Liquid State Machine (LSM) is an adaptive neural computational model with rich dynamics to process spatio-temporal inputs. These machines are extremely fast in learning because the goal-oriented training is moved to the output layer, unlike conventional recurrent neural networks.
The capability to multiplex at the output layer for multiple tasks makes LSM a powerful intelligent engine. These properties are desirable in several machine learning applications such as speech recognition, anomaly detection, user identification etc. Scalable hardware architectures for spatio-temporal signal processing algorithms like LSMs are energy efficient compared to the software implementations. These designs can also naturally adapt to dierent temporal streams of inputs. Early literature shows few behavioral models of LSM. However, they cannot process real time data either due to their hardware complexity or xed design approach. In this thesis, a scalable digital architecture of an LSM is proposed. A key feature of the architecture is a digital liquid that exploits spatial locality and is capable of processing real time data. The quality of the proposed LSM is analyzed using kernel quality, separation property of the liquid and Lyapunov exponent. When realized using TSMC 65nm technology node, the total power dissipation of the liquid layer, with 60 neurons, is 55.7 mW with an area requirement of 2 mm^2. The proposed model is validated for two benchmark. In the case of an epileptic seizure detection an average accuracy of 84% is observed. For user identification/authentication using gait an average accuracy of 98.65% is achieved.
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Polepalli, Anvesh, "Scalable Digital Architecture of a Liquid State Machine" (2017). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus