Is back-prop biologically plausible?

Question

One of the common criticisms of Deep Learning is it's training algorithms, back-propagation of error (back-prop), has no biologically plausible implementation, despite evidence of something like it occurring in the brain. The default implementation is considered biologically implausible due to it's reliance on bi-directional synapses. However, there's been a lot of publications published showing implementations that should be plausible. Specifically, from section 2.2.2 of the review "Towards an integration of deep learning and neuroscience" and some discussion I've had:

1. Random synaptic feedback weights support error backpropagation for deep learning
2. Direct Feedback Alignment (related to the previous random weights idea) and it's inclusion in the Superspike algorithm
3. Deep learning with segregated dendrites
4. Spike timing dependent plasticity (STDP) with iterative inference and target propagation which is based off Difference Target Propagation
5. Kickback from "Kickback cuts Backprop's red-tape: Biologically plausible credit assignment in neural networks"
6. Generalized Recirculation from "Biologically Plausible Error-driven Learning using Local Activation Differences: The Generalized Recirculation Algorithm"
7. Contrastive Hebbian Learning from "Equivalence of backpropagation and contrastive Hebbian learning in a layered network"
8. Complex neurons from "Supervised and unsupervised learning with two sites of synaptic integration"
9. Using an attention mechanism from "A Biologically Plausible Learning Rule for Deep Learning in the Brain"
10. Layer-specific target from "GAIT-prop: A biologically plausible learning rulederived from backpropagation of error"

Are any of these implementable in spiking neurons, able to scale to more than two layers and have been shown to work on normal Deep Learning benchmarks such as MNIST or ImageNet?

Answer 1

Biological Plausibility of Back-Prop

No, the algorithm of back-prop (BP) isn't biologically plausible. However, there are other means which involve propagating the error through multiple layers of neurons in a feed-forward network which are biologically plausible. But before we evaluate these substitutions, let's review why back-prop isn't biologically plausible [1]:

1. Weight Transport Problem

BP uses the same weights for forward-pass and backward error propagation. But synapses are uni-directional.

2. Derivative Transport Problem

The derivative of each neuron is used to modulate the error signal in BP. Additionally, the derivative is propagated through each layer. How this derivative calculated by neurons and then propagated, is unclear.

3. Linear Feedback Problem

The BP feedback path is linear, but neurons aren't linear.

4. Spiking Problem

Neurons spike. BP is defined for rates.

5. Timing Problem

The BP error signal is expected to propagate instantaneously. Calculations and signal transmission in the brain do not happen instantaneously.

6. Target Problem

Most applications of BP rely on many examples with given labels, but where do those labels come from?

Replacements for Back-Prop

Most replacements for back-prop solve a subset of these problems, however none of them see to solve all of them. In "Spiking Deep Neural Networks: Engineered and Biological Approaches to Object Recognition", a method is constructed using the NEF that solves all problems except for the Target Problem. This means there are no replacements for BP that work for all cases where BP is applied, which is to be expected given this is what usually happens when biological plausibility is emphasized.

Beyond Back-Prop

However, it should be noted that there is more than back-propagation through fully-connected networks involved in Deep Learning. Specifically, the are a number of different architectures used, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs).

• CNNs aren't biologically plausible due to their reliance on weights being exactly equal across multiple locations.
• RNNs rely on Back-Propagation Through Time (BPTT), of which there is currently no biologically plausible analogue [2].

This confusion between Deep Learning, neural network architectures and back-prop is reason that Yann LeCunn is trying to get people to adopt the term "Differentiable Computing" instead of "Deep Learning". At which point, the question must shift from "Is Deep Learning biologically plausible?" into "What aspects of Differentiable Computing are biologically plausible?" Which is much harder to answer, but at least we have a better question than when we started!

[1] Summarized from Eric Hunsberger's PhD Thesis "Spiking Deep Neural Networks: Engineered and Biological Approaches to Object Recognition"

[2] That being said, there are other means of learning (supposedly) biologically plausible recurrent neural network weights (FORCE training, Conceptors), however they don't resemble BPTT in any way. The discussion of these methods is outside the scope of this question.

Answer 2

I don't know much about this, but here goes anyway.

I heard that the reason backprop isnt biologically plausible is that it requires global control/coordination for the propagation of the gradients. (verification of this would be nice...)

The decoupled neural interface seems to solve this problem by making gradient propagation local (using a finite difference approximation to break the dependency). So could be biologically plausible, while keeping much of the flavour of backprop.

Also, in my mind the bi-directional argument isnt much of an issue. As you can just have another neuron/path doing the backward step/propagation?

Answer 3

While a backpropagating action potential can presumably cause changes in the weight of the presynaptic connections, there is no simple mechanism for an error signal to propagate through multiple layers of neurons, as in the computer backpropagation algorithm. However, simple linear topologies have shown that effective computation is possible through signal backpropagation in this biological sense.

From wikipedia https://en.wikipedia.org/wiki/Neural_backpropagation

It is a matter of time and more precise research. I am sure that this is the main clue