In general, I think the influence of neuroscience on neural network research has been subtle and perhaps underrated. For example, modern day convolutional neural networks Havea lineage going back to Fukushima's neocognitron, which was heavily inspired by Hubel an Wiesel's simple/complex cell model of visual cortex based on single cell electrophysiology.
2. https://en.wikipedia.org/wiki/Hierarchical_temporal_memory
Sure, actually it seems highly unlikely to me that they were designed without considering at all how they brain works. Let me clarify what I meant: given the definition of how a Boltzmann machine infers (i.e. how to determine which neurons are on or off), the training algorithm can be derived from purely mathematical considerations.
If you google "SPW neuroscience", you will find many resources.
https://en.wikipedia.org/wiki/Sharp_waves_and_ripples
Brain science is highly interdisciplinary. Each expert should spend time to familiarize themselves with the various approaches used by other experts. In particular, machine learning theorists should spend time to understand the latest results (their implications, at least) in neuroscience and neuron biophysics. At some point, an ML theorist is not too different than a biophysics theorist. They just use different abstractions. This is why I am particularly interested in Google's and Hinton's work. That work seems to be subtly motivated by neuroscience and natural science...
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2039924/
http://neuralcircuits.uwm.edu/sharp-wave-ripples/
In summary, the hippocampus will organize neurons into clumps which encode spatial dimension. As a rat runs around throughout a track, certain groups of neurons fire[1] depending on where the rat is located. Before the task even starts, hippocampal neurons in a rat actually fire in quick succession (e.g. ABCDEFG). When the task is completed, the same pattern fires in the reverse order (e.g. GFEDCBA). Scientists believe that these SPWs play an important role in memory consolidation.
I have always thought of this as "the rat is thinking about where it's going to run to". So I started learning machine learning and landed on things like RBMs. I was so excited because RBMs are reminiscent of this actual phenomenon that occurs in the brain. So, I asked Hinton about it because this is no mere coincidence, and he told me that he had a theory that required reverse order firing during sleep state.
If you get a ton of new memories and patterns during wake, sleep may be responsible for not only pruning less important memories, but strengthening more important memories. For example, sleep might be asking the brain for a ranked list of memories, and then strengthens the memories above some threshold, and prunes the rest (just an idea).
[1] Synchronized firing, see Kuramoto coupled differential equations: https://en.wikipedia.org/wiki/Kuramoto_model
Maybe jayajay needs a nap!8-))
Edit: SPWs, Sharp Waves, SWR, etc. are just labels for sharp waves. I can't believe I am getting attacked for using jargon. The names don't matter. The physics and science matters. Richard Feynman.
Edit 2: The link I sent uses the same jargon that I use. So, honestly mate, what is your deal here? Are you trolling?
You used a term (SPW) I'm not familiar with without defining the term, the three articles you referenced did not include the term, in the end we find the term is not an acronym, I have no knowledge of your training or background and my background is different from yours.
I was intrigued enough to briefly search for SPW but seeing nothing clearly enlightening, I asked for clarification.
My reference to your needing a nap is because you didn't initially define the term SPW, since most reading your post were likely unfamiliar with. No harm was intended. Not everyone knows what you know or thinks about what you think about, especially technical jargon.
It is true that the article you posted later in response to my queries does indeed contain references to the term "SPW". Thank you for providing that clarifying link.
Unfortunately the terminology and origin of the term "SPW" (or "SPW-R")remains unknown to me, despite the second article using it freely. The article says, for example:
"... John O'Keefe investigated SPW-Rs in more detail in 1978 while studying the spatial memory of rat..."
Queries on Google seem to indicate that the term "SPW-R" refers to the same concept as the acronym "SWR" (a true acronym), but I'll never know, being too impatient to consider the matter further.
Apparently someone else found your posts to be dissatisfactory.
> Anyways, I would appreciate if you removed the downvotes -- there is no good reason you have downvoted my responses to you.
As for this, obviously somebody thought your responses were inappropriate. I agree with them, and would have done the same if I had the necessary reputation.
Quoting the HN Guidelines: > Please resist commenting about being downvoted. It never does any good, and it makes boring reading.
> 18.2
Because the negative phase involves drawing samples from the model’s distri- bution, we can think of it as finding points that the model believes in strongly. Because the negative phase acts to reduce the probability of those points, they are generally considered to represent the model’s incorrect beliefs about the world. They are frequently referred to in the literature as “hallucinations” or “fantasy particles.” In fact, the negative phase has been proposed as a possible explanation for dreaming in humans and other animals (Crick and Mitchison, 1983), the idea being that the brain maintains a probabilistic model of the world and follows the gradient of log p ̃ while experiencing real events while awake and follows the negative gradient of log p ̃ to minimize log Z while sleeping and experiencing events sampled from the current model. This view explains much of the language used to describe algorithms with a positive and negative phase, but it has not been proven to be correct with neuroscientific experiments. In machine learning models, it is usually necessary to use the positive and negative phase simultaneously, rather than in separate time periods of wakefulness and REM sleep. As we will see in Sec. 19.5, other machine learning algorithms draw samples from the model distribution for other purposes and such algorithms could also provide an account for the function of dream sleep.
> 19.5.1 Wake-Sleep
One of the main difficulties with training a model to infer h from v is that we do not have a supervised training set with which to train the model. Given a v,we do not know the appropriate h. The mapping from v to h depends on the choice of model family, and evolves throughout the learning process as θ changes. The wake-sleep algorithm (Hinton et al., 1995b; Frey et al., 1996) resolves this problem by drawing samples of both h and v from the model distribution. For example, in a directed model, this can be done cheaply by performing ancestral sampling beginning at h and ending at v. The inference network can then be trained to perform the reverse mapping: predicting which h caused the present v. The main drawback to this approach is that we will only be able to train the inference network on values of v that have high probability under the model. Early in learning, the model distribution will not resemble the data distribution, so the inference network will not have an opportunity to learn on samples that resemble data.
Another possible explanation for biological dreaming is that it is providing samples from p(h,v) which can be used to train an inference network to predict h given v. In some senses, this explanation is more satisfying than the partition function explanation. Monte Carlo algorithms generally do not perform well if they are run using only the positive phase of the gradient for several steps then with only the negative phase of the gradient for several steps. Human beings and animals are usually awake for several consecutive hours then asleep for several consecutive hours. It is not readily apparent how this schedule could support Monte Carlo training of an undirected model. Learning algorithms based on maximizing L can be run with prolonged periods of improving q and prolonged periods of improving θ, however. If the role of biological dreaming is to train networks for predicting q, then this explains how animals are able to remain awake for several hours (the longer they are awake, the greater the gap between L and log p(v), but L will remain a lower bound) and to remain asleep for several hours (the generative model itself is not modified during sleep) without damaging their internal models. Of course, these ideas are purely speculative, and there is no hard evidence to suggest that dreaming accomplishes either of these goals. Dreaming may also serve reinforcement learning rather than probabilistic modeling, by sampling synthetic experiences from the animal’s transition model, on which to train the animal’s policy. Or sleep may serve some other purpose not yet anticipated by the machine learning community.
The training algorithm for Boltzmann machines has two phases:
1. (Awake phase) Feed the network training data, and increase the weights between neurons that tend to get activated simultaneously in this phase.
2. (Dreaming phase) Let the network run without any training input, and decrease the weights between neurons that tend to get activated simultaneously in this phase.
This training algorithm wasn’t designed to emulate how human brains work, it just sort of falls out of the math. And yet (1) resembles the Hebbian principle from neuroscience (neurons that fire together, wire together), and (2) resembles the process described in this article. Actually I suspect something stronger than what the article claims based on this: that the connections that get pruned are the ones connecting neurons that fire together while dreaming. It would be nice to have someone who knows stuff about neuroscience comment on this.