Simone, Gianluca e NicolòEntrepreneur and Statistiscian, Engineer, Self-driving car Engineer
"L'hype da AI"
"L'hype da AI"
"L'hype da AI"
Perché AI?Perché ora?
Un po' di storia: Dartmouth, 1956
What is intelligence?“The true sign of intelligence is not knowledge but
imagination”
Albert Einstein
AI: generale vs ristretta
Prima applicazione di AI ristretta:English-Russian translator
Risultati:English:"The spirit is strong, but the flesh is weak"After English - Russian > Russian - English:"The whiskey is strong, but the meat is rotten"
Primo "AI Winter" (1966 ~ 1980s):No potenza di calcolo.
No dati.
No metodi.
Secondo "AI spring":I sistemi esperti
Problemi Expert Systems:Costosi da realizzare
Molto settorializzati
Scarsa capacità di generalizzazione
Un "nuovo" approccio: il Machine Learning
What is Machine Learning?«A computer program is said to learn from
experience E with respect to some class of tasks Tand performance measure P if its performance at
tasks in T, as measured by P, improves withexperience E»
Funziona?"If one could devise a successful chess machine, onewould seem to have penetrated to the core of human
intellectual endeavor"
Allen Newell, 1958
Deep Blue vs Garry Kasparov,1997
Deep Learning: un nuovo (ultimo?) "AI spring"
Deep Learning: un nuovo (ultimo?) "AI spring"
Deep Learning: un nuovo (ultimo?) "AI spring"
Deep Learning: un nuovo (ultimo?) "AI spring"
"It may be a hundred years before a computer beatshumans at ’Go’, maybe even longer"
The New York Times, 1997
Alpha go vs Lee Sedol, 2016"Master of Go Board Game Is Walloped by Google
Computer Program."
The New York Times, 2016
Deep Learning & ReasoningLevel = Superhuman
Deep Learning & Computer visionLevel = Superhuman
Deep Learning & Voice RecognitionLevel = Same as human
Deep Learning & Speech SynthesisLevel = Close to human
Perchè ora?
AlgoritmiRumelhart et Al., Learning representations by back-propagating errors.1986Corinna Cortes and Vladimir Vapnik. Support-vector networks. 199550+ nuovi paper al giorno su arxiv.
Strumenti open sourceScikitTensorflowCaffeKerasTheano
Strumenti in pratica: dante-botdef build_graph(batch_size, seq_len, vocab_size, rnn_size): x = tf.placeholder(tf.int32,[batch_size, seq_len]) y = tf.placeholder(tf.int32,[batch_size, seq_len]) cell = rnn_cell.GRUCell(rnn_size) init = cell.zero_state(batch_size, tf.float32) embeddings = tf.get_variable('embedding_matrix',[vocab_size, rnn_size])
rnn_inputs = tf.nn.embedding_lookup(embeddings, x) rnn_outputs, final_state = tf.nn.dynamic_rnn(cell, rnn_inputs, initial_state = init)
with tf.variable_scope('softmax') as scope: W = tf.get_variable('W',[rnn_size, vocab_size]) b = tf.get_variable('b',[vocab_size], initializer=tf.constant_initializer(0.0))
rnn_outputs = tf.reshape(rnn_outputs, [-1, rnn_size]) y_ = tf.reshape(y, [-1]) logits = tf.matmul(rnn_outputs, W) + b
predictions = tf.nn.softmax(logits) cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits, y_) loss = tf.reduce_mean(cross_entropy)
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
Dati disponibili
Prezzo storage dati
Potenza computazionale disponibile
Prezzo CPU
"Many of the papers, data sets, and software toolsrelated to deep learning have been open sourced.[...] Software tools like Theano and TensorFlow,
combined with cloud data centers for training, andinexpensive GPUs for deployment, allow small teamsof engineers to build state-of-the-art AI systems."
Chris Dixon, A16Z partner
Come rispondono gliinvestitori?
Come rispondono le corporate?
Social: Pinterest deep-learning-powered reccommender: +30% repinsEcommerce: The Clymb ha avuto +175% revenue/1000 promo email,-72% churn (HBR)Customer service: 85% interazioni senza interazione umana nel 2020(Gartner).
Marketing & sales: 76% delle aziende che usano ML hanno aumentato leproprie revenue (Accenture)Fintech: Banche che usano ML per promuovere prodotti ottengono +10%sales e -20% churn (Accenture).Ingegneria: Sight ha ridotto downtime macchine 50% e aumentatoperformance del 25%...
Conclusioni"AI is the new electricity. Just as 100 years ago
electricity transformed industry after industry, AIwill now do the same."
Andrew Ng, Chief Scientist at Baidu
È il momento di passare da "ML as a product" a"ML as a feature".
Q&A