Research

• Neuromorphics

• EFRI BRAID: Using Proto-Object Based Saliency Inspired by Cortical Local Circuits to Limit the Hypothesis Space for Deep Learning Models ——–

Deep learning has achieved impressive performance in many tasks, which is driven by the capacity for backpropagation to “assign credit” to a vast array of parameters. Typical networks have immensely complex computational graphs, with many options to assign credit for every computation. A large number of options comes with the benefits of being very flexible in learning, but also with costs of large energy consumption and many needed examples for learning. A selection of important (salient) features will cause inductive biases in learning, but such biases, when appropriately conditioned, can be optimally selected, as is done in biology via evolution or development. For our project, the selection mechanism will be inspired by biology or learned, and will be instantiated in software and hardware. The process of selection is akin to the attention mechanisms of mammals. We previously developed state-of-the art models, based on neurophysiology, of bottom-up and top-down attention and suggested how perceptual organization can reshape and focus attention. We showed how such mechanisms of attention which can predict human behaviors can be implemented using local circuits in the cortex and in neuromorphic hardware. We propose to construct a hybrid architecture, where local circuits implement a bottom-up attention, or saliency module that provides a “gate” for selecting features for a global learning network with a convolutional architecture. The saliency module will decrease the number of features considered for inference and for learning by including a learned prior of what features are important. We hypothesize that after determining and implementing optimal attentional mechanisms for a set of tasks/input statistics, they will substantially reduce power requirement for both inference and learning, as well as allowing learning with considerably fewer examples than traditional methods. Such a model can also help answer the question of why some visual cortex neurons have their properties explained by convolutional neural networks while others mimic saliency models, and why biology learns with few examples. We can also answer determine optimal learning architecture for hardware and benchmark them against existing systems.

• Algorithmic and Hardware Co-Design of Event-based Local Synaptic Plasticity
  ——– The human brain expends a mere 20 watts^[1] in learning and inference, exponentially lower than state-of-the-art large language models (GPT-3 and LaMDA). There is the need to innovate sustainable AI hardware as the 3.4x compute doubling per month has drastically outpaced the Moore’s law, i.e., a 2-year transistor doubling^[2]. Efforts here are geared towards realizing biologically plausible learning rules such as the Hebb’s rule based Spike-Timing-Dependent Plasticity (STDP) algorithmically and corresponding in low-power neuromorphic circuits and systems implemented in mixed analog-digital VLSI. Analog subthreshold current-mode computation offers convenient realization of addition and multiplication arithmetic. Moreso, current-mode memory elements such as dynamic current mirrors can be leverage to arrive at low-power learning engines. This has been demonstrated in our recent work presented at NEWCAS. On the Algorithmic front, we have developed LODeNNS, a linearly optimized Dendrocentric Nearest Neighbor STDP suitable for existing large-scale digital neuromorphic platforms such as the Intel Loihi and can be deployed in spatiotemporal signal applications such as object tracking and audio recognition. Future work on this include validation and application on event-based data. On the hardware front, insights from the algorithmic level are being translated in mixed signal VLSI with implications for learning at the edge and on-the-fly, i.e. low-power intelligent sensors.
• VLSI implementation of Synaptic Memory Consolidation for Lifelong Learning —————- Our brains are capable of learning new information quickly and retaining them over long periods - from days to years. This plasticity-rigidity property is lacking in present day Machine Learning models as they are often riddled with catatrosphic forgetting arising from vanishing and exploding gradients. The goal here is implement computational prinicples (the Linear Chain Bicascade model) behind synaptic memory consolidation (SMC) presented by Benna and Fusi^[3] in mixed signal VLSI. Ultimately, deploying an array of synapse equipped with such a property will be useful for realizing energy-efficient hardware for continual learning.

• Neural Interfaces

  • HermEIS: A Parallel Multichannel Electrochemical Impedance Spectrometer for Rapid Neural Electrode Characterization