Part 2 Lecture 5: Uses of neuromorphic chips in neuroscience

  • Motivation: Why neuromorphic chips?
  • SPAUN by Chris Eliasmith
  • Bayesian brain hypothesis
  • Predictive processing (will be skipped probably)

Simulating the human brain

The original use of neuromorphic chips is in neuroscience to simulate the brain/how the brain computes.

Dharmendra Modha (IBM) did a 500 billion neuron simulation in 2012.

Blue Gene/Q sequoia supercomputer, used 8 MegaWatt power and 1.5 million GB, but came nowhere close in capabilities of a brain. Also 1500 times slower than human brain. In the meantime, the human brain operates on 20 Watt. Speeding up to real time would require 12 GigaWatt, that's the total amount of eneregy the whole of the Netherlands uses.

The ability to perform massive brain simulations is the driving force behind development of neuromorphic chips.

Loihi

"Loo-EEH-hie".

  • Doesn't use a traditional chip, but is silicon based; currently research-only.
  • Has online learning on the chip
    • To study plasticity
  • Funding, workshops, SDK, remote access (for this course as well!)
  • Loihi project leader: Mike Davies, see his paper on Brightspace.
  • Example specific neuroscientific project: Imam & Cleland, one-shot learning for recognizing odors based on the human olfactory bulb.

Questionable if they will find it marketable enough to go into widespread production.

Most applications are in robotics rather than neuroscience. Some are also in astronomy for finding exoplanets (recall 'event-based processes'; fire when something changes rather constantly capturing and analyzing full scale pictures).

See overview paper by Mike Davies. See Imam & Cleland paper on one-shot learning for a specific application.

Brain simulations

Human brain project (EU); Brain Initiative (US); are large fundings in developing hardware and generate simulation data for computational neuroscience.

Examples: BrainScaleS, and SpiNNaker.

Spaun project:

  • cognitive architecture including input ("eye") and output ("arm").
  • NEF Neural Engineering Framework (underlies Nengo)
  • Still runs on conventional hardware.
    • Porting to SpiNNaker shows non-trivial challenges
    • Lecture does not mention any challenge in particular

Bayesian Brain hypothesis

Bayesian brain hypothesis: the belief that the brain's information processing approximately adheres to the principles of Bayesian inference and Bayesian updating.

(Note that this assumes the Bayesian coding hypothesis; which just says that processing in the brain uses probability distributions).

For more detail, review TFCA-5-Bayesianism_in_Psychology_Cognitive_Neuroscience.md Levels of explanation:

  • Computational: "brain computation respects laws of probability"
  • Implementational: is this realistic on the level of biological neural networks ("wetware")?
    • We cannot currently measure on this level.

Bridging this gap on process-level/algorithmic level: can probability distributions be mapped on spiking neural networks and can we do inference on them? How can SNNs infer with probability distributions? (I think this is what Johan said; it wasn't very clear) So far, biological plausibility is not maintained at all levels. Most useful so far to consider this question on a "process-level".

So SNNs here are a proof-of-concept model on the process level of what may be happening in the brain. - Show that probability distributions can be encoded using neurons - Show that this allows computing a posterior over this distribution

We need to figure out if this is possible to do. If yes, then we can more meaningfully ask how the brain does this.

Maass: SNNs can represent probability distributions

Wolfgang Maass' research: SNNs can represent probability distributions using sampling methods.

  • LIF neurons realize binary stochastic variables.
    • They can be combined in Winner-takes-all (WTA) circuits to enforce a one-hot encoding of categorical variables.
  • The network structure represents (direct) probabilistic dependencies.
  • Nodes can be clamped to spike/not spike
  • Sampling: $v_k=1$ if node has spiked in a particular timewindow
    • Usual timewindow: 20ms
    • All samples together approximate a PD
  • Arbitrary PDs can be encoded as well, but at exponential cost.
    • But these auxiliary nodes are not that biologically plausible

The open questions are: how can we properly learn these probability distributions (in a biologically plausible way)?

Tossing coins example again

We can use SNNs to encode the hyperparameters of generative models.

We see the same pictures as in the second lecture. Again we have that the relative frequency of two neurons encode a probability. And again the actual rate of the neurons can be an indication of the hyperparameters; i.e. how certain are we of the estimated probability?

We also hypothesized that learning is easier done by pruning synaptic connections rather than making new ones, which is slow (inspired by learning processes in the brain). So (somewhat counter-intuitively) the situation where we are most certain of the encoded probability has less spikes in the same time window.

(Additional benefit is that the learned connections require less spikes and thus less energy; we assume connections are learned because they occur frequently, so this is a very good thing).

Some things are still unclear:

  • What's a suitable learning rule?
    • I.e. ok there is a suitable encoding, but how would we reach such an encoding given training data?
  • Is this biologically plausible?
  • How do we establish prediction error?
  • How do we weigh errors with precision?
  • How do we then process weighted prediction errors?
  • ... and so on.

Predictive processing and neuromorphic systems

(This section was not covered in the lecture, so everything here is my own interpretation and reconstruction)

Q: Why is neuromorphic computation an interesting addition to the predictive processing framework?

One answer is that predictive processing implies that you try to minimize the prediction error, either by taking action or by making a perceptual update. Such an minimization step assumes that the prediction error is available. The question then is: is there constant feedback communication (even when there is no error), or only event-driven feedback?

Well, predictive processing only makes a claim that not-anticipated information is processed (where "not expected" is defined with respect to a stochastic generative model).

We have seen that neuromorphic chips are suitable for event-driven computations, because spikes are discrete events. A spike could indicate a relevant (non-expected) event that should trigger processing.

However, even though there seems to be a natural match here, this match is not trivial, because prediction error is not just the same as "a spike". Spikes only operate on the most basic level, but the prediction error for e.g. vision relies on larger hierarchically organized brain areas.

Predictive processing is a high-level theory about information processing; it is a cognitive theory on the computational level, I think. It's therefore not simply understood by looking at neural networks.

However, SNNs may again be a proof of concept in the attempt to bridge the computational theory of predictive processing with implementations.

Above we saw that SNNs may be useful to model the hyperparameters of generative models as well.