From Deep Learning of Disentangled Representations to Higher-level Cognition

One of the main challenges for AI remains unsupervised learning, at which humans are much better than machines, and which we link to another challenge: bringing deep learning to higher-level cognition. We review earlier work on the notion of learning disentangled representations and deep generative models and propose research directions towards learning of high-level abstractions. This follows the ambitious objective of disentangling the underlying causal factors explaining the observed data. We argue that in order to efficiently capture these, a learning agent can acquire information by acting in the world, moving our research from traditional deep generative models of given datasets to that of autonomous learning or unsupervised reinforcement learning. We propose two priors which could be used by an agent acting in its environment in order to help discover such high-level disentangled representations of abstract concepts. The first one is based on the discovery of independently controllable factors, i.e., in jointly learning policies and representations, such that each of these policies can independently control one aspect of the world (a factor of interest) computed by the representation while keeping the other uncontrolled aspects mostly untouched. This idea naturally brings fore the notions of objects (which are controllable), agents (which control objects) and self. The second prior is called the consciousness prior and is based on the hypothesis that our conscious thoughts are low-dimensional objects with a strong predictive or explanatory power (or are very useful for planning). A conscious thought thus selects a few abstract factors (using the attention mechanism which brings these variables to consciousness) and combines them to make a useful statement or prediction. In addition, the concepts brought to consciousness often correspond to words or short phrases and the thought itself can be transformed (in a lossy way) into a brief linguistic expression, like a sentence. Natural language could thus be used as an additional hint about the abstract representations and disentangled factors which humans have discovered to explain their world. Some conscious thoughts also correspond to the kind of small nugget of knowledge (like a fact or a rule) which have been the main building blocks of classical symbolic AI. This, therefore, raises the interesting possibility of addressing some of the objectives of classical symbolic AI focused on higher-level cognition using the deep learning machinery augmented by the architectural elements necessary to implement conscious thinking about disentangled causal factors.

[flamingxombie]

The intuition for why we current speech models can't produce good unconditional samples (see wavenet) is simply mindblowing. Phonemes occupy a small number of bits as compared with the overall signal (~10/s as compared with 16 k/s)!

[johntanchongmin]

Interpolating in abstract space, exactly what Stable Diffusion is doing. This idea is really impactful.

[itaybenou]

Wonderful video. You can't help but admire his approach for what is AI, and the way he manages to convey these concepts. Brilliant!

[siarez]

Who is the gentleman at 1:09:35 asking a question, and bringing up gradual learning?

[tempvariable]

In 12:07, are cognitive states low dimensional if that is the case are they sparse? If they are both sparse and low dimensional it contradicts with what he said in his MSS talk in 2012, where he states high dimensional and sparse is better than low dimensional

[johntanchongmin]

In 1:09:33, there was a question on gradual change in hypothesis space from very few samples - theory revision. I feel like neural nets may be quite ill-suited for fast change of learnt knowledge as the weights take a long time to change by backpropagation. What I believe is necessary, will be to imbue some form of learnable external memory bank on which we draw our knowledge from (in addition to neural nets), so we can just change that knowledge bank and learn new concepts instantly.

[catsaresocute]

In Lexes Podcast just the mention of the example of speech as becoming unrecognicable gibberish (due to the amount of data) but when you seperate the gibberish to get a baisc feel for intonation or sound and speach as vocilisation of cerain tones that humans think of as speach you get a functional way to work it out would have totally suffieced to get the thing

[jonabirdd]

51:00 I like the idea of a two-level system but disagree with the mutual information criterion.

[bingeltube]

Thanks for the interesting talk! Please post the slides as well!

[muckvix]

Anyone has a link to the slides? And come on camera people, it's not a beauty pageant, it's ok if you show slides instead of the speaker's face :)

[silberlinie]

Deepening of learning into a higher cognitive level:
Very good.
What and where are the works, who is working on this approach?

[dr.mikeybee]

Doesn't translation into an abstract space necessitate a loss of information?

[juliocardenas]

The camera work negatively affects a wonderful lecture

[scose]

Sampling rate * bit depth is a big overestimate of the amount of information in speech audio signals - look at the compression ratios that audio codecs can achieve

[zlh]

46:00 re: attention as gating the conscious and unconscious thoughts - can you imagine a machine which can widen and narrow its aperture of attention to accomplish different tasks?

[runvnc]

Sounds right to me. But why do they assume that the traditional neural net and deep learning are the best or only possible fundamental structures and processes for a system with these capabilities of disentangled abstractions working together with granular representations?

[nguyenngocly]

You can turn artificial neural networks inside-out by using fixed dot products (weighted sums) and adjustable (parametric) activation functions. The fixed dot products can be computed very quickly using fast transforms like the FFT. Also the number of overall parameters required is vastly reduced. The dot products of the transform act as statistical summary measures. Ensuring good behavour. See Fast Transform (fixed filter bank) neural networks.
The variance equation for linear combinations of random variables is very useful for understanding dot products in neural networks especially in conjunction with cosine angle.
Also ReLU is a switch. The electricty in your house is a sine wave. Turn on a switch and the output is f(x)=x. Again the same sine wave as the input. Off(x)=0. A ReLU neural network then is a switched composition of dot products. If the switch states are known then there is a linear mapping between the input vector and the output vector which you can check out with various metrics.

[MartinLichtblau]

Humans use fuzzy approaches, while computers use precise numbers. Which one can work in this complex world?

[rahuldeora]

Someone should write a detailed blog explaining stuff in this

[ewfq]

I want to talk with the guy talking about barycentres and wasserstein distance!