Towards deep learning with segregated dendrites

Abstract

Deep learning has led to significant advances in artificial intelligence, in part, by adopting strategies motivated by neurophysiology. However, it is unclear whether deep learning could occur in the real brain. Here, we show that a deep learning algorithm that utilizes multi-compartment neurons might help us to understand how the neocortex optimizes cost functions. Like neocortical pyramidal neurons, neurons in our model receive sensory information and higher-order feedback in electrotonically segregated compartments. Thanks to this segregation, neurons in different layers of the network can coordinate synaptic weight updates. As a result, the network learns to categorize images better than a single layer network. Furthermore, we show that our algorithm takes advantage of multilayer architectures to identify useful higher-order representations—the hallmark of deep learning. This work demonstrates that deep learning can be achieved using segregated dendritic compartments, which may help to explain the morphology of neocortical pyramidal neurons.

Bridging Biological and Artificial Neural Networks with Emerging Neuromorphic Devices: Fundamentals, Progress, and Challenges

“…Recently, several researchers have proposed alternatives to the well-adopted backpropagation algorithm in ANN training based on this mechanism. [266,274] Using multi-terminal Ag-based CBRAM memristor, the extra terminal with electrical bias is shown to mimic the use of external variables to modulate its plasticity and can be wired into more complicated circuits to potentially mimic such control mechanisms in the brain in the future. [275]…”

Beyond the feedforward sweep: feedback computations in the visual cortex

“…Here again, there has been significant progress toward algorithms that can assign and propagate credits in more biologically palatable forms. 137,141,142 Another widely successful approach to tuning weights is via reinforcement learning. 143 Reinforcement learning algorithms have demonstrated seemingly magical performance in tasks, such as learning how to play games like chess, Go, or different types of video games, even beating world champions.…”

SuperSpike: Supervised Learning in Multilayer Spiking Neural Networks

“…While some of these studies have been performed using spiking networks, they still use effectively a rate-based approach in which a given input activity vector is interpreted as the firing rate of a set of input neurons (Eliasmith et al., 2012; Diehl & Cook, 2015; Guergiuev et al., 2016; Neftci et al., 2016; Mesnard, Gerstner, & Brea, 2016). While this approach is appealing because it can often be related directly to equivalent rate-based models with stationary neuronal transfer functions, it also largely ignores the idea that individual spike timing may carry additional information that could be crucial for efficient coding (Thalmeier, Uhlmann, Kappen, & Memmesheimer, 2016; Denève & Machens, 2016; Abbott, DePasquale, & Memmesheimer, 2016; Brendel, Bourdoukan, Vertechi, Machens, & Denéve, 2017) and fast computation (Thorpe, Fize, & Marlot, 1996; Gollisch & Meister, 2008).…”