Buongiorno, materia estremamente specialistica e davvero difficile da rendere in modo comprensibile a chi non conosce i meccanismi delle reti neurali. don Luca Peyron <dluca.universitari@gmail.com> writes: [...]
https://today.duke.edu/2020/12/accurate-neural-network-computer-vision-witho...
Vincendo la mia rabbia dopo aver letto nel sommario dell'articolo: «New research offers clues to what goes on inside the minds of machines as they learn to see». Minds of machines, learn, see... oh mamma?!? AFAIU i ricercatori propongono di ispezionare la "black box" delle reti neurali inserendo una _sorta_ di breakpoint [1] (lo chiamano modulo di concept whitening) nella rete neurale che consente ai ricercatori non solo di verificare l'ambiente in quell'istante (variabili) ma anche di alterarne il contenuto (decorrelated and normalized) in modo controllato, ottenendo così informazioni su come le singole parti della rete stabiliscono correlazioni e da esse inferiscono risposte. Il succo è che *se* i ricercatori hanno ragione allora potrebbe essere possibile trovare un modo per ottenere l'intepretabilità delle inferenze automatiche eseguite dalle reti neurali, che sarebbe una specie di rivoluzione copernicana dell'AI, per quanto ne capisco io ad oggi. Io, che NON sono esperto in reti neurali, mi domando semplicemente se da un'infinita catena di correlazioni tra dati selezionati, misurati e *annotati* da esseri umani e NON campionati automaticamente "dalla realtà" (perdonatemi la semplificazione) sia oggettivamente possibile ricavarne un nesso di causa... ma sono ignorante appunto. La sintesi giornalistica dice: --8<---------------cut here---------------start------------->8--- The method controls the way information flows through the network. It involves replacing one standard part of a neural network with a new part. The new part constrains only a single neuron in the network to fire in response to a particular concept that humans understand. The concepts could be categories of everyday objects, such as “book” or “bike.” But they could also be general characteristics, such as such as “metal,” “wood,” “cold” or “warm.” By having only one neuron control the information about one concept at a time, it is much easier to understand how the network “thinks.” [...] “Our method revealed a shortcoming in the dataset,” Rudin said. Perhaps if they had included this information in the data, it would have made it clearer whether the model was reasoning correctly. “This example just illustrates why we shouldn’t put blind faith in “black box” models with no clue of what goes on inside them, especially for tricky medical diagnoses,” Rudin said. --8<---------------cut here---------------end--------------->8--- La ricerca pubblicata su Nature, da cui è tratto l'articolo, è questa: https://www.nature.com/articles/s42256-020-00265-z.epdf?sharing_token=rnt3d7... --8<---------------cut here---------------start------------->8--- What does a neural network encode about a concept as we traverse through the layers? Interpretability in machine learning is undoubtedly important, but the calculations of neural networks are very challenging to understand. Attempts to see inside their hidden layers can be misleading, unusable or rely on the latent space to possess properties that it may not have. Here, rather than attempting to analyse a neural network post hoc, we introduce a mechanism, called concept whitening (CW), to alter a given layer of the network to allow us to better understand the computation leading up to that layer. When a concept whitening module is added to a convolutional neural network, the latent space is whitened (that is, decorrelated and normalized) and the axes of the latent space are aligned with known concepts of interest. By experiment, we show that CW can provide us with a much clearer understanding of how the network gradually learns concepts over layers. CW is an alternative to a batch normalization layer in that it normalizes, and also decorrelates (whitens), the latent space. CW can be used in any layer of the network without hurting predictive performance. --8<---------------cut here---------------end--------------->8--- Io non l'ho ancora letta tutta (e farò fatica a interpretarla nella sua complessità). Saluti, Giovanni. [1] https://en.wikipedia.org/wiki/Breakpoint -- Giovanni Biscuolo