Georgia Institute of Technology
Processing in memory architecture for neural networks and on-chip learning acceleration
Abstract
dc:description.abstractThese days, the demand for efficient and adaptive neural network models has grown significantly. Two primary approaches to meet this demand are on-server learning and on-chip learning. On-server learning offers high-precision computing but consumes energy in server communication and raises privacy concerns. On the other hand, on-chip learning quickly adapts to dynamic environments and efficiently uses personal data, reducing the need for extensive datasets. However, on-chip learning has resource overhead. This research focus on applications where on-chip learning is useful or necessary. Real-time learning with limited resources is a common theme in these scenarios, which requires high performance for adapting to environments. However, traditional general-purpose processors can be slow or resource-intensive, making them unsuitable. Furthermore, real-time learning often lacks associated labels for input data, necessitating unsupervised learning. To overcome these challenges, we propose the use of Processing-in-Memory (PIM) architecture. PIM architecture enables computation within the memory itself, reducing data movement, lowering energy consumption, and improving performance. On-chip learning allows for the local training of machine learning models, mitigating the issues that come with server-side training. Combining PIM architecture with on-chip learning, we aim to provide an efficient solution for machine learning acceleration on resource-limited devices. First, we a propose mixed-signal spike accumulation method to eliminate the need for ADCs in PIM architecture. We also employ the Spike-Time-Dependent Plasticity (STDP) algorithm for on-chip learning. The proposed method allows high performance while keeping low power and area footprint. Subsequently, we address larger classification problems using a convolutional spiking neural network and resolve STDP inefficiency and biasing issues in PIM-based ConvSNN accelerators. Lastly, we extend our target application scope to optimization problems. Unsupervised learning algorithms and a variant of the recurrent stochastic neural network architecture is proposed to solve the optimization problem.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kim, Daehyun
- Advisor dc:contributor.advisor
-
- Mukhopadhyay, Saibal
- Committee members dc:contributor.committeemember
-
- Yu, Shimeng
- Krishna, Tushar
- Sathe, Visvesh
- Kumar, Satish
Subjects
dc:subject × 6Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/72663