Current students


INDIRLI FABRIZIOCycle: XXXVI

Section: Computer Science and Engineering
Advisor: SILVANO CRISTINA
Tutor: ALIPPI CESARE

Major Research topic:
Hardware Acceleration of Neural Networks through In-Memory Computing

Abstract:
In recent years, Deep Neural Networks are quickly becoming leading edge solutions in a broad variety of computational tasks, such as image classification and speech recognition. The rapid proliferation of pervasive IoT devices and ubiquitous cognitive computing applications is pushing the industry towards performing Machine Learning inference on edge devices. These embedded platforms pose stringent constraints on power consumption, latency and memory footprint, which has led to the development of new training and mapping techniques for TinyML applications. 
Furthermore, the scaling of performance and efficiency of traditional hardware architectures is being hindered by the slowdown of Moore’s Law; thus, the research focus in this field is shifting towards novel methodologies, such as Heterogeneous platforms and Neuromorphic computing. In this context, hardware accelerators based on FPGA platforms and ASICs (such as Google’s TPU) are emerging as promising solutions. However, these devices often have limited I/O bandwidths, which worsens the negative impact of the transactions between the processing units and the external memory. To overcome this problem, the In-Memory-Computing (IMC) paradigm proposes to perform the calculations in the same cells in which the data is stored, reducing the accesses to the off-chip RAM. This approach can be particularly beneficial for neural networks inference, where the same weights or activations can be reused across several Multiply-And-Accumulate (MAC) operations.  
Further advancements in computational efficiency can be achieved by integrating analogic elements inside IMC accelerators, leveraging novel memristive technologies such as ReRAM or Phase-Changing Memory. These devices, however, pose new challenges because of their lower resolution and higher noise sensitivity, and they require ad-hoc training, quantization and deployment methodologies for optimal results. 
This PhD thesis research aims to develop novel techniques and tools to optimally train and deploy neural networks on Heterogeneous and Neuromorphic hardware accelerators, in collaboration with STMicroelectronics SRA department (in Cornaredo). In particular, the following contributions are proposed:  ;
    ;
  • Selection and study of applicable neuromorphic and/or heterogeneous hardware accelerators for neural networks; 
  • ;
  • Identification of neural networks suitable for low-precision IMC acceleration, and development of ad-hoc quantization techniques and topological transformations to prepare the input models for lowering on the target hardware; 
  • ;
  • Exploration of different scheduling and binding strategies to map the proposed models on the target neuromorphic accelerators; 
  • ;
;
    ;
  • Comparison of the results of different hardware accelerators and of several mapping techniques, in terms of latency, accuracy, precision and power consumption; 
  • ;
  • Design of novel acceleration kernels for specific nodes and prototyping on FPGAs. 
  • ;