Nowadays, Machine Learning (ML) empowers many different consumer applications, where Artificial Intelligence (AI) is well supported by statistical models and mathematical algorithms that allow computer systems to accurately perform specific tasks. Deep learning methods have dramatically enhanced classification and recognition operations by exploiting a general-purpose learning procedure in a multiple layer’s architecture. However, this technology is not able to perform continual learning, as it suffers from catastrophic forgetting. This means that artificial neural networks (ANN) forget previous stored information when new data is learnt.
During my research, I have introduced bio-inspired algorithms that rely on unsupervised learning in order to improve the lifelong learning performance of ANNs (that perform supervised learning). STDP (Spike Timing Dependent Plasticity) is a protocol used in neuromorphic engineering able to accomplish recognition tasks exploiting neural plasticity, that means, modulating the strength of the synaptic connections (potentiation or depression). The combination of both approaches allows to create a novel network that stores previous information on a stable way, while can get new data (plasticity) by changing the relative synaptic weights following unsupervised learning.
The developed network has several advantages compared to standard deep learning algorithms: (i) it can be implemented in hardware by using Phase Change Memory (PCM) devices, so exploiting in-memory computing, (ii) it is energy efficient, (iii) it can recognize up to 30% not trained digits in MNIST dataset and up to 20% classes in CIFAR10, (iv) it is possible to carry out a fully digital design of the network by means of programming standard HW/SW System On Chips (SoC).
This work has been developed in different steps that can be appreciated in a journal and in a conference papers (which are specified in the following sections of this document). In addition, I have participated at the Demo Session of 2019 Symposium of VLSI Technology and Circuits, where I have explained the complete operating mode of the system and all the technical details. For this demonstration, the network has been implemented in a SoC of Xilinx with the purpose of establishing a friendly human-computer interaction.
In addition, I have collaborated at the development of other project (which is under review for the 2019 International Conference on Electron Devices) related to the introduction of biological schemes like homeostasis in neuromorphic circuits. Moreover, we have deepened the use of bio-inspired homeostasis in the framework of reinforcement learning exploiting PCM devices. We demonstrate that PCM-homeostasis is a key point in the so called recurrent neural networks (RNNs) because it can provide a self-controlled internal state by the simple implementation of homeostasis via multi-resistive PCM devices. Thus, the theorization and consequent experimental demonstration of this brain-inspired homeostatic RNN has given us the possibility to achieve solutions of reinforcement learning tasks like maze navigation.
During these two years of PhD, I have acquired a solid background in deep learning algorithms, neuromorphic networks, brain inspired techniques and industrial tools as SoC programming. In the last third year, my aim is to make a step forward and continue improving the state of the art of machine learning. The goal is to design an integrated circuit using Cadence Virtuoso, capable of performing inference operations using an array of integrated PCM devices. The project includes also the design of a specific board capable of storing/writing/reading the correct testing data and results. Since July 2019 I have been working on this project in collaboration with other members of the Electron Devices group.