338 

   339 <h3 id="sec2.23">Gael Varoquaux(Affiliation: INRIA Parietal, Neurospin, Saclay, France): Machine learning as a tool for Neuroscience</h3>

   340 <h4>Abstract</h4>

   341 <p>For now two decades, functional brain imaging has provided a tool for

   342 building models of cognitive processes. However, these models are

   343 ultimately introduced without a formal data analysis step. Indeed,

   344 cognition arise from the interplay of many elementary functions. There

   345 are an exponential amount of competing possible models, that cannot be

   346 discriminated with a finite amount of data. This data analysis problem is

   347 common in many experimental science settings, although seldom diagnosed.

   348 In statistics, it is known as the <b>curse of dimensionality</b>, and can be

   349 tackled efficiently with machine learning tools.</p>

   350 <p>

   351 For these reasons, imaging neuroscience has recently seen a

   352 multiplication of complex data analysis methods. Yet, machine learning is

   353 a rapidly-evolving research field, often leading to impenetrable

   354 publication and challenging algorithms, of which neuroscience data

   355 analysts only scratch the surface. 

   356 </p>

   357 <p>

   358 I will present our efforts to foster a general-purpose machine-learning

   359 Python module, <b>scikit-learn</b>, for scientific data analysis. As it aims

   360 to bridge the gap between machine-learning researchers and end-users, the

   361 package is focused on ease of use and high-quality documentation while

   362 maintaining state-of-the-art performance. It is enjoying a growing

   363 success in research laboratories, but also in communities with strong

   364 industrial links such as web-analytics or natural language processing. 

   365 </p>

   366 <p>

   367 We combine this module with high-end interactive

   368 visualization using <b>Mayavi</b> and neuroimaging-specific tools in <b>nipy</b> to

   369 apply state of the art machine learning techniques to neuroscience:

   370 learning from the data new models of brain activity, focused on

   371 predictive or descriptive power. These models can be used to perform

   372 "brain reading": predicting behavior our thoughts from brain images. This

   373 is a well-posed <b>supervised learning</b> problem. In <b>unsupervised</b>

   374 settings, that is in the absence of behavioral observations, we focus on

   375 learning probabilistic models of the signal. For instance, interaction

   376 graphs between brain regions at rest reveal structures well-known to be

   377 recruited in tasks. 

   378 </p>

   379 <p>

   380 Optimal use of the data available from a brain imaging session raises

   381 computational challenges that are well-known in large data analytics. The

   382 <b>scipy</b> stack, including <b>Cython</b> and <b>scikit-learn</b>, used with care, can

   383 provide a high-performance environment, matching dedicated solutions. I

   384 will highlight how the *scikit-learn* performs efficient data analysis in

   385 Python. 

   386 </p>

   387 <p>

   388 The challenges discussed here go beyond neuroscience. Imaging

   389 neuroscience is a test bed for advanced data analysis in science, as it

   390 faces the challenge of integrating new data without relying on

   391 well-established fundamental laws. However, with the data available in

   392 experimental sciences growing rapidly, high-dimensional statistical

   393 inference and data processing are becoming key in many other fields.

   394 Python is set to provide a thriving ecosystem for these tasks, as it

   395 unites scientific communities and web-based industries.

   396 </p>

   397 <h4>Slides</h4>

   398 <p>To be uploaded</p>