Hello everybody,
Do you consider that Deep Learning machines and IA which generate realistic representations are "(Scientific) realism machines" vs "human subjectivity"?
Machine Learning (ML) and Deep Learning (DL) solutions like Generative Adversarial Networks (GAN) generate realistic representations of real objects or subjects (person face like with the now famous https://thispersondoesnotexist.com/).
StyleGAN are based on data and statistics and is closed with a kind of logico-mathematical realism where humans use semantic references, different memory types, etc. to perceive, understand or imagine representations and could be seen as less realistic.
An interesting reference about this question:
Machine Learning and the Future of Realism
Article Machine Learning and the Future of Realism
Laurent Berry ,
"StyleGAN are based on data and statistics and is closed with a kind of logico-mathematical realism where humans use semantic references, different memory types, etc. to perceive, understand or imagine representations and could be seen as less realistic."
I cannot find enough justification for this statement. Lets take a step back and analyse for a moment that the mathematics and hardware are equivalent to the biological components of the NN's. Then stating that semantics comes from the sodium potassium channel of the biological neuron is kind of strange but not in the same manner as ANN's, and that is an asymmetric comparison. This privileged treatment of biological neurons is not justified.
How do the cumulative independent values equate to a representation? I think there are arguments for and against emergent behavior in neurons. if we take a single neuron it has a weight stored in memory, but how does the representation emerges when the relevant neurons are activated in a generative model. The mathematics of NN's still cannot account as a reasonable explanation beyond neurons fire and according to the weights the image comes out. Great so what! does it answer the question?
BTW, great question
Arturo Geigel thank you for your answer.
I found a few publications or communications on Realism Machines of data-realism on the web.
The arguments about the form of "realism productions" made by Artificial Neural Networks could be summarized by saying that:
- artificial devices can produce realistics representations of existing things like human but without the same architecture
- artificial devices can compute (infered) new versions of existing things in a way which mimics human creativity without proof of semantic knowledges or autobiographical memory requirement
- However, Mathematics (like Statistics) are anti-realistic
- And Data are not always objective (so not realistic by default) due to the collect or Analog to Digital Conversion
- Deep Learning seems to work like an artifical intuition if you are ok with Herbert A Simon when he stated that intuition was “nothing more and nothing less than recognition”. Because of this knowledge and recognition capability, experts can respond to new situations very rapidly and usually with considerable accuracy. https://www.youtube.com/watch?v=1UqekPMfNk4
- About 2 to 3% of human population encountered Aphantasia (like a ‘blind imagination’) and can't visualize mental images so machines without a mental image capability could be a potential model of living entity https://www.youtube.com/watch?v=KuWSh4n5AiI and it's possible that we won't know when a machine will be more than a Realism artefact like we don't know that a lot of people have Aphantasia and thing that everybody is like them.
- And, it's not sure that images (body, mental, art etc.) can exist only in a Realism approach and maybe do they need an Anti-Realism approach.
Pradeep Paraman Thank you for your feedback.
Yes, DL is like a revival of Neural Networks and an incredible new kind of Machine Learning.
Laurent Berry ,
My concern is another that takes a more nuanced stance.
If you take an image and store it in a file image.png then you can then map the values of the stored images to a particular pixel on the screen (disregard for now screen resolution). Now for neural networks there are several issues:
- When one is encoding the data there is a transformation and abstraction of the data. Then at the output it is decoded into a meaningful representation
- all inputs to the neural network whether using Hebbian learning, backpropagation, annealing or contrastive divergence fix the weights of the neuron. In this instance the weight is shared among different inputs on the network
Under these two conditions the data has been transformed into a"representation" but we do not know its content or form. We are limited in knowing the nature of the underlying representation. What NN theory tells us is the procedure, but does not provide an explanation of the representation itself.Take for example [1], where does it provide a discussion on how representations are actually stored (besides being characterised as distributed)even for generative networks, there is an assumption that the image is stored on the weights, because it produces an output "representation" of what is stored. Now the question is: was this representation actually stored or reconstructed or recalled? Is the representation a fragment of what is stored or the complete representation? what are the qualities and characteristics of the stored information?
[1] Bengio, Y., Courville, A., & Vincent, P. (2013). Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8), 1798-1828.
Yes, I think about the same point:
- NN can't store things which are not included in a picture (human has experience, memories emotion swith multiple contextual information etc.)
- and they could deduct attributes or informations based on Learning
- but we don't know if "reality" is only what is given (data) or measured as they work like an unidimensional machine (they play only with a "flat representation" of things for me)
It's why people wrote that DL is a kind of "Realism machine" but it's not so simple especially since the natural model id a bit more complexe if we investigate complexity of a synpase and the fact that "Brain Communicates in Analog and Digital Modes", https://news.yale.edu/2006/04/12/brain-communicates-analog-and-digital-modes-simultaneously
Article What are the mechanisms for analogue and digital signalling ...
Article Modulation of intracortical synaptic potentials by presynapt...
Laurent Berry ,
While I think we are on the same page on the problem of representation by NNs, I still find hard to justify the statement
"StyleGAN are based on data and statistics and is closed with a kind of logico-mathematical realism where humans use semantic references, different memory types, etc. to perceive, understand or imagine representations and could be seen as less realistic."
The reason being that the reproduction such as the example you provided are outputs of the NN. I think that there is still much to be researched on whether the outputs represent the internals of the networks representation. I find it really hard to make attributions to a black box internal state (the same with humans). Though, this is more of a philosophical rather than scientific stance.
The inputs (dataset of pictures) and the outputs are the same category of things.
The operations and data in the latent space are different. I would say that the latent space is a virtual model and the outputs actualizations as representations of this space as the model.
Interesting article on DL as Learning how to Represent https://www.bbvadata.com/not-deep-learning-learning-represent/
Article Representation Learning: A Review and New Perspectives
- Badges
- Science topic