AI and Deep Learning Yoshua Bengio A new revolution seems to be in the work after the industrial revolution. And Machine Learning, especially Deep Learning, is at the epicenter of this revolution. Deep Learning Breakthroughs Computers have made huge strides in perception, manipulating language, playing games, reasoning, ... 3 Intelligence Needs Knowledge • Learning: powerful way to transfer knowledge to intelligent agents • Failure of classical AI: a lot of knowledge is intuitive • Solution: get knowledge from data & experience Deep Learning Machine Learning Artificial Intelligence 4 Bypassing the curse of dimensionality We need to build compositionality into our ML models Just as human languages exploit compositionality to give representations and meanings to complex ideas Exploiting compositionality can give an exponential gain in representational power Distributed representations / embeddings: feature learning Deep architecture: multiple levels of feature learning Prior assumption: compositionality is useful to describe the world around us efficiently 5 Deep Representations: The Power of Compositionality • Learned function seen as a composition of simpler operations, e.g. inspired by neural computation • Hierarchy of features, concepts, leading to more abstract factors enabling better generalization • Again, theory shows this can be exponentially advantageous Why multiple layers? The world is compositional 6 What’s New with Deep Learning? • Progress in unsupervised generative neural nets allows them to synthesize a diversity images, sounds and text imitating unlabeled images, sounds or text GANs (NIPS’2014) Random Generator Fake Vector Network Image Discriminator Network Random Training Real Index Set Image 7 What’s New with Deep Learning? • Incorporating the idea of attention, using GATING units, has unlocked a breakthrough in machine translation: Neural Machine Translation (ICLR’2015) Higher‐level Softmax over lower locations conditioned on context at lower and higher locations Lower‐level • Now in Google Translate: current n‐gram neural net human translation translation translation Human evaluation 8 Medical Image Classification @ Montreal Detect cancer cells by deep learning Accuracy > 90%, real Imagia time GI Experts (Key ~ 90% Opinion Leaders)* GI Doctors Trained by ~ 75% KOLs* *(D. Rex, 2015) Still Far from Human-Level AI • Industrial successes mostly based on supervised learning • Learning superficial clues, not generalizing well enough outside of training contexts, easy to fool trained networks: • Current models cheat by picking on surface regularities Humans outperform machines at unsupervised learning • Humans are very good at unsupervised learning, e.g. a 2 year old knows intuitive physics • Babies construct an approximate but sufficiently reliable model of physics, how do they manage that? Note that they interact with the world, not just observe it. Latent Variables and Abstract Representations P(h) • Encoder/decoder view: maps Q(h|x) Abstract representation between low & high‐levels space • Encoder does inference: interpret encoder decoder P(x|h) the data at the abstract level • Decoder can generate new configurations • Encoder flattens and disentangles the data manifold data space 12 Maps Between Representations x and y represent different modalities, e.g., image, text, sound… Can provide 0‐shot generalization to new categories (values of y) (Larochelle et al AAAI 2008) 13 How deep learning could be used for medical applications • Handle high‐dimensional inputs which doctors do not have time to look at • Genome, mRNA expression levels • 3‐dim images (e.g. brain scans) or videos (e.g. intestine) • Handle patient history data • Handle textual data (e.g. doctors’ reports) • Handle missing data and censored data • Predict future outcomes or imitate doctors’ decisions • Estimate probabilities of future events • BUT THERE MUST BE ENOUGH CASES 14 Challenges • Collecting enough data, examples • Data may be incompatible or inconsistent (different devices, hospitals, etc.), but there are ways to learn invariant representations • Confidentiality, anonymity, legal permission issues, there is research on anonymized representations • Knowledge gap between healthcare professionals and AI experts • Incompatible value systems for publications • Difficulty of interpreting the results of complex predictive models (high‐order interactions between many variables) and yield causal conclusions when confounders are not observed 15 The So-Called Black Box • Problem with complex machine learning like DL: no simple interpretation of the learned predictor’s mechanism • But would you rather use a very precise diagnostic tool which saves your life but is not fully explained or one that is very simple but makes wrong predictions and lets you die? • Doctors are not always able to completely justify their decisions either, because their brain is performing a complex computation 16 From Controlled Groups to Population Data and Randomized Interventions • The traditional control‐group setting does not have enough power to allow learning complex dependencies and personalized predictions. • How to avoid confounders? • This is an active research area • Randomized interventions in principle solve the problem • Simulated randomized intervention might be possible if we model doctors’ decisions as P(treatment|patient) 17 Deep Data Fusion • Deep nets are very good at combining multiple sources of data, multiple sensors or modalities • With convnets, can have separate pre‐processing stages for each modality, then CONCATENATE the representations before continuing processing Need to map to the same spatial scale, or ‘copy’ a non‐spatial modality at all positions. 18 Combining Heterogeneous Sources with Missing Modalities • Different measurements may have been made for different patients, each with a different subset of modalities • Variable number of inputs! Cases • Handled by HeMIS: Havaei et al MICCAI 2016 19 HeMIS Backend Abstraction Frontend Only available modalities are provided for a given case HeMIS Backend Abstraction Frontend Only available modalities are provided for a given case HeMIS Backend Abstraction Frontend Only available modalities are provided for a given case HeMIS Backend Abstraction Frontend Only available modalities are provided for a given case Results (BRATS) FLAIR T2W T1W T1c HeMIS Truth Results (BRATS) FLAIR T2W T1W T1c HeMIS Truth Results (BRATS) FLAIR T2W T1W T1c HeMIS Truth Results (BRATS) FLAIR T2W T1W T1c HeMIS Truth Results (BRATS) FLAIR T2W T1W T1c HeMIS Truth ults (BRATS) onclusions & Future Perspective The deep learning literature is rich and growing fast and an mportant source of inspiration for many medical applications Current research in unsupervised deep learning could help to handle large quantities of unlabeled data or to relate multiple modalities to each other thanks to their ability to discover good representations With the MILA, Montreal has become an international hub for deep learning & AI research, both scientifically and in terms of nnovation and industrial investment Montreal Institute for Learning Algorithms.