing in memory by exploiting some physical attributes of the storage mechanism. In the case of brain architecture, that’s done through the synapses. The way in which we store information in our computational memory units is through a type of memory called phase change memory.” In a phase change memory device, information is stored in terms of the atomic configuration of certain types of materials. If the atoms are in an ordered state, it is logic 1, and if it is in a disordered state the logic is 0, while there are also intermediate states. “The idea is to use the physical properties
AI in everyday life. While some of us may still think of AI as being quite a futuristic idea, Dr Sebastian says that in fact it is in everyday use today. “Web companies use it for search, mobile cameras use it for identifying objects in pictures. There are many applications, including in robotics, the Internet of Things (IoT) and healthcare,” he outlines. For example, Dr Sebastian’s colleagues are using AI and sensors to understand the progression of lung diseases based on the sound of a cough and the colour of a patient’s salvia. “The AI uses this data to potentially predict when the patient is about to have an acute event, called exacerbations, where they nearly suffocate and require re-hospitalization. Unless the doctor is monitoring them 24/7 this would be impossible,” he points out. Many of these types of AI applications are currently run in the cloud, so when an individual uses their mobile phone to translate some text for example, the translation and the data processing are done elsewhere. While the cloud
PROJESTOR Projected Memristor: A nanoscale device for cognitive computing Project Objectives
We are entering the era of cognitive computing, which holds great promise in terms of deriving intelligence and knowledge from huge volumes of data. However, it is becoming clear that to build efficient cognitive computers, we need to transition to non-von Neumann architectures where memory and logic coexist in some form. The main goal of the Projestor project is to explore such a memory device concept
ERC Consolidator Grant https://www.ibm.com/blogs/research/2016/03/ ibm-scientist-abu-sebastian-develops-futurememory-computer-paradigms-prestigiouseuropean-grant/
Dr Abu Sebastian Principal Research Staff Member IBM Research - Zurich Säumerstrasse 4 8803 Rüschlikon Switzerland T: +41 44 724 8684 E: ASE@zurich.ibm.com W: http://www.erc-projestor.eu/ W: https://researcher.watson.ibm.com/ researcher/view.php?person=zurich-ASE M. Salinga, B. Kersting, I. Ronneberger, V.P. Jonnalagadda, X.T. Vu, M. Le Gallo, I. Giannopoulos, O. Cojocaru-Miredin, R. Mazzarello, A. Sebastian, “Monatomic phase change memory,” Nature Materials, 17 681–685, 2018 (Cover).
We are trying to come up with new physical computing systems which are, in a sense, inspired by the architecture of the brain. We are looking at having co-located memory and processing for example. researchers are still only scratching the surface when it comes to understanding the architecture of the brain, Dr Sebastian says it’s still possible to derive useful insights. “The brain seems to compute
will still be required for more advanced AI applications, improving the energy efficiency of computing systems could open up the possibility of performing some tasks in a mobile device and widening the use of AI. “The idea is to make AI more pervasive – to make your mobile phone or your healthcare device more intelligent for example. So we really want to bring down the energy consumption of these computers that are doing the AI,” says Dr Sebastian. A lot of progress has been made in these terms, with the project making an important contribution to the development of a new computing architecture. “The goal is to establish in-memory computing as a post von-Neumann computing paradigm,” continues Dr Sebastian. This could eventually lead to AI becoming a part of more everyday devices. It is already quite commonly used today of course, but in order to widen its applications still further, Dr Sebastian says continued technological development is required.
M. Le Gallo, D. Krebs, F. Zipolli, M. Salinga, A. Sebastian, “Collective structural relaxation in phase-change memory devices,” Adv. Electronic Materials, 2018. M. Le Gallo, A. Sebastian, R. Mathis, M. Manica, H. Giefers, T. Tuma, C. Bekas, A. Curioni, E. Eleftheriou, “Mixed-precision in-memory computing,” Nature Electronics 1, 246–253, 2018. A. Sebastian, T. Tuma, N. Papandreou, M. Le Gallo, L. Kull, T. Parnell, E. Eleftheriou, “Temporal correlation detection using computational phase-change memory,” Nature Communications 8, 2017.
of these devices to perform computational tasks. This is all operating at the nanoscale - the devices are typically around tens of nanometres in length,” outlines Dr Sebastian. Research is progressing well, and Dr Sebastian says work has already reached quite an advanced stage. “We have developed a computational memory platform with these new devices that I have talked about,” he continues. “Recently, we have shown how in-memory computing can tackle problems such as solving systems of linear equations as well as unsupervised learning of temporal correlations between event-based data streams. A significant part of the research conducted within Projestor is also aimed at improving the properties of the memory devices such as increasing the precision of computation when using these devices.”
T. Tuma, A. Pantazi, M. Le Gallo, A. Sebastian, E. Eleftheriou, “Stochastic phase-change neurons,” Nature Nanotechnology 11(8), 2016. (Cover)
Dr Abu Sebastian
Dr Abu Sebastian is a Principal Research Staff Member and Master Inventor at IBM Research - Zurich. He is actively researching the area of non-von Neumann computing with the intent of connecting the technological elements with applications such as artificial intelligence. He has published over 150 articles and holds over 40 granted patents.
Artificial intelligence The wider background to this research is the start of the cognitive computing era and the increasing pervasiveness of