Learn by consuming the brains of your enemies

A few people have sent this my way and asked about it:

In a paper published Monday in the journal eNeuro, scientists at the University of California-Los Angeles reported that when they transferred molecules from the brain cells of trained snails to untrained snails, the animals behaved as if they remembered the trained snails’ experiences…

In experiments by Dr. Glanzman and colleagues, when these snails get a little electric shock, they briefly retract their frilly siphons, which they use for expelling waste. A snail that has been shocked before, however, retracts its siphon for much longer than a new snail recruit.

To understand what was happening in their snails, the researchers first extracted all the RNA from the brain cells of trained snails, and injected it into new snails. To their surprise, the new snails kept their siphons wrapped up much longer after a shock, almost as if they’d been trained.

Next, the researchers took the brain cells of trained snails and untrained snails and grew them in the lab. They bathed the untrained neurons in RNA from trained cells, then gave them a shock, and saw that they fired in the same way that trained neurons do. The memory of the trained cells appeared to have been transferred to the untrained ones.

The full paper is here.

Long and short of this is that there is a particular reflex (memory) that changes when they have experienced a lot of shocks. How memory is encoded is a bit debated but one strongly-supported mechanism (especially in these snails) is that there are changes in the amount of particular proteins that are expressed in some neurons. These proteins might make more of one channel or receptor that makes it more or less likely to respond to signals from other neurons. So for instance, when a snail receives its first shock a neuron responds and it withdraws its gills. Over time, each shock builds up more proteins that make the neuron respond more and more. These proteins are built up by the amount of RNA (the “blueprint” for the proteins, if you will) that are located in the vicinity of the neuron that can receive this information. There are a lot of sophisticated mechanisms that determine how and where these RNAs are built and then shipped off to the place in the neuron where they can be of the most use.

This new paper shows that in these snails, you can just dump the RNA on these neurons from someone else and the RNA has already encoded something about the type of protein it will produce. This is not going to work in most situations (I think?) so it is surprising and cool that it does here! But hopefully you can begin to see what is happening and how the memory is transferring. The RNA is now in the cell, it is now marked in a way that will lead it to produce some protein that will change how the cell responds to input, etc, etc.

One of the people who asked me about this asked specifically in relation to AI. Could this be used as a new method of training in Deep Networks somehow? The closest analogy I can think of is if you have two networks with the same architecture that have been trained in the same way (this is evolution). Then you train a little more, maybe on new stimuli or maybe on a new task, or maybe you are doing reinforcement learning and you have a network that predicts a different action-value pair. Now the analogy would be if you chose a few units (neurons) and directly copied the weights from the first network into the second network. Would this work? Would this be useful? I doubt it, but maybe? But see this interesting paper on knowledge distillation that was pointed to me by John O’Malia.


Controversial topics in Neuroscience

Twitter thread here.

  • Do mice use vision much?
    • They have pretty crappy eyesight and their primary mode of exploration seems to be olfactory/whisker-based
  • How much is mouse cortex like primate cortex?
    • Mouse cortex is claimed to be more multimodal than primate cortex which is more specialized
  • “The brain does deep learning”
    • Deep learning units aren’t exactly like neurons, plus we resent the hype that they have been getting
  • Is there postnatal neurogenesis? Is it real or behaviorally relevant?
    • See recent paper saying it doesn’t exist in humans
  • Brain imaging
    • Does what we see correlate with neural activity? Are we able to correct for multiple comparisons correctly? Does anyone actually correct for multiple comparisons correctly?
  • Bayesian brain
    • Do people actually use their priors? Does the brain represent distributions? etc
  • Konrad Kording
    • Can neuroscientists understand a microprocessor? Is reverse correlation irrelevant?
  • Do mice have a PFC
    • It’s so small!
  • STDP: does it actually exist?
    • Not clear that neurons in the brain actually use STDP – often looks like they don’t. Same with LTP/LTD!
  • How useful are connectomes
    • About as useful as a tangled ball of yarn?
  • LTP is the storage mechanism for memories
    • Maybe it’s all stored in the extracellular space, or the neural dynamics, or something else.
  • Are purely descriptive studies okay or should we always search for mechanism
    • Who cares about things that you can see?!


  • Does dopamine have ‘a role’?
    • Should we try to claim some unified goal for dopamine, or is it just a molecule with many different downstream effects depending on the particular situation?
  • Do oscillations (‘alpha waves’, ‘gamma waves’, etc) do anything?
    • Are they just epiphenomenon that are correlated with stuff? Or do they actually have a causative role?

What HASN’T Deep Learning replicated from the brain?

The brain represents the world in particular ways. Here are a few:

1. The visual world on the retina

The retina is thought to whiten images, or transform them so that they always have roughly the same average, maximum and minimum (so that you can see in very bright and very dark environments. This was originally shown very nicely in two papers from Atick and Redlich (1990, 1992). Essentially, you want to smooth the visual scene around each point depending on the noise. You get receptive fields that look something like this:

Or more generally this:

A denoising autoencoder – a network that tries to replicate a corrupted image which smooths locally – has neural representations that look similar:

2. The visual world in first order visual cortex

Similarly, if you want to efficiently represent the visual world (once it is denoised) you want to represent things sparsely or independently. This was shown by Olshausen and Field 1996  and Bell and Sejnowski 1997 and is equivalent to doing ICA on natural images. Note that doing PCA on natural images will give you Fourier components.

If you train a Deep Network on ImageNet (AlexNet example below), the filters on the first layer look similar:

3. The auditory world

The best representation of the auditory world is also efficiently encoded. Lewicki 2002 show that if you run ICA on acoustic data you get filters that look like nearly identical to the sounds neurons respond to (wavelet basis functions).

I have not seen a visualization of the first few layers of a neural network that classifies speech (for instance) but I would guarantee it has features that look like wavelets.

4. Spatial cells

Your sense of place in the world is encoded by a series of grid cells – which are a periodic representation of place – and place cells, which are precise locations in space. Dordek et al 2016 showed that non-negative PCA on place cells will give you grid cells. This is similar to the result that PCA on images gives you Fourier components. Note that Dordek et al also use a single-layer feedforward neural network and show that it has a similar property.

It turns out if you train a Deep recurrent network on network navigation task, you get grid cells (once you have assumed place cells).


What else is left? Olfaction is a mess and doesn’t have a coherent coding principle as far as I can tell (the olfactory space is not clearly defined). Mechanosensation (touch) has been hard to define but Zhao et al 2017 can find first-order touch receptive fields with an autoencoder (like with vision). You can get CPGs (oscillatory movement generators) with recurrent neural networks by training an input signal to be associated with a particular sequence of movements. I’m struggling to think of other internal representations that are well understood.

A long-term principle in neuroscience has been that successive layers of the brain are attempting to decorrelate their responses to produce ever-finer features. Tishby and Zaslavsky 2015 suggest that a similar principle applies to Deep Networks: you have a constrained input output and networks are trying to find the representations that encode the most information between input and output given the limited bandwidth that they have (numbers of layers, numbers of units). It should not be surprising that this entails something like different forms of PCA or ICA or other signal-detection framework.

One of the nice things about Deep Networks is that you do not have to explicitly code for this in order to find these features – they are costless in a way. You can train for a particular task – a visually-driven one, a path-driven one, an acoustic-driven one – and these features will just fall out. Not only will these features fall out, but neurons which are deeper in the pathway will also have similar activity. This is a much harder problem and one in which “run PCA again” or “run ICA again” will not give a good answer to.

What other neural representations have we not yet seen in neural networks?