Il deep learninged una nuova generazione di AI
Presenter: Simone Scardapane
Data Driven Innovation Conference, 24 Febbraio 2017
Il paradosso di Moravec
"In the 60s, Marvin Minsky assigned a couple of undergrads [to program] a computer to use a camera to identify objects in a scene.
He figured they'd have the problem solved by the end of the summer.
Half a century later, we're still working on it."
https://xkcd.com/1425/ (2014)
Non solo riconoscere: descrivere
Download TensorFlow Code
Non solo descrivere: inventare!
An introduction to Generative Adversarial Networks
Inventare è semplice…
Una rete "generativa" crea un'immagine verosimile a partireda rumore.Una seconda rete cerca di discriminare fra immagini reali edimmagini sintetiche.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A. andBengio, Y., 2014. Generative adversarial nets. In Advances in Neural Information ProcessingSystems (pp. 2672-2680).
Immagini, musica…
L’intelligenza artificiale che scrive musica come i Beatles e Duke Ellington[La Stampa, 24/09/2016]
« Scientists at SONY CSL Research Lab have created the first-ever entire songs composed by Artificial Intelligence: "Daddy's Car" and "Mister Shadow".
The two songs are excerpts of albums composed by Artificial Intelligence to be released in 2017. »
… giochi!
How AlphaGo Mastered the Game of Go with Deep Neural Networks
Un passo indietro: reti neurali artificiali
Ogni connessione è un parametro: adattandole in base ai dati, possiamo "apprendere" dagli errori.
Cosa ci dice la biologia?
Un elemento essenziale:strati multipli di elaborazione
Urbanski, M., Coubard, O. A., & Bourlon, C. (2014). Visualizing the blind brain: brain imaging of visual field defects from early recovery to rehabilitation techniques. Frontiers in integrative neuroscience, 8.
Inception
Gooing Deeper Into Convolutions [Google Research Blog]
Ragione 3: software!
Creazione di un modello in Keras:
model = Sequential() model.add(Dense(20, input_dim=16, init='uniform', activation='relu'))model.add(Dense(1, init='uniform', activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam')model.fit(X, Y, nb_epoch=50, batch_size=100)
Allenamento:
La rivoluzione del ML
“People worry that computers will get too smart and take over the world, but the real
problem is that they're too stupid and they've already taken over the world.”
[Pedro Domingos, The Master Algorithm]
Machine Bias
Machine Bias[ProPublica]
Equal Opportunity
Equality of Opportunity in Machine Learning[Google Research Blog]
Machine Trust
A known urban legend: neural networks "learns" to recognize tanks because all training data from one class was taken on sunny days.
"Spiegare" un classificatore
Ribeiro et al. (2016): "Why should I trust you?"Lipton (2016): "The Mythos of Model Interpretability"
Adversarial ML
Attacking machine learning with adversarial examples[OpenAI blog]
Mancanza di senso comune
Lake, B.M., Ullman, T.D., Tenenbaum, J.B. and Gershman, S.J., 2016. Building machines that learn and think like people. arXiv preprint arXiv:1604.00289.
Mancanza di senso comune (2)
Movie written by algorithm turns out to be hilarious and intense[ArsTechnica, 06/09/2016]
Imparare il senso comune
Lerer, A., Gross, S. and Fergus, R., 2016. Learning Physical Intuition of Block Towers by Example. arXiv preprint arXiv:1603.01312.
Privacy
Narayanan, A. and Shmatikov, V., 2008. Robust de-anonymization of large sparse datasets. In 2008 IEEE Symposium on Security and Privacy, 2008 (pp. 111-125). IEEE.
Privacy Concerns Put the Kibosh on the Netflix Prize [Mashable]
Approfondimenti
Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J. and Mané, D., 2016. Concrete problems in AI safety. arXiv preprintarXiv:1606.06565.
Crawford, K. and Calo, R., 2016. There is a blind spot in AI research. Nature, 538(7625), p. 311.