Today, we are entering the era of cognitive computing, which holds great promise in deriving intelligence and knowledge from huge volumes of data. In today’s computers based on von Neumann architecture, huge amounts of data need to be shuttled back and forth at high speeds, a task at which this architecture is inefficient.
It is becoming increasingly clear that to build efficient cognitive computers, we need to transition to non-von Neumann architectures in which memory and processing coexist in some form. At IBM Research–Zurich in the Neuromorphic and In-memory Computing Group, we explore various such computing paradigms from in-memory computing to brain-inspired neuromorphic computing. Our research spans from devices and architectures to algorithms and applications.
About the IBM Research–Zurich
The location in Zurich is one of IBM’s 12 global research labs. IBM has maintained a research laboratory in Switzerland since 1956. As the first European branch of IBM Research, the mission of the Zurich Lab, in addition to pursuing cutting-edge research for tomorrow’s information technology, is to cultivate close relationships with academic and industrial partners, be one of the premier places to work for world-class researchers, to promote women in IT and science, and to help drive Europe’s innovation agenda. Download factsheet
Hybrid AI Systems (HAS)
Neither symbolic AI nor neural networks alone has produced the kind of intelligence expressed in human and animal behavior. Why? Each has a long and rich history, but has addressed a relatively narrow aspect of the problem. Symbolic AI focuses on solving cognitive problems, drawing upon the rich framework of symbolic computation to manipulate internal representations in order to perform reasoning and inference. But it suffers from being non-adaptive, lacking the ability to learn from example or by direct observation of the world. Neural networks on the other hand have the ability to learn from data, and derive much of their power from nonlinear function approximation combined with stochastic gradient descent. But intelligence requires more than modeling input-output relationships. Without the richness of symbolic computation, neural nets lack the simple but powerful operations such as variable binding that allow for analogy making and reasoning, which underlie the ability to generalize from few examples.
We approach the problem from a very different perspective, inspired by the brain’s high-dimensional circuits and the unique mathematical properties of high-dimensional spaces. It leads us to a novel information processing architecture that combines the strengths of symbolic AI and neural networks, and yet has novel emergent properties of its own. By combining a small set of basic operations on high-dimensional vectors, we obtain hybrid AI system (HAS) that makes it possible to represent and manipulate data in ways familiar to us from symbolic AI, and to learn from the statistics of data in ways familiar to us from artificial neural networks and deep learning. Further, principles of such HAS allow few-shot learning capabilities, and extremely robust operations against failures, defects, variations, and noise, all of which are complementary to ultra-low energy computation on nanoscale fabrics such as phase-change memory devices. Exciting further research (listed in below table) awaiting in this direction spans high-level algorithmic exploration all the way to efficient hardware design for emerging computational fabrics.
- Robust high-dimensional memory-augmented neural networks, arXiv, 2020
- In-memory hyperdimensional computing, Nature Electronics, 2020.
- A wearable biosensing system with in-sensor adaptive machine learning for hand gesture recognition, Nature Electronics, 2020.
- Hyperdimensional computing: an introduction to computing in distributed representation with high-dimensional random vectors, Cognitive Computation, 2009.
- Background in machine learning (recommended)
- Experience with any deep learning framework such as TensorFlow or PyTorch (recommended)
- VLSI I (recommended)
In-Memory Computing (IMC)
For decades, conventional computers based on the von Neumann architecture have performed computation by repeatedly transferring data between their processing and their memory units, which are physically separated. As computation becomes increasingly data-centric and as the scalability limits in terms of performance and power are being reached, alternative computing paradigms are searched for in which computation and storage are collocated. A fascinating new approach is that of computational memory where the physics of nanoscale memory devices are used to perform certain computational tasks within the memory unit in a non-von Neumann manner. Computational Memory (CM) is finding application in a variety of areas such as machine learning and signal processing. In addition to novel non-volatile memory technologies like PCM and ReRAM, also conventional SRAM has been proposed as underlying storage technology in Computational Memories.
- An SRAM-Based Multibit In-Memory Matrix-Vector Multiplier With a Precision That Scales Linearly in Area, Time, and Power
- Memory devices and applications for in-memory computing
- Deep learning acceleration based on in-memory computing
- General interest in Deep Learning and memory/system design
- VLSI I and VLSI II (recommended)
- Analog Circuit Design (recommended)
Specific requirements for the different projects vary and are generally negotiable.
We are inviting applications from students to conduct their master’s thesis work or an internship project at the IBM Research lab in Zurich on this exciting new topic.
|MA||Face Recognition at Scale||Link to description||HAS||algorithmic design|
|MA||Accelerating Transformers with Computational Memory||Link to description||HAS||Hardware/algorithmic design|
|MA||Developing Efficient Models of Strong AI for Intelligence Quotient (IQ) Test||Link to description||HAS||algorithmic design|
|MA||Lifelong learning challenge||Link to description||HAS||algorithmic/hardware design|
|MA||Machine learning based on optimal transport using in-memory computing'||Link to description||HAS||hardware design|
|MA||Accurate deep learning inference using computational memory||Project description and application||IMC||algorithmic design|
|MA||ADC design for computational memory||Digital-to-Analog converters (DACs) and Analog-to-Digital converters (ADCs) are extensively employed in Computational Memory (CM) to handle the crossing between the digital and analog domain, in which computationally expensive tasks, like Matrix-Vector Multiplications (MVM), are carried out with O(1) complexity. Each conversion costs a certain amount of energy and its precision can only be guaranteed up to the Effective Number of Bits (ENOB) of the employed data converter.
The research focus will be on understanding the system level requirements on ADC and DAC for optimal performance of Deep Neural Network inference using CM. Furthermore, the effects of noise, non-linearity and manufacturing tolerances shall be examined and counter measurements, like for example periodic digital ADC recalibration and digital post processing, shall be evaluated with regards to effectivity and energy costs.
|IMC||analog circuit design|
|SA||Testing of a computational memory chip||This project is about building a Microprocessor/FPGA-based test platform around a novel IMC chip. After commissioning the chip, DL workload tests can be run to characterize its throughput and energy-efficiency.||IMC|| PCB Design |
|MA/SA|| Neural Network Training on a
computational memory chip
|TBA||IMC|| algorithm/system design|
analog circuit design
- Thesis will be at IBM Zurich in Rüschlikon
- Hybrid AI Systems (HAS) projects
- Contact (at ETH Zurich): Dr. Frank K. Gurkaynak and Michael Hersche
- Contact (at IBM): Dr. Abu Sebastian
- Contact (at IBM): Dr. Abbas Rahimi
- Professor: Luca Benini
- In-Memory Computing (IMC) projects
- Contact (at IBM/ETH Zurich): Riduan Khaddam-Aljameh
- Contact (at IBM): Dr. Abu Sebastian
- Professor: Luca Benini