BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260531T154615EDT-3532PapuTB@132.216.98.100
DTSTAMP:20260531T194615Z
DESCRIPTION:Abstract\n\nArtificial Intelligence (AI) has become profoundly 
 embedded in contemporary life\, with its applications proliferating across
  a wide array of domains. Central to AI are neural networks\, which have m
 arkedly enhanced the capabilities of AI in areas such as computer vision a
 nd natural language processing. As neural networks scale in both size and 
 computational complexity\, the intelligent devices tasked with executing t
 hese networks face growing demands for computational and energy resources 
 to ensure efficient and reliable performance. Consequently\, resource-limi
 ted embedded devices\, such as smartphones\, encounter significant challen
 ges in deploying state-of-the-art AI models. These devices frequently reso
 rt to cloud-based platforms\, which necessitate continuous internet connec
 tivity. This dissertation seeks to address these challenges by reducing th
 e computational complexity of neural networks. Specifically\, it targets t
 he primary source of computational burden and a major contributor to energ
 y consumption in neural networks: high-precision multipliers (e.g.\, 16-bi
 t or 8-bit multipliers). We propose novel implementations of neural networ
 ks that either markedly reduce the bit-width of multipliers (to 4 bits or 
 fewer) or entirely replace them with simpler logic operations (e.g.\, XNOR
  and shift operations). In our initial implementation of neural networks\,
  we present a novel approach for training multi-layer networks utilizing F
 inite State Machines (FSMs). In this approach\, each FSM is interconnected
  with every FSM in both the preceding and subsequent layers. We demonstrat
 e that the FSM-based network can effectively synthesize complex multi-inpu
 t functions\, such as 2D Gabor filters\, and perform non-sequential tasks\
 , such as image classification on stochastic streams\, without the need fo
 r multiplications\, given that FSMs are implemented solely through look-up
  tables. Building on the FSMs' capability to handle binary streams\, we pr
 opose an FSM-based model specifically designed for handling time series da
 ta\, applicable to temporal tasks such as character-level language modelin
 g. In our second implementation\, we introduce an advanced stochastic comp
 uting (SC) representation termed the dynamic sign-magnitude (DSM) stream. 
 This representation is specifically designed to enhance the precision of s
 hort-sequence SC-based multiplication. The DSM framework facilitates the s
 ubstitution of conventional neural network multiplications with more effic
 ient bitwise XNOR operations. By employing DSM\, we achieve a reduction in
  the sequence length for SC-based neural networks by a factor of 64\, whil
 e maintaining accuracy levels comparable to existing methodologies. In our
  third implementation\, we propose an innovative base-2 logarithmic quanti
 zation scheme for neural networks. This scheme quantizes weights into disc
 rete power-of-two values by leveraging information about the network’s wei
 ght distribution. This method allows us to replace computationally intensi
 ve high-precision multipliers with more efficient shift-add operations. Co
 nsequently\, our quantized networks exhibit approximately eight times fewe
 r parameters compared to high-precision networks\, without compromising cl
 assification accuracy. In our latest implementation\, we introduce a novel
  training framework that utilizes quantization techniques to facilitate th
 e conversion between quantized networks and spiking neural networks (SNNs)
 . SNNs are inherently devoid of multiplications\, relying instead on addit
 ion and subtraction. This new framework offers an alternative approach for
  training SNNs. Specifically\, we modify the SNN algorithm and mathematica
 lly demonstrate that after T time steps\, the modified SNN approximates th
 e behavior of a quantized network with T quantization intervals. This allo
 ws for the replacement of any SNN with its corresponding quantized network
  for training purposes and then transfer the parameters from the trained q
 uantized network to the SNN without additional steps.\n
DTSTART:20250324T143000Z
DTEND:20250324T163000Z
LOCATION:Room 603\, McConnell Engineering Building\, CA\, QC\, Montreal\, H
 3A 0E9\, 3480 rue University
SUMMARY:PhD defence of Amir Ardakani – Towards Multiplier-less Implementati
 on of Neural Networks
URL:https://www.mcgill.ca/ece/channels/event/phd-defence-amir-ardakani-towa
 rds-multiplier-less-implementation-neural-networks-364321
END:VEVENT
END:VCALENDAR