Random feedback weights support learning in deep neural networks

By Timothy P. Lillicrap, Daniel Cownden, Douglas B. Tweed, Colin J. Akerman. 2014

186 citations recorded by the source corpus.

The brain processes information through many layers of neurons. This deep architecture is representationally powerful, but it complicates learning by making it hard to identify the responsible neurons when a mistake is made. In machine learning, the backpropagation algorithm assigns blame to a neuron by computing exactly how it contributed to an error. To do this, it multiplies error signals by matrices consisting of all the synaptic weights on the neuron's axon and farther downstream. This operation requires a precisely choreographed transport of synaptic weight information, which is thought to be impossible in the brain. Here we present a surprisingly simple algorithm for deep learning, which assigns blame by multiplying error signals by random synaptic weights. We show that a network can learn to extract useful information from signals sent through these random feedback connections. In essence, the network learns to learn. We demonstrate that this new mechanism performs as quickly and accurately as backpropagation on a variety of problems and describe the principles which underlie its function. Our demonstration provides a plausible basis for how a neuron can be adapted using error signals generated at distal locations in the brain, and thus dispels long-held assumptions about the algorithmic constraints on learning in neural circuits.

Subfields: Artificial Intelligence, Cognitive Neuroscience, Electrical and Electronic Engineering, Optimization & Training, Neuroscience

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Cognitive Neuroscience

In Electrical and Electronic Engineering

In Optimization & Training

In Neuroscience

Citations in the corpus

Deep Information Propagation (2016, neutral): “This approximation is similar in spirit to the vanilla mean ﬁeld approximation and is reminiscent of work on feedback alignment (Lillicrap et al., 2014).”
Direct Feedback Alignment Provides Learning in Deep Neural Networks (2016, background): “For FA the feedback weights are fixed, but if the forward weights are adapted, they will approximately align with the pseudoinverse of the feedback weights in order to make the feedback useful [9].”
Toward an Integration of Deep Learning and Neuroscience (2016, background): “…Recirculation (O’Reilly, 1996), Contrastive Hebbian Learning (Xie and Seung, 2003), random feedback weights together with synaptic homeostasis (Lillicrap et al., 2014; Liao et al., 2015), spike timing dependent plasticity (STDP) with iterative inference and target propagation (Bengio et al.,…”
Learning in the Machine: Random Backpropagation and the Deep Learning Channel (2016, background): “However, recent simulations [24] surprisingly indicate that such symmetry may not be required after all, and that in fact backpropagation works more or less as well when random weights are used to backpropagate the errors.”
Equilibrium Propagation: Bridging the Gap between Energy-Based Models and Backpropagation (2016, background): “Also, Lillicrap et al. (2014) show that the backpropagation algorithm for feedforward nets also works when the feedback weights are random, and that in this case the feedforward weight tend to ’align’ with the feedback weights.”
Assessing the Scalability of Biologically-Motivated Deep Learning Algorithms and Architectures (2018, background): “It is once again natural to wonder whether some approximation of BP might underlie learning in the brain [22, 5].”
Backpropagation and the brain (2020, background): “In this sense, the circuit learns to learn, a phenomenon common to many proposed approximations for backprop 34,90,92,93,101 .”
Training Neural Networks with Local Error Signals (2019, background): “The L pred loss can be made more biologically plausible by using feedback alignment (Lillicrap et al., 2014) to transport the prediction error back to the hidden layer.”
Towards a Biologically Plausible Backprop (2016, background): “rror between two successive layers of rectifying (ReLU) units, suggesting that symmetry may arise as the result of an additional objective function making successive layers form an autoencoder. Also, Lillicrap et al. (2014) show that the backpropagation algorithm for feedforward nets also works when”
Decoupled Greedy Learning of CNNs (2019, background): “One class of alternatives to standard back-propagation aims to avoid its biologically implausible aspects, most notably the weight transport problem (Bartunov et al., 2018; Nø kland, 2016; Lillicrap et al., 2014; Lee et al., 2014; Ororbia et al., 2018; Ororbia & Mali, 2019).”
Synaptic Plasticity Dynamics for Deep Continuous Local Learning (DECOLLE) (2018, background): “An increasing body of work is showing that approximations to the backpropagated errors in SNNs can allow local learning, for example in feedback alignment (Lillicrap et al., 2014).”
Difference Target Propagation (2014, background): “Actually, if f i and g i are linear mappings and g i has a random matrix, difference target propagation is equivalent to feedback alignment (Lillicrap et al., 2014) which works well on many datasets.”
Unsupervised learning by competing hidden units (2018, background): “If those small weights increase and approach the maximal weight, however, they begin to contribute strongly to the normalization condition [4].”
Training Spiking Neural Networks Using Lessons From Deep Learning (2021, methodology): “While this does not fully solve the spatial credit assignment problem, it quells the weight transport problem [138], which is specifically concerned with a weight update in one layer depending upon the weights of far-away layers.”